This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON ROBOTICS 1

Soft Robots Modeling: A Structured Overview

Costanza Armanini , Member, IEEE, Frédéric Boyer, Anup Teejo Mathew ,
Christian Duriez , Senior Member, IEEE, and Federico Renda , Member, IEEE

Abstract—The robotics community has seen an exponential

growth in the level of complexity of the theoretical tools presented
for the modeling of soft robotics devices. Different solutions have
been presented to overcome the difficulties related to the modeling
of soft robots, often leveraging on other scientific disciplines, such
as continuum mechanics, computational mechanics, and computer
graphics. These theoretical and computational foundations are
often taken for granted and this leads to an intricate literature
that, consequently, has rarely been the subject of a complete review.
For the first time, we present here a structured overview of all
the approaches proposed so far to model soft robots. The chosen
classification, which is based on their theoretical and numerical
grounds, allows us to provide a critical analysis about their uses
and applicability. This will enable robotics researchers to learn the
basics of these modeling techniques and their associated numerical
methods, but also to have a critical perspective on their uses.
Index Terms—Dynamics, flexible robots, kinematics, modeling,
control, and learning for soft robots. Fig. 1. Some resolution methods that have been distinguished in this re-
view. (a) FEM, (see Section IV-A). (b) Shooting method (see Section V-A2).
(c) Relative modal-Ritz reduction (Section V-B4). (d) Pseudorigid approach (see
Section VII-B).
I. INTRODUCTION
HE term soft robot appeared for the first time in a scientific
T paper in [1], describing a McKibben pneumatic artificial
muscle a family of braided pneumatic actuators developed in
in almost every possible technological field, from biomedical
engineering to aerospace and underwater robotics. The increas-
the 50 s to assist polio patients. Even though they were not ex- ing interest in soft robotics is demonstrated by the vast number
plicitly called soft robots, McKibben actuators represent the first of review papers that have been published to summarize the
example of a robotic device exploiting its compliance to achieve employed techniques, the achievements and the future prospects
improved performances with respect to their exclusively rigid of this research field [2], [3], [4], [5]. Among the available
counterpart. Since then, soft robotics has been one of the fastest reviews on the theoretical modeling components, some have
growing research community in the last decades. Many different been published for specific groups of approaches [6], [7], [8], [9],
soft robotics devices and actuators have been presented, ranging [10], some for specific application fields [11], [12], and others
for specific families of robots [13].
Unlike traditional rigid robots, soft robots are infinite-
Manuscript received 29 September 2022; accepted 8 December 2022. This dimensional systems whose time evolution is governed by highly
work was supported in part by the US Office of Naval Research Global under nonlinear partial differential equations (PDEs), generally not
Grant N62909-21-1-2033 and in part by the Khalifa University of Science and analytically integrable. Based on this observation, the search for
Technology under Grants CIRA-2020-074 and RC1-2018-KUCARS, and in
part by the French National Research Agency (ANR) through the COSSE- a tradeoff between accuracy and numerical usability for robotic
ROOTS research under Project ANR-20-CE33-0001 (2020-2024). This paper purposes (such as control and optimization), is undoubtedly the
was recommended for publication by Associate Editor C. Rucker and Editor A. great challenge of soft robot modeling. For the first time, we
Menciassi upon evaluation of the reviewers’ comments. (Corresponding author:
Costanza Armanini.) present here a comprehensive literature review of the techniques
Costanza Armanini, Anup Teejo Mathew, and Federico Renda are with the that have been presented to model soft and continuous robots.
Department of Mechanical Engineering, Khalifa University of Science and Given the enormous quantity and diversity of contributions on
Technology, Abu Dhabi 127788, UAE, and also with the Khalifa University
Center for Autonomous Robotic Systems, Khalifa University of Science and this theme in recent years, choices had to be made to classify
Technology, Abu Dhabi 127788, UAE (e-mail: [email protected]; them. The chosen classification is based on the mathematical
[email protected]; [email protected]). techniques themselves (see Fig. 1), and not on their uses (de-
Frédéric Boyer is with the LS2N Laboratory, Institut Mines Telecom Atlan-
tique, 44307 Nantes, France (e-mail: [email protected]). sign, simulation, control). Although this choice may seem a bit
Christian Duriez is with the Inria, CNRS, Centrale Lille, Team DEFROST, arbitrary at first sight, it has the great advantage of highlighting
University Lille, 59000 Lille, France (e-mail: [email protected]). the structural similarities and differences that gather and/or
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TRO.2022.3231360. discriminate all the modeling approaches proposed so far in
Digital Object Identifier 10.1109/TRO.2022.3231360 the field. Moreover, it allows us to go back to the theoretical
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON ROBOTICS

roots of the different approaches and to give their historical X that identifies each particle. In particular, X can be defined
origins, sometimes from other communities. We believe that as the coordinates of the particles of the body when it is in
this will enable robotics researchers to learn the basics of these a stress-less configuration, called reference configuration. The
modeling techniques and their associated numerical methods, motion of a continuum body is defined by a continuous sequence
and to have a critical perspective on their uses. The rest of of configurations along time. A change in the configuration of
this article is organized as follows. In Section II, some general a continuum body results in a displacement, which usually has
concepts are drawn and the classification employed to discuss two components: a rigid-body displacement and a deformation.
all the modeling approaches is presented. In Section III, we To describe the internal deformation state of the body, it is
present the formulation of models that are based on continuum necessary to define its strain (time) rate, which is a combination
mechanics theories, while their numerical resolution is presented of the gradient of the velocity fields. The definition of the strains
in Sections IV and V. In Section VI, we present the approaches, needs to be objective, i.e., observer independent. Finally, beyond
which rely on the hypothesis that the deformed shape of the kinematics, the closed formulation describing the time evolution
robot follows a specific geometrical curve. Proceeding with the of a continuum medium is given by the following.
level of discretization, in Section VII, the approaches that arise 1) a principle of the dynamics (as before), providing the
from an a priori discretization of the soft body are presented, PDEs relating the stress with the acceleration of the parti-
while in Section VIII, data-driven models are discussed. Sec- cles and the external forces applied inside and across the
tion IX presents the software implementation of the proposed boundaries of the medium;
approaches. Finally, Section X concludes this article. 2) a set of geometric BCs;
3) a constitutive law that relates the time evolution of the
stress to that of the strains.
II. BASIC CONCEPTS AND CLASSIFICATION The resulting dynamic equations are, in general, highly non-
Mechanical modeling is a set of mathematical techniques, linear and characterized by an infinite-dimensional configura-
allowing us to represent the evolution of a material system. It is tion space. The soft robotics modeling literature could be viewed
defined by the following key concepts [14], [15]. as the story of how these extremely complex equations, when
1) Configuration: The set of position of the system’s particles applied to soft robots, can be discretized and solved. It should be
compatible with the internal (joints) and the external [geo- noted that discretization and reduction are two distinct concepts
metric boundary conditions (BCs)] kinematic constraints. that should not be confused. A discretization is employed to
At any given time, the configuration of a mechanical obtain a numerical solution to the nonlinear problem and, for
system defines a subset of the ambient Euclidean space. example, the finite differences methods fall in this category. On
2) Generalized coordinates: A set of parameters describing the other side, a reduction consists in the depiction of a basis of
any configuration. They are the coordinates of a unique functions allowing a proper description of the kinematic fields of
point in an abstract space or “configuration space.” They the body. These features are not always clearly expressed in the
can be absolute when they refer to a fixed inertial frame, developing of a theoretical model, or they might be implicitly
or relative when they refer to a moving frame covarying imposed by the assumptions of the model itself or its numerical
together with the system. resolution. One of the aims of this work is to shed light on
3) Kinematic map: It takes as input the generalized coor- the underline structure of the proposed models using a unified
dinates and it returns the configuration of the system. language, which facilitate their understanding and comparison.
This is also called forward kinematics, while the inverse With this as a guideline, we classified the modeling of soft robots
kinematics represents exactly the opposite process, i.e., as follows.
the calculation of the coordinates (typically, the actuation 1) Continuum mechanics models: They are characterized by
ones) required to obtain a specific configuration of the a continuous (infinite-dimensional) configuration space, and on
system. physical considerations about the soft bodies deformations. As
4) Dynamic principle: Such as Newton’s laws, d’Alembert’s such, they benefit from a physically rigorous definition of the
principle, or Hamilton’s principle. Such a principle pro- kinetic and potential energy of the system. This family of models
vides the equations of motion (EoM) of the system gov- can employ both absolute and relative coordinates reduction and
erning the temporal evolution of its configurations. they are derived from continuum mechanics theories. When no
The abovementioned characteristics of a mechanical model specific assumptions are made, they are based on the classical
can be identified relatively easily for a traditional rigid robot. three-dimensional continuum mechanics theory, while other
However, this is not the case for soft robots, due to their contin- approaches have been presented for surface structures (shells,
uum nature. In fact, while traditional rigid robots can be fully membranes) or slender structures (beam, rods). In particular, the
represented by some finite discrete set of frames, in soft robotics latter include Cosserat, Kirchhoff, and nonlinear Euler Bernoulli
the robot is a continuum of particles. Let us draw some general beam theories, which are frequently employed in soft robotics.
considerations on the main steps that are required to obtain a 2) Geometrical models: They are based on geometrical as-
model for a continuum body. In the Lagrangian description of sumptions on the deformed shape undertaken by the soft body
solid continuum mechanics, the body’s configuration is param- when specific loads are applied. For this group, the central role
eterized through positional fields, which depend both on time is taken by the generalized coordinates, on which the system’s
and on the material coordinates, i.e., a set of continuous labels kinetic and potential energy are defined. In particular, two main
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 3

groups emerged. The first one includes the “functional” models,

which all rely on the assumption that the deformed shape of
the body resembles a theoretical space curve represented by
a specific mathematical function. In this case, the generalized
coordinates are usually absolute in nature. The second group
includes the widely known piecewise-constant-curvature (PCC)
models, which are based on the discretization of the continuous
soft body in a finite number of sections having circular arc shape,
with intrinsically relative coordinates.
3) Discrete material models: As the name suggests, these
models are based on a discretization of the continuous body in
a finite number of discrete material components. As such, they Fig. 2. Scheme of the considered body.
are characterized by an a priori finite-dimensional configuration
space of absolute and/or relative coordinates, the relative ones
being usually preferred in practice. normal n ∈ R3 , Fig. 2. In general, volume and surface forces
4) Surrogate models: The configuration of the system is are applied upon B, and modeled by some volume and surface
obtained using sets of data and a learning process. The great vector densities, b ∈ R3 and t ∈ R3 , respectively. While the first
majority of the approaches falling in this group use neural represents the external forces acting on the volume, the latter
networks (NNs) models and machine learning algorithms. represents the forces that B exchanges with its surroundings (the
This classification has been defined to ease the explanation other parts of Ω) across its boundary ∂B, which are also known
of the models, in an effort to group them based on the different as internal forces. In particular, t is called stress vector and we
parametrization path that they employ (see Fig. 1). Clearly, this admit the Cauchy theorem, i.e., t = σn, where σ ∈ R3 × R3
is only one of the classifications that can be conceived and, in is the Cauchy’s stress tensor. It is then possible to express the
some cases, it is possible to note some overlapping between the balance equations for any part B of Ω, as
models that we will try to underline while going through them  
and which are summarized in Section X. (b − ρv̇) dV + tdS = 0 (2)
B ∂B
III. CONTINUUM MECHANICS MODELS where ρ is the mass density, v is the body velocity, and “·”
Soft robotics is an extremely interdisciplinary field and con- denotes the partial time-derivation ∂./∂t. Using divergence the-
tinuum mechanics represents the most influential community orem, and remarking that (2) holds for any material subdomain
for the theoretical modeling of soft robots. Classical elasticity B, yields the Cauchy equilibrium equations
theories have been used for centuries to precisely model the ∇ · σ + b = ρv̇ (3)
mechanics of continuum bodies and they offer an established
and general framework that is already available to the soft where “∇·” stands for the divergence operator. In order to
robotics community. Thus, one of the main goals of using these close the formulation, (3) needs to be supplemented with a
approaches in the robotics community is to make them compu- definition of the strains and a constitutive relation characterizing
tationally efficient for the purpose at hand, while maintaining the response of the body’s material under external forces (or any
the realism of the models produced. external stimuli), i.e., a relationship between the stress tensor σ
In the following, we will start considering in Section III-A, the and an objective strain measure as the Green–Lagrange tensor
models that are obtained from the three-dimensional elasticity field
theory, treating the soft body as a continuum medium. We will 1 T 
then move in Section III-B to the so-called “director approaches” E= ∇r ∇r − ∇r To ∇r o (4)
2
[16]. For slender bodies, these approaches are all based on the
where ∇r is the gradient (the Jacobian) of the nonlinear map (1),
Cosserat rod theory which encompasses the Kirchhoff and the
while the index o denotes the positional field over a reference
large deflections (nonlinear) Euler Bernoulli theories.
configuration. There are different ways to define a constitutive
relation and the first and most famous example is the Hooke’s
A. Classical 3-D Models
law for linear elastic materials, which in the 1-D case reads
Defining a (soft) body as a set of material particles Ω labeled σ = E. One other more general constitutive relation has been
by their (material, in general curvilinear) coordinates X, the aim proposed by Green, employing the concept of strain energy
of any three dimensional theory is to predict the time-evolution function. A Green elastic material (also known as hyperelastic
of its configuration r defined by material) is a type of medium for which the stress–strain rela-
tionship can be derived from a strain energy density function (or
r(·) : X ∈ Ω → r(X) ∈ R3 . (1)
stored-energy function). This function depends symmetrically
In order to proceed, the first step consists in obtaining the on the principal stretches (the eigenvalues of (4)) λ1 , λ2 , and
balance equations of the system. Shall we consider a material λ3 , which, for incompressible materials satisfy the constrain
subpart of the body B ⊂ Ω, having frontier ∂B and outward unit λ1 λ2 λ3 = 1. It is worth to remind that the stretch ratios λi
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON ROBOTICS

are defined as the ratio between the stretched length li and the
undeformed one. Many soft robotics devices and components are
realized with rubber-like materials, which, in the case of static
deformations, are often treated as hyperelastic (while, in general,
they can exhibit other nonlinear behaviors, such as hysteresis,
visco-elasticity, and stress softening). Beyond Green’s model,
the Ogden [17] material model represents one of the most general
framework for the modeling of hyperelastic materials. In such
a material, the strain energy density is expressed in terms of
the principal stretches of the left Cauchy–Green strain tensor
B = (∇r∇r −1 −1 T
o )(∇r∇r o ) as

n
μ j  αj α α 
U (λ1 , λ2 , λ3 ) = λ1 + λ2 j + λ3 j − 3 (5) Fig. 3. Cosserat rod representation.
α
j=1 j

where n, μj , and αj are material constants. In particular, for

n = 1 and α = 2, the Neo–Hookean model is obtained, while cross-sections (supporting the directors), and labeled by a single
the Mooney–Rivlin model for an incompressible material is material coordinate X. In this case, the rod is called a Cosserat
obtained for n = 2, α1 = 2, α2 = −2. Once the strain energy rod [21]) and its configuration is defined by the positional vector
function is defined, the principal Cauchy stresses σ1 , σ2 , and σ3 field r ∈ R3 of its center-line, and the orientation (rotation) field
are related to the principal stretches through the equations R ∈ SO(3), of its (mobile) cross-sectional basis (of directors).
Both these fields are functions of X ∈ [0, l], where l is the rod
∂U length at rest. With this parametrization, the balance equations
σ i = λi −p (6)
∂λi of the rod can be derived with the Newton’s law and Euler’s
where p represents the Lagrange multiplier (or pressure), associ- theorem, by considering an arbitrary subinterval [a, b] ⊂ [0, l]
ated with the incompressibility constraint. It is worth to remind of the rod (see Fig. 3). Across any X-section splitting the rod
that the Cauchy principal stresses are related to the correspond- in two parts, the action of the side Y ≤ X upon the other side
ing nominal (or engineering) stresses σe by the relationship is modeled by the stress resultant vector n(X) and the couple
σe,i = σi λ−1
i . The great majority of the materials that are em-
stress m(X) in the global frame. The mass density of the rod
ployed within the soft robotics community can be considered as is ρ, I is its cross-sectional inertia tensor, while v and ω are
hyperelastic. However, most soft robotics modeling approaches the linear and angular velocities of cross-sections in the global
still rely on a linear-elastic material assumption, as they are frame. The rod is also subject to some external forces f¯, and
mostly focused on the description of the large deformation of moments l̄ densities per unit of X, in the global frame. Stating
the body, rather than the large strains. the linear and angular (Newton–Euler) balance equations over
the subdomain [a, b], and taking the limit of the interval going
B. Directors Approaches to 0 provides the dynamic equilibrium [16]
⎧ 
The abovementioned 3-D model applies to material bodies of ⎨n + f¯ = ρAv̇
any geometry (at rest). However, there are many situations where m + r  × n + l̄ = ∂t (ρI · ω) (7)
⎩
the considered bodies are flattened or elongated, i.e., where two n(0) = −f 0 , n(l) = f l , m(0) = −l0 , m(l) = ll
or one of the material dimensions dominate the others. In the
first case, we speak of plates or shells, in the second of beams where from now on, a “ ” denotes the space-partial derivative
or arches. In these cases, the full 3-D position field of (1) can be ∂(·)/∂X, while l0,l f 0,l are the moments and forces applied in
developed in Taylor series around a reference material surface X = 0 and X = l, respectively. In order to close the formula-
for shells, or a material line for beams, and the dependence on tion, (7) needs to be supplemented with a definition of strains
small orthogonal material dimensions analytically preintegrated and a stress–strain law. In the assumption of finite deformations
into the thickness. At leading order, this provides reduced mod- and small strains, one can use as strain measurements the two
els where the positional field r of (1) is replaced by that of 3 × 1 vectors
the reference surface or line, plus vector fields supporting the  = RT (r  − r o ) , a = K − K o (8)
remaining small dimensions, named directors. For both 2-D or
1-D reduced media, the resulting models are due to Kirchhoff where the index o denotes a reference configuration of the rod,
and Reissner depending whether transverse shearing is neglected K = (RT R )∨ is the vector of material torsion and curvature in
or not. In robotics, except for modeling soft robots inspired of the cross-sectional frame, where ∨ changes any skew symmetric
octopus [18], [19], [20], the great majority of these reduced matrix W ∈ R3 × R3 into W ∨ ∈ R3 s.t. ∀ V ∈ R3 , W V =
models are beam models. To obtain such models, one can advan- W ∨ × V , while conversely ∧ is s.t. Ŵ ∨ = W . In this context,
tageously replace the limit process applied to the 3-D model [16], the constitutive law takes the usual linear Hookean form
by a simpler approach consisting in directly modeling the rod as
a material line along which is continuously stacked a set of rigid n = RH  , bmm = RHa a (9)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 5

with H = diag(EA, GA, GA) and Ha = diag(GI1 , EI2 , the constitutive law (9). In both approaches, the tendons are ide-
EI3 ) defining the usual cross-sectional linear and angular stiff- alized as inelastic and friction-less force transmitters. Initiated
ness matrices in the cross-sectional frames, respectively. In the in the planar case in [25] and extended to the three-dimensional
recent years, the abovementioned model has been reformulated case in [26], the first approach is based on the application of the
and exploited in the modern language of Cosserat rod theory Newton’s laws to the rod and to each of the tendons, isolated
on Lie groups. Introduced in biorobotics [22], [23] to study separately. Then, using the action-reaction principle provides
dynamics of hyper-redundant locomotors, the configuration of the internal forces and moments exerted by a set of na tendons
a soft body is directly defined as a curve onto the rod
  na  
g(·) : X ∈ [0, l] → g(X) ∈ SE(3) (10) l̄ di × tci
=− ui (14)
f̄ tci
i=1
where g(X) denotes the 4 × 4 homogeneous matrix of the X-
cross sectional frame, with positional and rotational components where di (X) and tci (X) are, respectively, the position vector
denoted r(X) and R(X), respectively. The space variations of of the tendon routing i with respect to the rod’s backbone and
the field g can be entirely described by the field of space-twists its unit tangent vector, both expressed at X, while ui (t) is the
ξ = (K T , ΓT )T ∈ R6 , which defines a continuous geometric uniform tension (ui = 0) transmitted along the tendon. In the
model of the rod second approach [27], [28], applying the virtual work principle
to the whole arm, allows modeling the effect of tendons by
g = g
ξ , g(0) = g 0 . (11) changing the constitutive law (9) into the active one [28]
With these definitions, the strain state (9) can be expressed      na  
m RHa a di × tci
as a unique field of twist  = (Ta , T )T = ξ − ξ o . Similarly, = + ui (15)
n RH  tci
replacing X by the time variable t, yields i=1

where the derivatives of tci (X) appearing in (14) are now taken
ġ = g η (12)
in charge by the space derivations of the equilibrium equations
which defines the field of velocity twist η = (ΩT , V T )T ∈ R6 , (13).
where Ω(X), V (X) ∈ R3 are the angular and linear velocity of Remarkably, other nonlinear beam theories can be defined as
the X-cross-sectional frame in its mobile basis, which are related submodels of the abovementioned theory, often referred to as
to their inertial counter-part through: Ω = RT ω, V = RT v. the Reissner beam model [29]. The Kirchhoff rod theory can
On SE(3), the dynamic equilibrium equations of a rod subject be obtained by fixing in ξ of (11), the first component Γ1 of
to a density of external wrench F̄ on ]0, l[ and two tip external Γ to one, and the two others to zero, which prevents the beam
wrenches F 0 and F l at X = 0 and X = l, respectively, are from stretching and shearing, respectively. By imposing that the
expressed in the cross-sectional frames, in the form [23], [24] motion of a Kirchhoff rod is planar and ignoring its thickness,

the stress equilibrium (13), restricted to statics, can then be
Mη̇ − adTη Mη = Λ − adTξ Λ + F̄ integrated explicitly. The model is then that of the elastica, i.e.,
(13)
Λ(0) = −F 0 , Λ(l) = F l an elastic line of minimal curvature energy, and we have
where ad is the adjoint action of the Lie algebra, M is the EI(X)θ = Mext (16)
6 × 6 cross-sectional inertia matrix, and Λ models the stress
field along the beam (it is the dual counterpart of the strain where θ (X) is its curvature in the plane, EI(X) its bending
field). The balance (13) can be derived by applying Hamilton’s stiffness, and Mext (X) is the integral of the external bending
principle in the framework of Lagrangian reduction theory [19], moments about the center of the X-cross section, acting on the
or deduced from (7), by using the relations between vectors remaining part (Y > X) of the beam.
of components in the inertial and cross-sectional frames: η =
((RT ω)T , (RT v)T )T , Λ = ((RT m)T , (RT n)T )T , F 0,l = IV. NUMERICAL RESOLUTION OF 3-D CONTINUUM
((RT l)T0,l , (RT f )T0,l )T , and F̄ = ((RT l̄)T , (RT f¯)T )T . Equa-
FORMULATIONS
tion (13) is first-order PDEs that govern the time-evolution of To integrate the dynamics of a soft manipulator modeled by
velocities η along the rod. Thus, they need to be supplemented 3-D continuum mechanics, one needs to solve for each material
with the kinematic model (12) allowing us to reconstruct (by subdomain a closed formulation obtained by the PDEs (3) with
integration) the time-evolution of the configuration g. a definition of strains, a constitutive law of the type of (6) and
To apply the Cosserat rod model to soft robots, an actuation BCs (some of them imposed by the connection of the bodies).
model must be introduced into the abovementioned formulation. The usual numerical resolution methods use finite differences to
In general, this has to be done on a case by case basis, depending discretize the time axis through explicit or implicit integration
on the specific technology and design of the robot. However, this schemes. However, the spatial discretization can be achieved
task has been fully accomplished in the case of tendon-driven differently by using finite differences, boundary elements, or
robots. Two equivalent approaches have been proposed in this finite-elements method. The first approach is based on Taylor
case, depending whether the effect of actuation is introduced approximations of the PDEs, but it is difficult to apply to complex
through the densities of external wrenches (l̄, f̄ ) or through a shaped objects, while the second is based on Green’s theorems
field of internal stress Λ added to the elastic restoring ones of and restricted to linear problems. In contrast, the finite-elements
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON ROBOTICS

method has the advantage of inheriting the variational struc- (Abaqus), [33] (ANSYS). However, their generality intrinsically
ture of the original formulation, which is transformed into a entail an increased computational cost and the simulations can
robust nonlinear optimization problem in the context of the Ritz take a lot of time to converge. There are some popular exam-
method. ples of ad hoc finite-element approaches (and softwares) that
have been specifically proposed for (soft) robotics applications.
A. Finite-Element Method (FEM) for 3-D Formulations In 2007 [34], a group of researchers from different institutes
released simulation open framework architecture (Sofa), an
FEM is a well-known and widely-spread numerical technique
open-source C++ library, which was originally presented as a
for finding approximate solutions to PDEs. It was established
computational environment for medical simulations [35], [36].
by a set of scientific papers in the 40 s and it soon became one
In the following years, Sofa became a comprehensive high-
of the most commonly employed technique for the modeling
performance library that have been widely implemented for
of a wide range of engineering problems. The main property
different application fields, and a SoftRobots plugin was also
of this technique is the subdivision of the problem’s domain
created for the design [37], modeling and control [38], [39]
into a set of smaller parts, called finite elements. These are
of soft robots, including self-collision scenarios [40]. More
obtained by the construction of a mesh [see Fig. 1(a)], which
recently, a model combining the FEM and the discrete Cosserat
represents the numerical domain for the solutions. Once this
approach of Section III-B was presented in [41]. A discussion
space discretization is performed, the continuous field r is
on the architecture of the software is addressed in Section IX,
approximated on each element by a polynomial interpolation of
while here we mostly focus on the modeling part for the soft
its values at the edges of the elements, called nodes, according
bodies. In Sofa, a deformable continuum is modeled using a
to a Ritz reduction approach. Introducing this discretization in
dynamic or quasi-static system of simulation nodes. The node
the above continuous formulation and projecting it on the same
coordinates are the independent DOFs of the object, and they are
trial (polynomials) basis, changes the PDEs (3) into a set of
typically governed by equations of the type (17) and (18). Some
algebraic discrete equations for steady-state problems, or ODEs
approaches have also introduced reduced coordinates and these
for transient problems. In the former case, the static equilibrium
are presented in Section VIII-B. With regards to the actuation, a
becomes
constraint-based approach is employed. Imposing a linearization
Qint (q) = Qext (q) (17) of the internal forces Qint (q i ) ≈ Qint (q i−1 ) + K(q i−1 )dq and
considering the contribution of the actuation constraints, the
where Qint (q) and Qext (q) represent the internal and external static equilibrium at each ith step (17) becomes:
generalized forces, respectively, while q are the generalized
coordinates, i.e., the nodal positions. On the other hand, the Qint (q i−1 ) + K(q i−1 )dq = Qext (q i−1 ) + B(q i−1 )T λ (19)
dynamic equilibrium is given by a system of ODEs

where the term B T λ models the contributions of the Lagrange
q̇ = v
(18) multipliers (i.e., those of the actuators), and B represents the
M v̇ + Qint (q, v) = Qext (q, v)
Jacobian of the constraint equations imposed by the actuators.
where M is the (constant) generalized mass matrix. Note that The following three steps are then performed.
the terms appearing in (17) and (18) are computed integrating 1) a free configuration q free is found for λ = 0 and, for each
the distributed variables element-by-element according to the constraint, the violation δ free is estimated;
selected shape functions, which play the role of the kinematic 2) the solver computes the value of λ through a projection
map, while the position of each element in the mesh is ensured of the mechanics into the constraint space, obtaining the
by the assembly process. Since nodal coordinates are of absolute smallest possible projection space δ = BK −1 B T λ +
nature, the nonlinearities appear in Qint (q, v). They are named δ free ;
geometric or material, depending whether they come from the 3) the final configuration is corrected using the value of the
large displacements (and namely the rotations) of the elements, constraint response q = q free + K −1 B T λ.
or from the constitutive laws. To capture the local finite rotations For a fluidic actuation, λ and δ represent the cavity pressure
along bodies, the corotational approach is often used in three- and volume. For a cable, λ and δ are the cable tension and length.
dimensional finite element (FE) software. One other solution to While in the inverse model case, λ is unknown and can be find
represent the nonlinearities is to use an absolute nodal coordinate by solving a QP type problem [36], in the direct problem case,
formulation (ANCF-FEM) [30]. λ can be unknown too, as this is the case of a cable driven arm
The FEM represents a standard for many popular simulation where the nonnegative tensions condition is modeled as a set of
software, such as Abaqus, ANSYS, and COMSOL. The main unilateral and complementarity (nonholonomic) constraints. Be-
advantage of these software is that they provide a powerful and cause standard finite elements are not differentiable at element
general tool that can be applied to a wide spectrum of physi- boundaries, some authors have recently implemented a moving
cal problems ranging from structural dynamics, fluid-structure least squares (MLS-FEM) formulation, making the deformation
interactions, contact, and thermodynamics. Moreover, they pro- gradient twice differentiable [42], with the aim of applying usual
vide a readymade and user-friendly framework that can be optimization gradient descent techniques to different design
easily employed for specific studies on multiphysics systems. problems related to automated actuation routing [43] and sensor
They have been widely used to model soft robots [31], [32] design [44] for soft robots.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 7

V. NUMERICAL RESOLUTION OF DIRECTORS-BASED algorithm), depending on whether an explicit or implicit time

FORMULATIONS finite-difference scheme is used. In continuum and soft robotics,
The closed formulation of a soft manipulator modeled by finite-differences method has been applied to both statics [48],
Cosserat rods consists, for each rod, of the kinematic model [49] and dynamics [24], [50]. One canonical example of such
(12), the stress balance (13) (where the BCs can depend on the a numerical resolution applied to soft robotics is presented
connection between the rods), and a constitutive law of the type in [24], where the dynamic modeling and simulation of a cable
of (9), with the associated strain definition (8). This formulation driven multibending soft robot arm is addressed. Gathering the
is consistent with the definition of the configurations (10) in geometric model (12), with that of velocities (20)1 , and the
terms of the absolute fields g. For soft manipulators, one can balance of stress (13), in which the model of tendons actuation
(14) and the constitutive law (9) are introduced, allows first to set
alternatively parameterize the configuration with the fields ξ or
 and use (11) as a space-reconstruction equation of g along the forward dynamics BVP in the form ż = f (z, z  z, t), where
z is composed of the angular and linear components of fields g, ,
the rod. Time-differentiating twice this geometric model (11)
and η, while the time dependency is due to the actuation. Then,
provides two further models [23] [45]
approximating, the space derivatives of the kinematic field η
η  = ξ̇ − adξ η with a forward (from the base to the tip) finite-difference scheme,
(20)
η̇  = ξ̈ − adξ̇ η − adξ η̇ and those of the strains field , with a backward one (from tip
to the base), this BVP is changed into a finite-dimensional state
whose spatial integration allows us to reconstruct the absolute
system of the usual form ẋ = f (x, t) that can be time-integrated
velocity and acceleration fields along the rod, from the strain ve-
with a standard explicit Runge–Kutta integrator.
locities and accelerations ξ̇, and ξ̈. In this case, the parametriza-
2) Shooting Methods: The shooting method represents one
tion of configurations is said to be relative and opposed to the
of the most popular numerical approach to solve a BVP along
classical absolute one based on g. In this section, we discuss
one dimension of space or time. In the context of Cosserat rods,
the most important examples of approaches using Cosserat rod’s
it has been introduced in the oceanic engineering community
theory for modeling soft robots. They are separated into “nonen-
of submarine cables as an alternative way to finite differences
ergetic” and “energetic” methods. In the first case, the original
for solving the forward dynamics of these systems [51]. In the
formulation is first set in the form of a space boundary value
case of soft arms, the archetypal BVP is that of the forward
problem (BVP) with BCs partly defined at one end, and partly
static model of a Cosserat rod. It is obtained by gathering (11)
at the other. Then, this BVP is directly discretized on a spatial
with the stress balance (13), in which we remove the velocities
grid with no further reduction. In the second case, often referred
and accelerations. Then, expressing the strains in terms of stress
to as the “Ritz” or “Ritz–Galerkin” method, the configuration
by inverting the constitutive law (9) provides a (forward static)
needs first to be parameterized by some vector fields (absolute
BVP in the form z  = f (z) that can be solved by the shooting
or relative) that are reduced on a truncated functional basis of
algorithm, i.e., by integrating a sequence of initial (proximal)
space-dependent vectors. The components along these vectors
value problems, whose unknown proximal conditions are first
define a finite set of generalized coordinates governed by a set
predicted and then iteratively corrected (with a Newton type-
of Lagrange ODEs in time. In the following, Ritz methods are
algorithm), until a solution of the BVP that fulfills the known
classified into nodal and modal, depending whether the basis is
distal BCs is found [see Fig. 1(b)].
defined by some nodal interpolation polynomials over the finite
In [26], Cosserat theory is used for static and dynamic mod-
elements of a mesh, or defined over the full rod domain.
eling of continuous robots, driven by tendons. The effect of
tendons on the rod is modeled as some densities of external
A. Nonenergetic Approaches forces and moments defined by (14), in which the routing path of
In practice, these approaches consist in extracting from the each tendon is expressed as a function of the rod strain variables.
previous model a closed formulation (absolute, relative, or After some clever algebraic manipulations, an explicit BVP
mixed), which is reshaped into one of the standard forms of of the forward statics in the form z  = f (z, t) is obtained, to
numerical analysis. which the shooting method is applied for the purpose of cal-
1) Finite Differences: The finite-difference method is one of culating the static equilibrium configurations of these systems.
the simplest and oldest methods to solve PDEs. In the case of In [52], the shooting-based resolution was extended from statics
Cosserat beams, it has been widely used by the ocean engi- to the forward dynamics of several designs of continuum robots
neering community to predict the temporal evolution of towed for simulation purposes. Inspired by [51] and [53], an implicit
submarine cables, subjected to low tensions that cause a dynamic time-integrator is used in order to remove the time derivatives of
singularity in the standard catheter model. Developed by Burgess the system of PDEs (12), (20)1 , and (13). Using the constitutive
and his successors [46], [47], the direct dynamics of the cable law (9) and the model of actuation (14), provides a BVP in space
modeled by a shear-less Cosserat rod is formulated as a nonlinear in the usual form z  = f (z, t), that is, solved at each time-step of
BVP in space and an initial value problem (IVP) in time. The the simulation by the shooting method starting from a prediction.
application of a spatial finite-difference scheme transforms this With the rapid expansion of continuum robotics, the approach
continuous formulation into a set of time-ODEs at the grid has been successfully extended to several multisegment designs,
nodes, which can then be solved explicitly, or by a predictor– ranging from multisegment tendon robots [54], to parallel con-
corrector strategy (with correction by the Newton–Raphson tinuum robots [52], [55], to concentric tube robots [56] and [57].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON ROBOTICS

In the first case, the BVPs along segments are connected in series is introduced at the same level as that of positions, by replacing
by their BCs and the approach is reapplied in cascade from the material points of classical medium by the rigid microstruc-
the base to the tip of the arm. In the second case, it is applied tures of a Cosserat medium. The GE character, then imposes to
along each legs and the residual vector of distal conditions must apply the FEM without resorting to any simplifications on finite
integrate the equilibrium equations of the rigid platform. In rotations, except the unavoidable space and time discretizations
the third case, the superimposed tubes must be considered as required by the numerical resolution. The GE-FEM was devel-
nonmaterial domains of varying length, along which the method oped for both shells [63], and rods [64], [65], [66]. The method’s
is applied in cascade. tour de force was to generalize all the key operations of the
3) Continuous Newton–Euler-Based Method: The Newton– FEM (interpolation and space and time-integration), from the
Euler formalism of rigid multibody systems has produced fast linear vector space of positional fields to the curved manifold
and simple algorithms for solving the direct and inverse dy- SO(3). Rooted in Lagrangian mechanics, the resulting dynamic
namics of rigid robots [58]. In [22], the Luh “computed torque equations take a form similar to that of (18), except that q and
algorithm” has been extended to solve the inverse dynamics of v are now defined by the sets of (r, R) and (ṙ, Ω) at the nodes,
fish-inspired hyper-redundant robots modeled as Cosserat rods while an additional Coriolis term survives due to the intrinsic
whose shape is controlled by the internal torque field RT m. curvature of SO(3)
Concatenating (13) with (11), and its first and second time-
derivatives (20), defines a continuous Newton–Euler model, M v̇ + C(v)v + Qint (q, v) = Qext (q, v). (21)
where the rod sections take the place of the rigid bodies and K
Finally, (21) is solved iteratively at each time step according
that of the joint variables of a rigid discrete system. Imposing the
to a standard prediction–correction strategy, in which the time-
curvature time law in this model provides an inverse dynamics
dependence is removed with an implicit geometric time inte-
BVP, which, in contrast to the forward dynamics one, can be
grator on SO(3) × R3 , which preserves the Lie group structure
solved in two decoupled pass (i.e., without resorting to shooting).
and the geometrically exact character of the approach. This
Solving the (locomotion) dynamics of net motions, this algo-
approach, which is probably the most advanced for modeling
rithm has been extensively exploited for the study of bio-inspired
mechanisms of rods subject to rigid motions and finite deforma-
swimming [59]. Although first designed for hyper-redundant
tions, is today used as a reference for others, and commercialized
systems, this algorithm has been recently applied to soft and
in the SAMCEF software through its plugin MECANO [67].
continuum robotics [28]. Remarkably, when pose-dependent
In this software, the model of rods is the full-Cosserat model
external forces (such as gravity) are neglected, the forward static
of Reissner. However, its dynamic resolution is ill-conditioned
BVP enjoys the same decoupling property as the inverse BVP,
when the aspect ratio of the rod increases, i.e., for very slender
a feature that was exploited in [60] for quasi-static simulation
rods. To overcome this limitation, several FEM based on the
of tendon-driven robots with the same two pass approach. Note
nonlinear Euler–Bernoulli or Kirchhoff rod models, were pro-
that in such cases, the Newton–Euler and finite-difference-based
posed, such as the absolute nodal FEM [68] or the Kirchhoff
approaches can lead to the same numerical solutions but from
GE-FEM [65], [69]. Despite their power, these methods have
different points of view.
not yet been adapted to the specific needs of soft robotics as was
4) Collocation Methods: Alternatively to shooting, the for-
done for Sofa. As a result, specific applications, more or less
ward static BVP of soft arms can also be solved using the collo-
related to the original spirit of Simo’s GE-FEM were proposed
cation method. The main idea is to replace the unknown strain
for these systems. In [70], a GE-FEM approach for modeling
field with a polynomial and setting the vanishing of the residual
inflatable robots is presented. A strain field measuring radial
between the two in a finite set of points on the domain, also called
inflation is added to the usual 6 strains of the Reissner model
collocation points. Considering m collocation points, a total of
and the weak form of the static equilibrium in this augmented
m + 1 equations is obtained, providing the parameters that are
space is derived and solved in COMSOL.
required to define an mth order polynomial. In [61], the ODEs
2) Absolute Modal-Ritz Reduction: The Ritz method can be
describing the kinematics of a Cosserat rod are directly solved
applied directly on the full domain of the rod and, in this case, we
in terms of transformations on SE(3), using a combination of
refer it to the modal-Ritz method. The approach can be applied
orthogonal collocation and forward integration through Magnus
both to absolute or relative vector fields. In the first case, one
Expansion on SE(3). In particular, the unknown ξ(X) in (11) is
needs first to parameterize the rod configuration with a set of
expressed as a set of three Chebyshev polynomials.
vector fields (e.g., the positional field along the backbone r and
the vector field of its 3 cross-sections Euler’s angles θ). Then,
B. Energetic Approaches applying the separation of variables, the components of these
fields can be approximated on a truncated basis of Ritz functions
Energetic approaches seek to reduce the continuous models
compatible with the BCs
of Section III-B to a finite set of Lagrangian ODEs in time, thus
retaining the variational (energetic) structure of the modeling. r(X, t) = Φ(X)q r (t) , θ(X, t) = Ψ(X)q θ (t) (22)
1) Absolute Nodal-Ritz Reduction: One of the most powerful
method to capture in an exact manner the geometric nonlinear- where Φ(X) and Ψ(X) are matrices of spatial shape functions,
ities of soft robots is the geometrically exact FEM (GE-FEM) or “modes” (stacked in columns), while q = (q Tr , q Tθ )T is a
introduced by Simo [62]. In this approach, the model of rotations vector of time-dependent generalized (modal) coordinates.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 9

In [71], the Ritz–Galerkin method is used to reduce the Magnus expansions, yielding to the geometric map
weak-form of a Cosserat rod in both statics and dynamics. The 
g(X) = exp Ω(q, X) (24)
kinematics of the rod is obtained expressing equation (11) in an
inertial reference system, placed at the base of the manipulator.
where exp represents the exponential map in SE(3), and
Then, the positional field of the backbone can be approximated
Ω(q, X) is the Magnus expansion of ξ, truncated at a desired
as in (22), while using Kirchhoff assumption and Bishop frames,
order of approximation. The strain reduction and (24) are then
the rotation field is parameterized with a unique further field of
introduced into the expressions of η̇ and η obtained by analytical
torsion angle. Applying (22) to this field and introducing all these
integration of the model of velocities and accelerations along the
approximations into the static balance equations projected onto
rods (20). This provides the kinematic map η = J (q)q̇ and its
the same truncated basis, provides a set of algebraic equations
time derivative η̇ = J (q)q̈ + J̇ (q)q̇, with J (q) the Jacobian
in the usual form (17). Using a numerical root finder finally
between the twist and q̇-spaces. Finally, applying this change
provides the qi that minimize the approximation error. In the
of space in the contributions of external and inertial forces of
dynamic case, applying the same reduction process provides
the weak form of the virtual works, while those of the elastic
ODEs of the form (21) where M and C depend on q due to
and actuation ones are directly deduced of the projection of the
the geometric nonlinearities introduced by the parametrization
active constitutive law (15) on the strain modes, we obtain the
of SO(3). The coefficients qi (t) are then found by explicit
EoM in the classical Lagrangian form (23). Alternatively, to
time-integration of these ODEs.
this analytical Jacobian-based approach, in [28], a numerical
3) Relative Nodal-Ritz Reduction: After the works of Simo,
Newton–Euler-based approach has been proposed to calculate
a small group of authors proposed an alternative GE-FEM based
(23). In this case, the complex double space integrals required
on the polynomial interpolation of strains instead of poses [72].
by the strain-based parametrization are automatically performed
This approach is more accurate, but also more computation-
through the two passes of the continuous computed torque
ally complex than conventional GE-FEM. In the field of soft
algorithm of [22], whose inputs and outputs are reduced on the
robotics, a GE-FEM based on strain parametrization was pre-
strain modes. In [76], the GVS approach is further extended
sented in [73]. Inspired by [66], in [73], a helical shape function
to the modeling of concentric tube robot systems, including
defined with the exponential map is used to express the shape of
the modeling of the tube’s insertion motion in quasi-steady
a manipulator. It is given in terms of the pose at the base (X = 0)
conditions. In [77], the same approach is applied to the un-
and the strain .
steady case to address the “paradoxical” dynamics of the sliding
4) Relative Modal-Ritz Reduction: The modal Ritz approach
spaghetti [78]. The GVS approach has been implemented in a
has been recently applied to the strain fields of Cosserat rods
MATLAB Toolbox, SoRoSim, for the simulation of soft, rigid,
in the context of the geometric variable-strain (GVS) ap-
and hybrid robots [75]. A special case of the GVS is the piece-
proach, [28], [74], [75]. Equation (22), thus, becomes (X, t) =
wise constant-strain approach (PCS) [45], [79], where the strains
Φ(X)q(t) [see Fig. 1(c)]. In this case, the Ritz coefficients
in the sections are assumed to be constant, an approach which
of the strains stand for the generalized coordinates of a set of
has been extended to closed-chain geometries in [80]. In [81],
homogeneous transformations along the soft robot, similar to the
a similar approach to PCS is presented. Based on the constant
joint transformations for rigid ones [15]. This (relative) strain
curvature (CC) assumption, it also incorporates hyperelastic and
parametrization, produces a highly reduced set of ODEs in the
viscoelastic material behaviors. A planar restriction of the GVS,
classical form
named polynomial curvature approach was introduced in [82]
and [83] to address the control of planar soft manipulators. A
M (q)q̈ + C(q, q̇)q̇ + Qint (q, q̇) = Qext (q, q̇) (23)
similar approach was also used in [84] to model planar elastic
flexure joints.
where q is then comparable to the vector of joint angles of a
rigid robot. In [28], the GVS was successfully validated against
C. Analytical-Based Resolutions
the GE-FEM with standard benches in statics and dynamics.
The results showed that the approach can provide very good In some specific loading conditions (especially when the rod
results in terms of accuracy with a few number of generalized is subject to concentrated external loads, and not distributed
coordinates. Moreover, in contrast to absolute Ritz methods, the ones), some analytical solutions to (16) can be obtained in terms
strain functions do not need to fulfill any BCs. On the other hand, of elliptic functions. To illustrate this, let us consider the case of a
unlike the usual absolute Ritz reduction (such as the FEM), it simply supported beam subjected to a concentrated compressive
handles double space integrals which are not easy to calculate. force F applied at its tip. The analytical solution of the Euler
Recent progress has been made to overcome this difficulty and Bernoulli equation then yields [85]
obtain the reduced model (23). In [75], D’Alembert’s principle   2
4EI θ(0)
of virtual works is used to project the Cosserat model (12), (13), F = 2 K sin (25)
from the space of pose fields g onto that of strain coordinates q. l 2
This approach has been achieved in a purely analytical way. It where an uniform cross-section along the length of the beam is
first exploits the fact that since (11) is a system of homogeneous assumed, and K is a complete elliptic integral of the first kind.
first order ODEs of matrix form Y  = Y A(X), the exponen- It should be noted that (25) provides an implicit expression
tial representation of its solutions can be obtained through the of the base angle θ(0) as a function of the applied load F .
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON ROBOTICS

Once θ(0) calculated for a given F , the rotational field θ can The equilibrium is then given by the Euler–Lagrange equations
be obtained through integration of (16), and used to get the or other equivalent principles of dynamics. Note that all these
displacement field of the rod in terms of other elliptic functions. approaches can be considered as some approximations of
When applicable, this approach can offer great advantages for the Cosserat model starting from the model of a sequence of
the study of some specific phenomena related to soft robots deformable curves. As a result, depending on their material
design. EB rod theory was applied to statics and dynamics of consistency, the resulting models can be equivalent to the ones
a soft arm with a continuously rotating clamp subject to a tip derived in the previous section, with particular reference to
dead load for the purpose of analysing snap-back phenomenon the Ritz-based approach of Sections V-B2 or V-B4. Two main
and post buckling behavior [86]. In [87], kineto-static modeling groups of approaches fall in this category: the functional models
of multiple-backbone continuum robots is addressed with EB and the PCC models.
solutions able to express the deflected shapes of the backbones
over a subsegment of the robot. In the paper [88], planar EB A. Functional Approaches
theory is applied to the modeling of continuous tendon-driven
robots, and the static equilibrium of end and spacer disks is Functional approaches probably represent one of the first
derived. In [89], EB theory is applied to the modeling of a attempts to model soft robotic devices. Their main characteristic
fiber-reinforced bending actuator. The soft actuator is modeled is that they all use a chosen mathematical function to describe
as a Neo–Hookean material, where the strain energy functional the desired space curves representing the geometry of the robot.
of (5) reduces to U = μ2 (λ21 + λ22 + λ23 − 3), with μ, the shear One of the first example falling in this category is developed
modulus. The principal nominal stresses are deduced from the in [92], which employs a serpenoid curve to describe the kine-
principal Cauchy stresses of (6). Considering the fiber rein- matics of snake-like robots. Then, the kinematics of the robot is
forcement on the circumferential direction (λ2 = 1) and the obtained by simple forces and torques equilibrium, considering
material incompressibility, provides the three nominal stress and the action of the ground on the robot and the internal force
the Lagrange multiplier p in terms of μ and λ1 . The bending components [93]. One other important example is the so-called
moment is then deduced from stresses at the top and bottom modal approach that have been originally presented for the
layers, and set equal to that of the internal air pressure against modeling of hyper-redundant robots both in statics [94] and in
the distal cap of the actuator. The actuator force is obtained from dynamics [95], while a survey of the approach is presented in [7].
the torque balance, assuming that the actuator is constrained in The central concept of the approach is that of “backbone curve,”
i.e., a 3-D curve r parameterized by the integral representa-
a flat configuration and that no internal moments is generated X
under pressurization. In [90], the approach is extended in order tion r(X, t) = 0 λ(s, t)t(s, t)ds, where λ(X, t) represents a
to capture the effect of pressure on the lateral surface of the inner length scaling factor, while t(X, t) is the unit tangent vector,
chamber of the actuator. In [91], a soft arm made of longitudinal parameterized by any spherical kinematics representation (e.g.,
pneumatic actuators is described in terms of its curvature with Euler angles, quaternions). In words, the deformed axis of the
(16), a stretch field along the center line, and the bending plane robot is represented as a curve growing along a direction field
defined by an angle at the base. The balance of internal and the t(X, t) from the base to its tip, with a magnitude rate λ(X, t). In
external loads is then formulated, together with that of the cross order to obtain the complete description of the robot geometry,
section. The model is finally solved numerically, discretizing the the curve is equipped with a set of orthonormal frames and a roll
arm into a finite number of sections and driving the residual of distribution that describes the twist of the robot. At the end, the
the static balance to zero. robot geometry is defined by a reduced set of shape functions,
describing the backbone curve itself and a model of the twist.
In [94], a set of four independent shape functions Si (X, t) =
VI. GEOMETRICAL MODELS {λ(X, t) θ(X, t) φ(X, t) ψ(X, t)} is used, in which θ(X, t) and
φ(X, t) are the two angles defining t(X, t) and ψ(X, t) is the
The main characteristic of the approaches in this section is
roll distribution function. The inverse kinematic problem then
that they rely on the assumption that the deformed shape of
consists in finding the set of shape functions satisfying the task
the soft body resemble a specific geometrical shape. All these
constraints, which are usually represented by the end-effector
approaches are based on a representation of the soft body which
positioning. For this purpose, a modal approach is employed and
falls in the definition of a Cosserat rod. The main difference  NS
from the approaches described in Section III-B is that, when each Si (X, t) takes the form Si (X, t) = j=1i Φij (X)qij (t),
moving from the kinematics to the static and dynamic equations, where qij are modal participation factors and the Φij (X), NSi
these methods do not rely on the PDEs (7), but built from the modal functions, which are chosen by the robot programmer to
generalized coordinates specifically used to represent the body fulfill the task constraints. At the end, the robot shape is fully de-
geometry. We can then conclude that, while the configuration scribed by the modal factors qij , which represent the generalized
of these models can vary, being discrete or continuous, they all coordinates of the system q in (26). In [6], it has been proved
have their roots in some set of shape coordinates mapped to the that, when the curvature of the body is assumed to be constant,
configuration by this approach provides a transformation which is equivalent to
(27), presented in the following section. In one other work by
the same authors [7], variational approaches are presented in
q ∈ Rn → g(q) ∈ SE(3). (26) order to compute the optimal curve shapes that comply with both
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 11

joint and task constraints. The same kinematic model was also a survey on some of the main PCC approaches can be found
employed in [96] and extended to the dynamical modeling of in [6] and [8].
coiling continuum robots. Another possible function that can be 1) Kinematics-Based CC Models: Once the continuous body
used to represent the robot geometry consists in the pythagorean is represented as a finite set of CC segments, each of these can be
hodograph curves [97]. The backbone is represented using represented by a finite set of arc parameters and it is possible to
quadratic polynomials, which are functions of the chosen control obtain a map from them to the task space of the robot. Different
points, representing the generalized coordinates of the system. In parameters can be used to describe a CC segment, yielding
order to obtain the optimal quadratic polynomials to compute the to different kinematics maps. One of the most popular sets of
control points, the minimizing of the potential energy is applied. arc parameters that have been proposed consists of triplets of
Finally, an NN model is employed to predict the effects of curvature κ, the angle of the plane containing the arc φ, and arc
loads on the position of the robot, including the case of variable length l, which define q in (26). Different approaches have been
loads. In [98], the kinematics of a soft pneumatic actuator is proposed to obtain the kinematic map from these arc parameters.
modeled representing its backbone with cubic Hermite splines In [6], it has been proved that they all provide an identical
(also called cspline). In this way, the backbone is defined by transformation from the arc base to any point s ∈ [0, l] of the
two control points and two control orientations (vectors). The robot arm (noting cos = c and sin = s)
control points are obtained experimentally and an optimization ⎡ ⎤
procedure is carried out to fit the cspline with the given backbone. cφcκs −sφ cφsκs k1 cφ(1 − cκs)
⎢sφcκs cφ sφsκs 1 sφ(1 − cκs)⎥
To complete the description of the robot configuration, the ⎢ k ⎥
T (κ, φ, s) = ⎢ 1 ⎥.
orientation along the curve is finally obtained assuming the ⎣ −sκs 0 cκs κ sκs
⎦
minimizing of the torsion of the actuator’s backbone. In [99], 0 0 0 1
the backbone of tendon driven continuum robots is represented (27)
as a sequence of Euler arc splines, which are directed arcs One way to obtain (27) is based on the use of Denavit–
whose curvature varies in arithmetic progression. This article Hartenberg (D–H) parameters [101], [102], [103].
represent an example of a modeling technique that is based on In [101], under the CC assumptions, the continuous backbone
the CC assumption, as the models presented in Section VI-B, but of the robot is fitted with a virtual conventional rigid-link ma-
employs a functional representation of the robot’s kinematics. nipulator and modified D–H parameters are obtained to consider
In [100], elliptic Fourier descriptors are employed to describe the coupling imposed by the curvature in a continuum system,
soft deformable morphologies. Compared to the other methods providing the standard homogenous transformation matrix (27).
described in this section, this approach does not model the soft In a following work [102], a similar, improved, approach is pre-
robot through its backbone, but uses a closed curve to represent sented, finally providing the transformation from three parallel
the boundary of the 2-D regions that it occupies. This is obtained actuating tendons to the arc parameters:
thanks to elliptic Fourier descriptor, through a procedure that √ 
fits a closed curve to a set of 2-D points with arbitrary precision. l1 + l 2 + l 3 3 l2 + l3 − 2l1
−1
In particular, the image of the region occupied by the robot is l= , φ = tan
3 3 l2 − l 3
extracted from experimental recordings and its contour is iden-

tified through a discrete representation. Finally, the coefficients 2 l12 + l22 + l32 − l1 l2 − l1 l3 − l2 l3
of the Fourier series are computed, providing the description κ= (28)
d l1 + l 2 + l 3
of the shape.
where (l1 , l2 , l3 )T are actuator’s lengths, while d is the distance
from the center of the section to the actuator. This approach has
B. CC Models
been applied for the modeling of a great number of continuum
CC is often viewed as a desirable characteristic in contin- robots [104], [105] and [106], [107].
uum robots, due to the simplifications it enables in kinematic One of the main restrictions of these PCC models is that the
modeling as well as in real-time control and other useful com- used parametrization and kinematic maps can implicitly provide
putations. This is motivated by the fact that actuators with a numerical singularity that occurs when the curvature tends
a path parallel to the mid-line on a cylindrical manipulator to vanish (κ → 0), resulting in an infinite or undefined radius
produce a CC shape, in the absence of external forces. For these of curvature. In order to overcome this limitation, different
reasons, the CC assumption has been successfully applied in solutions have been introduced. In [108] and [109], the rotational
a great number of continuum robots modeling approaches. In and position components of the homogeneous transformation
these models, a soft body is represented by a finite number (27) are represented in modal form, similarly to the models
of circular arcs, each having a curvature that is constant in presented in Section VI-A. In particular, the entries in (27)
space. The coordinates q in (26) are specifically obtained to are numerically approximated using multivariate Taylor series
describe the circular arcs geometry. We can distinguish two expansions for actuator’s lengths variables at 0. Based on this
main groups of PCC approaches: the Kinematics-based models formulation, a Lagrangian approach was developed in [110]
are developed from a kinematical relation between the actuator and [111] for the spatial dynamics of a single section continuum
and the arcs parameters, while the mechanics-based models are arm and this was further extended to multisection arms in [112].
based on the mechanical description of the problem. Finally, Other approaches have been mostly focused on the definition of
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON ROBOTICS

a different parametrization and transformation map to describe the generalized coordinates q describing the configuration of the
the geometry of the soft bodies. One alternative is provided by arm are the curvature κ and an axial stretch measurement a .
the exponential map of [113], [114], and [115]. In particular, The static equilibrium leads to the description of the mechanical
(27) can be obtained using the exponential coordinates from response of the manipulator in the form Kq = B T u, where
the lie group theory seen in Section V-B. Considering that the B T represents the tendon moment arms and the axially directed
transformation for a circular arc is the composition of a rotation tangent. Relating the tendon tension and displacements is then
ζ rot with an in-plane ζ inp transformation, one has possible to obtain the mechanical response of the tendons in
 terms of their displacements y, y = C m u, where C m is a com-
T (κ, φ, s) = exp(ζ̂ rot φ)exp ζ̂ inp (κ)s . (29) pliance matrix, which is function of B T and of the manipulator
and tendon stiffness matrices and lengths. Finally, this yields to
In [114], this approach is applied to the modeling of concentric
tubes robot. It is assumed that the two tubes have the same y = A† q (30)
stiffness and are torsionally rigid, while applying torques to one where A = K −1 B T C −1 m is the forward kinematics transforma-
another. Given the arc assumption, these torques can be con- tion matrix. Equation (30) provides a mechanics-based relation-
sidered to be uniform. The in-plane bending model is obtained ship between the beam and the tendon configuration, analogous
employing EB linear equation and the resultant curvature of to transformations (28). In a following work [118], the 3-D static
two overlapping tubes is obtained through a force balance in model is also developed.
analogy with linear springs connected in parallel. Finally, the In [119], the statics of a braided pneumatic continuum ma-
mapping from the arc-parameters to the Cartesian poses of the nipulator is addressed. In this case, the effect of cross-sectional
tube cross-sections is obtained with (29). Similarly, in [116], the deformations on the arm deformation is studied using a strain
arcs are expressed with an axis-rotation parametrization. The energy function based on the Cauchy–Green stretch. The study
origin of the axis ω is positioned at point ρ on the horizontal shows that the PCC parametrization fails while releasing the
plane at the base of the section and ω is perpendicular to ρ. ρ PCC constraints gives better results. Another source of me-
is the radius of arc curvature, while ω = l/ρ is equal to chanical inconsistency in the PCC is related to the modeling
the angle between the proximal and distal positions when ρ is of torsion, which is absent from the model mainly concerned
used as the vertex of the angle. The CC arc is then parameterize with bending. In order to solve this issue, several developments
by (ρ, l) or equivalently (ω, l) which defines q. The forward of the PCC were proposed. In [120] and [121], the virtual
kinematics can, thus, be obtained through the exponential maps. power method is used to obtain the TDCRs robot dynamics.
Straightforward geometry finally provides the mapping between The considered robot geometry is that of [88], i.e., an elastic
the configuration parameters and the length of the actuating backbone connected by rigid disks and three tendon cables. The
tendons (28). In essence, this is a screw theory method, since disks are considered rigid, while the backbone and the cables
ω is exactly the axis of the screw motion required to cover the subsegment are treated as circular arcs. The relative pose of
arc bend, as shown in [27]. two adjacent discs is obtained from the two constant orthogonal
One other example of a parametrization that does not entail curvatures (β, γ) and an additional lumped torsional angle  of
the singularity in the proximity of a null curvature configuration the subsegment, which define the q-vector.
is provided in [117], where for each CC section composing The torsion model has become particularly critical in con-
the robot, the configuration of a segment is defined as linear tinuum robotics of CTR, where it is at origin of the so called
combinations of four arcs included in the section’s volume. “snapping” phenomenon. In [122], the authors provide first the
In particular, given the difference in length between opposite design conditions necessary for the validity of PCC approach,
arcs, Δx,i and Δy,i , through geometrical considerations, it is and conclude with a general model for CTRs. In this further
possible to obtain a transition map from the “conventional” arc model, the relative twist angle between two tubes is defined
parameters (κ, φ, l) and the new parametrization (Δx , Δy , δl), as a function of the arc length and then used to write the
where δl represents the change in the length of the section with moment equilibrium and to impose the compatibility equations,
respect to the at rest position. Algebraic steps are finally used to enforcing the coincidence of the tube centerlines. The torsional
obtain the explicit expression of the analogous of (27), in terms strain is then deduced from the Cosserat rod equilibrium (7).
of the new parameters. The resulting BVP was then solved by exploring the entire
2) Mechanics-Based CC Models: Some authors have sought input–output set. Based on this model, the stability analysis was
to draw mechanical consequences, by applying static laws to a performed in [123] and [124].
soft arm whose shape is described by the PCC. In [25], both
the forward and the inverse kinematics for a tendon driven
manipulator with parallel routines are derived in this way. More VII. DISCRETE MODELS
precisely, considering a single tendon that can experience only a In this section, we discuss the approaches where the configu-
constant tension u along its length, force, and moment balances ration of the system is discrete from the very beginning (i.e., the
are used to define the internal reaction forces acting on a section system is not discretized at the resolution level or through some
of the beam. This model is then extended to consider the presence assumptions on its backbone’s geometry). In particular, we can
of a finite number of tendons, based on an analogy with a system distinguish the following three main groups: the lumped-mass,
of linear springs acting in parallel. In the considered planar case, the pseudorigid, and the discrete rods models.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 13

perform parallel computing, resulting in good computational

performances.

B. Pseudorigid Models
The opportunity to exploit established rigid robotics theories
for soft robots motivated the family of models that are described
here. The soft bodies are represented as series of rigid links,
which are connected by revolute, universal, or spherical joints
[see Fig. 1(d)]. Thus, the configuration space of the soft body is
given in the form

Q ⊂ SE(3) × SE(3) × SE(3) · · · × SE(3). (32)

While these approaches can provide satisfactory results for the

modeling of hyper-redundant or snake-like robots, when dealing
Fig. 4. Lumped mass representation of a continuous rod. with continuous elastic structures they provide a low order of
spatial approximation accuracy, in addition to the expensive
identification procedure. In [128], a dexterous catheter manipu-
A. Lumped-Mass Models lator is represented as a series of pin joints connected by rigid
links. In [129], a pseudorigid 3-D approach is applied for the
One of the simplest approaches to modeling a continuous modeling of a steerable ablation catheter. The catheter is treated
body is to represent it as a set of discrete masses, dampers, and represented as four rigid links connected by three revolute joints
springs (see Fig. 4). The governing equations can be obtained and three torsional springs. Venkiteswaran et al. [130] presented
by energy methods or applying classical Newton equations to a pseudorigid model for continuum manipulators subject to
describe the propagation of the forces from one particle to the multiple external loads. In this case, a flexible manipulator
other in the configuration space is represented as four rigid links of given length, which are
Q ⊂ R3 × R3 × R3 · · · × R3 . (31) connected by three joints having 2DOF each. In [131], a flexible
beam is represented as a mass-less rigid body with a torsion
The main advantage of lumped-parameter mechanics models is spring attached at one end, providing the dynamical modeling of
that they are characterized by a simple structure, easy to adapt to compliant mechanisms. In [132], the modeling of hyperflexible
the capture of complex phenomena, such as nonlinear friction, manipulators is addressed using a serial rigid chain, where the
material hysteresis, and inertial dynamics. On the other hand, to number of kinematic DOF goes to infinity. The backbone curve is
reach the same fidelity of the continuum mechanics models or first described as a continuous curve with extended Frenet frames
FEM, they require a high number of DOFs and a data-expensive attached to each point along its length. For the numerical sim-
system identification procedure. ulations, the backbone curve is approximated by a serial chain
In [125], mass-spring-damper arrays are used to model the of rigid bodies. In [133], a steering catheter is represented as a
large deformations of a continuum surface (LDCS) actuated by combination of three sections: the virtual base of the distal shaft,
continuum arms. In particular, the 3-D problem of a system represented as a prismatic joint; the bending section of the distal
composed of two interconnected layers of LDCSs is considered, shaft, represented as two revolute joints, one prismatic joint and
where bending and shear deformations are allowed. The surface two revolute joints; the distal end of the catheter, which is treated
is modeled as a lattice where the masses are defined as the as a rigid body. In [134], a pseudorigid approach is proposed for
nodes of the system and they are connected through linear modeling an octopus inspired swimmer. In particular, each arm
spring and dampers. Assuming a uniformly stiff surface, the of the octopus is modeled as a kinematic chain of cylindrical
elastic constant of the springs and the damping coefficients are rigid segments, which are connected by planar rotatory joints.
assumed to be constant. Considering an isotropic linear elastic The first joints in the chain, connecting the arms to the main
material, the EoM for each mass is obtained by combining body, are modeled as actuated rotatory joints, while the other
direct application of the Newton’s second law and Hooke’s law. joints are modeled as (unactuated) rotatory linear spring and
In [126], a lumped-mass approach is presented for the modeling damper elements.
of tendon driven medical robotic catheters. The axial and the
bending stiffness of the model are adjusted through the selection
of the spring stiffness and their radial location, while the smooth C. Discrete Rods
motion of the catheter is modeled through the dampers. Because The computer graphics research community represents an
of their simplicity, lumped mass models are particularly suited important source of inspiration for soft robotics modeling. This
for large robotics simulations library, such as the one presented is the case of the approaches that are described here, which are
in [127]. Titan is a GPU-accelerated C++ software library for all based, with different extent, on the discrete elastic rod (DER)
the modeling of soft bodies and multi-agent robots at massive approach, originally introduced in [135], [136], and [137]. In this
scales. The main benefit of this approach is the possibility to formulation, the inextensible Kirchhoff rod assumptions (9) are
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON ROBOTICS

Fig. 5. Representation of a rod in the DER approach.

considered and, similarly to the Cosserat rod model, the config- Fig. 6. Scheme of an NN.
uration g of an elastic rod is given by a curve r(X), representing
the centerline, and a material frame, where the first axis is always
tangent to the curve. The Bishop (natural) frame {t, v, w} is then tree-like architecture, which are frequently encountered in the
introduced, providing, for a given centerline, the most relaxed soft robotics community. In [141], the DER-based formulation
frame, i.e., with zero twist. Finally, parallel transport allows us to is further enhanced to include frictional contact, inelastic col-
define the evolution of the Bishop frame along the mid-line. This lisions and inertial effects. In particular, considering a planar
model provides a simple parametrization of the cross-sectional case, a Rayleigh damping matrix is defined to formulate the
poses, given by the curve r(X) and an angle θ(X) between the internal damping forces vector, which is added to the external
Bishop and the cross-sectional frame. The rod is then discretized forces. In order to model the contact and the friction between
into a series of finite nodes (or vertices) connected by straight the robot and an unstructured ground, whose normal direction
segments (or edges). Each node is characterized by a position can vary with the horizontal axis, a Coulomb law is employed.
vector r i , while each segment is associated to the edge vector In [142], the DER is hybridized with a pseudorigid body model is
ei = r i+1 − r i and its tangent unit vector ti (see Fig. 5). In this used to represent compliant mechanisms. Based on this model,
way, the discrete curvature associated with the ith vertex takes the authors provided a computational design tool to perform
the form different optimization tasks related to soft robotics.
2 sin φi φi
κi = = 2 tan (33)
1 + cos φi 2 VIII. SURROGATE MODELS
where φi represents the turning angle between two consecutive With respect to the models presented so far, a complete
edges. Each ith vertex is also characterized by a total mass Mi , different approach to tackle the modeling of a soft robot consists
which is the average mass of the edges meeting in the vertex, in using large sets of data that are derived from various forms
while the mass moment of inertia for each edge can be obtained of external sources. These approaches are often referred to as
through volume integration. Finally, a discrete bending energy surrogate or data-driven models and a survey on some of them
and a discrete twist energy can be obtained in terms of the is presented in [9] and [10]. While one of the main benefits of
generalized coordinates (r i , θi , φi ). these solutions is that they do not require a physical model, on
Following the DER formulation, in [138], the three- the other hand they rely on large amounts of representative data
dimensional space of a soft filament is represented by a set of that are sometimes difficult to collect.
vertices r i (t) and a set of material frames Ri (t). Each vertex
is characterized by a linear velocity, a concentrated mass, and a A. Neural Networks
set of concentrated external forces. Extension, shear and axial NNs have been proved to be an effective tool to solve many
deformations are considered. Finally, through the spatial inte- kinds of nonlinear problems in different application fields, in-
gration, the discrete governing equations are obtained and they cluding robotics. As the name suggests, they are inspired from
are solved using a symplectic, second-order scheme. Similarly to the biological NNs that operate animal brains: the artificial
other approaches, the external physical interactions are included neurons, which represent the elementary units of NNs, can
in the external forces and moments vectors, while the internal transmit a signals to the other neurons. The signal is usually a real
physical effects (such as those modeling muscular activities) number and the output of each neuron is computed as a nonlinear
are added to the internal forces and moments resultants. In combination of all the inputs. Neurons are connected by edges
a recent work [139], this model has been implemented in an and they are all characterized by weights adjusted during the
open-source simulation environment, Elastica, for the modeling learning process. Usually, neurons are gathered in layers: the
and simulation of the dynamics of slender rods. first layer is also called input layer, the last one is called output
In [140], the planar case of the DER formulation is con- layer, and the intermediate layers are called hidden layers (see
sidered, where the rod is free to move on a plane, while the Fig. 6). NN results in a high dimensional set of nested functions
torsional deformations are neglected. Moreover, the notion of
folding for a straight rod is introduced to apply the approach to y = fM (AM , . . . f2 (A2 , f1 (A1 , x)) . . . ) (34)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 15

where x are the values of the input nodes, Ai are the edges forward dynamics of system (18) is formulated using only the
weights, fi are the activation functions, and y are the values task space variables and the direct mapping between the states
of the output nodes. In most soft robotics applications, the of the task space variables and the control inputs can be ob-
input and the output layers represent the actuation variables tained. Once the training is completed, an open-loop predictive
u and the shape parameters q. The learning process results in controller is developed through a trajectory optimization that is
an optimization over the network weights, which is performed carried out with an iterative sequential quadratic programming.
through back-propagation, a form of chain rule, where, after each In [150], the closed-loop controller is also implemented. The in-
forward step through the network, a backward pass is performed verse kinematic modeling of a bionic trunk is addressed in [151].
to compute the network Jacobian and to adjusts the model’s In particular, a learning phase is carried out considering a volume
weights accordingly. NNs have been used to approximate the of desired cartesian position for the robot’s tip, defining a finite
kinematics, the statics and also the dynamics description of set of target vertices. The inverse model is asked to estimate a
soft robots. Typically, Feed-forward neural network (FNN) are posture that allows us to move the effector to each vertex, and
employed for kinematics and static modeling. On the other hand, the training is carried out until the distance between the target
recurrent neural networks are usually preferred for dynamical and the actual position for each vertex is minimized.
modeling.
The first examples of the employment of NN algorithms B. Data-Driven Order Reduction
for the control of continuum robots were presented in [143].
Some modeling tools employ different forms of data-driven
In particular, a controller for continuum robots uses an FNN
order reduction to efficiently approximate the physical model.
component to compensate the dynamic uncertainties of the
In [152], a Koopman operator theory is employed for a data
system, in an attempt to reduce the uncertainty bound. In [144],
driven controller of soft robots. A dynamical system is repre-
the forward kinematic modeling of a bionic assistant trunk is
sented in an infinite function space F, which is composed of real-
obtained through multilayer perceptron (MLP) and radial basis
valued functions inside the state of the system domain [153]. The
function (RBF) NNs, which are both class of FNNs. The input
elements q ∈ F are called observables. The Koopman operators,
neurons propagate the input variables to the following layers,
denoted by Ut , are defined as the linear transformation Ut q =
while each neuron in the hidden layer is associated with an RBF
q ◦ Tt , where ◦ indicates the composition operation, while Tt
kernel (usually Gaussian). The data are obtained experimen-
is the flow (or dynamic) map of the system. In other words, the
tally, measuring the arm’s tip position at varying the actuation
Koopman operator lifts the dynamics of the system from the state
pressure, using a trilateration algorithm. In one other paper by
space to the space of the observables, describing the evolution
the same authors [145], the same handling assistant trunk is
of the observables q along the trajectories of the system. Its main
modeled, both with a data-driven and a pseudorigid modeling
advantage is that it provides a linear representation of the flow of
approach. With regards to the latter, the trunk is modeled as a
a nonlinear system, but in the infinite-dimensional space of the
series of rigid vertebrae connected by a total of four prismatic
observables. The discrete approximate of the Koopman operator
joints. For the data-driven approach, a modified Elman NN is
can be obtained from a set of experimentally measured state,
employed. FNNs are also employed in [146] and [147], which
given in the form of snapshot pairs. Some approaches have been
deal with the inverse kinetics of a cable driven soft manipulator.
focused on the model reduction of FEM that were presented
While the kinematical model describes the configuration of the
in Section IV-A. Model reduction methods are based on the
robot without considering the involved loads, the kinetics model
projection of the FE EoM to attractive subspaces of smaller
relates the motion of the robot with its actuation forces. The
dimensions. In this way, the size and the computational time of
direct kinetics model is obtained using a geometrically exact
the simulation are drastically lowered, allowing the application
model [60]. Thus, an FNN, taking the tip position as an input
of FE methods for control purposes. In [39], a snapshot proper or-
and giving the cable tension as an output, is employed. An
thogonal decomposition is employed to generate relevant bases
experimental data collection phase is carried out, using a set
to be used for the order reduction. More into details, the solution
of cable tensions and obtaining the tip position with an infrared
q of (19) is expressed as a truncated expansion of orthonormal
vision system. These data were used for the optimization and
vectors, which depend on the constraints λ. The orthonormal
training of the FNN. After the training phase, the performance
basis is then set to minimize the sum of all the errors that are
of the FNN is measured on test sets, using the output of the FNN
generated by the projection of the exact solution onto the basis.
as the input of the direct kinetics model. One other example of a
In [154], this approach is further extended for the development of
machine learning formulation for the global inverse kinematics
a low-order controller and observer, while in [40], it is applied for
of continuum manipulators was presented in [148]. The data
the reduction of self-collision contact forces. One other example
samples are generated by continuous motor babbling and a
of an FE order reduction is presented in [155] for modeling soft
single hidden-layer artificial NN is employed to learn directly
robots made of hyperelastic materials and actuated by cables or
the mapping (xi+1 , q i ) → q i+1 , where x is the pose of the end
tendons, with a special focus on contact problems.
effector. In [149], a dynamic model for open-loop control of soft
robotics manipulators is presented. The PCS model described in
Section V-B4 is employed to obtain the dynamics of a cable IX. SOFTWARE IMPLEMENTATION
driven robot operating underwater. Considering the case where An important aspect of the above models and their numerical
the robot and the task space have the same number of DOF, the resolution, is related to their software implementation. There
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

16 IEEE TRANSACTIONS ON ROBOTICS

TABLE I
SUMMARY OF THE MAIN MODELING FAMILIES, THEIR FUNDAMENTAL PROPERTIES, AND THEIR DESIGN APPLICATIONS

are several industrial as well as open-source soft robotic soft- the result of a compromise between realism (accuracy) and
ware/toolboxes available. Here, we briefly summarize the most computational efficiency, which is largely dictated by the user’s
popular ones. In Section IV-A, we introduced Sofa, one of the specific needs. For example, designers may be more interested
earliest open-source platforms for physics-based simulation. in realism in order to capture the concept of a new actuator in
Its SoftRobots plugin employs a multimodal description of a first phase, and in a second phase by computational efficiency
a problem, allowing the presence of several representations of forward (kinematics, statics, dynamics) algorithms that can
(mechanical, thermal, and visual) of the same object. This help optimization techniques requiring many simulations with
multimodal representation allows the simulation of scenarios different sets of design parameters. In the same way, closed-loop
involving the interaction of different components (rigid and/or control strategies can allow some relaxation of accuracy in favor
soft bodies, fluids). PyElastica is the Python implementation of of designing inverse algorithms (kinematics, statics, dynamics)
Elastica [139], described in Section VII-C. SoRoSim [75] is a capable of running online. Based on this tradeoff, here follows
MATLAB toolbox that uses the geometrically exact PVS model a critical analysis of the modeling approaches presented above.
described in Section V-B4. One other example of a MATLAB Surrogate models of Section VIII have the great advantage of
toolbox is TMTDyn, which employs discretized lumped systems being conceptually applicable to all systems, without physical
with reduced-order models [156]. SoMo (Soft Motion) [157] modeling effort, and can provide fast dedicated algorithms com-
couples the pseudorigid model discretization of Section VII-B patible with real-time control. However, for each robot design,
with a rigid body physics engine of Python to model soft robots. a huge database of measurements has to be built to feed the
ChainQueen [158] and incremental potential contact simulator learning process.
(IPC-Sim) [159] are simulators oriented toward computer graph- As regards physics-based models, the only approaches that are
ics rather than mechanical systems. ChainQueen is a Python physically consistent at all levels of modeling, (geometry, kine-
toolbox that uses the moving least square material point method, matics, and dynamics), are those based on continuum mechanics
a hybrid Eulerian/Lagrangian FEM, which uses both particles of Sections III–V, as well as the discrete rods method of Section
and grids to simulate soft bodies. IPC-Sim solves extreme non- VII-C. Potentially, they can be used to solve both direct problems
linear volumetric elastodynamic models using FEM. Finally, (simulation) and inverse problems (control). Like data-driven
some toolboxes are developed for specific applications, such as models of Section VIII, commercial FEM software have the
the DiffAqua [160], an optimization toolbox for soft underwater advantage of being applicable to complex systems of arbitrary
swimming robots. 3-D geometry in multiphysics contexts. For these reasons, they
are well suited to the design of new actuation principles or to the
study of specific complex phenomena. However, they are often
X. CONCLUSION
expensive in terms of computation time and poorly adapted to
Providing realistic models able to be operated computation- robotics specific control problems. To speed up the computa-
ally in real time or faster, to solve direct and inverse problems tions, the direct material discretizations in Sections VII-A and
of the wide variety of soft robots designed so far is an unattain- VII-B, may be appealing at first glance, since they allow any
able dream. Therefore, the choice of a modeling technique is continuous system to be replaced by a discrete one with lumped
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 17

parameters. However, although they can provide mechanically and from this point of view, it represents a promising perspective
consistent discrete formulations suited to fast simulation, they for soft robotics. Finally, Table I summarizes the main modeling
require complex identification methods to relate the lumped families in relation to their uses and robot designs.
parameters to the distributed original ones. In this regard, it is Soft robotics was developed in the last two decades and
worth mentioning that it is one of the historical motivations of the main efforts so far have been focuses on the simulation
the FEM to have overcame this “naive” discretization. With the of the robot’s static and dynamic behavior. Now the efforts
same goals in mind, Sofa has made great efforts in recent years of many researchers are mostly toward the extension of these
to revisit and optimize 3-D FEM for user-friendly modeling approaches for control [161], [162] and for their integration with
and rapid simulation of soft robots interacting with complex optimization tools [42], [163]. The main sources of inspiration
environments. However, due to the large number of nodal dof, all toward the achievements of these goals are more likely to come
these approaches remain difficult to exploit for control purposes, from other research disciplines. While going through the papers
and one of the challenges they face is to develop efficient reduc- of this survey, we tackled a great number of disciplines that
tion numerical methods compatible with the highly nonlinear represent their theoretical foundations: continuum and solid me-
character of soft robots (see Section VIII-B). When possible, a chanics, computational mechanics, machine learning, computer
natural reduction is to consider special geometries such as shells graphics, just to name a few. This interdisciplinary is probably
and rods. While the former remain little represented in robotics, what makes soft robotics so interesting, attracting scientists from
the latter are widely used to model multibody systems composed different research fields. On the other side, being the topic so
of rods and rigid bodies. In FEM, the use of beam-specific widespread, it is not easy to grasp it and this constitutes the main
finite elements then becomes natural, and the GE-FEM based motivation behind this manuscript. In this way, we were able to
on Cosserat beams of Section III-B, seems to be a good choice, recognize more clearly the uniqueness and the commonalities
as it is very efficient both in terms of accuracy and computational between the different techniques that have been presented so
cost. However, it still suffers from a lack of existing commercial far, in the effort to untangle such a vast research topic.
software dedicated to robotics and is difficult to implement from
scratch. REFERENCES
Based on the same models (based on 1-D directors), the other [1] B. Tondu and P. Lopez, “Modeling and control of Mckibben artificial mus-
methods in Section IV inherently enjoy similar accuracy to that cle robot actuators,” IEEE Control Syst. Mag., vol. 20, no. 2, pp. 15–38,
of GE-FEM, and represent reliable alternatives for soft robots Apr. 2000.
[2] C. Laschi, B. Mazzolai, and M. Cianchetti, “Soft robotics: Technologies
that can be described as beams (or a combination of beams). and systems pushing the boundaries of robot abilities,” Sci. Robot., vol. 1,
This includes robots with concentric tubes, continuous parallel no. 1, 2016, Art. no. eaah3690.
robots, or soft arms that do not involve large cross-sectional [3] D. Trivedi, C. D. Rahn, W. M. Kier, and I. D. Walker, “Soft robotics:
Biological inspiration, state of the art, and future research,” Appl. Bionics
inflation, but also systems with thread-like actuators, such as TD- Biomech., vol. 5, no. 3, pp. 99–117, 2008.
CRs. Moreover, unlike FEM, they were specifically developed [4] D. Rus and M. T. Tolley, “Design, fabrication and control of soft robots,”
for robotic purposes, whether for real-time simulation based on Nature, vol. 521, no. 7553, pp. 467–475, 2015.
[5] T. Thuruthel, Y. Ansari, E. Falotico, and C. Laschi, “Control strategies
the shooting method, or for the design of model-based control for soft robotic manipulators: A survey,” Soft Robot., vol. 5, no. 2,
laws. In the latter case, modal Ritz approaches as the GVS can pp. 149–163, 2018.
provide highly reduced dynamic models in the usual explicit La- [6] R. Webster and B. Jones, “Design and kinematic modeling of constant
curvature continuum robots: A review,” Int. J. Robot. Res., vol. 29, no. 13,
grangian form of rigid robots, and thus, facilitate the transfer of pp. 1661–1683, 2010.
methods from rigid to soft robotics. On the other hand, although [7] G. S. Chirikjian, “Conformational modeling of continuum structures in
computationally very efficient and ideally suited for kinematic robotics and structural biology: A review,” Adv. Robot., vol. 29, no. 3,
pp. 817–829, 2015.
control, the geometric approaches of Section VI-B. are based [8] A. Chawla, C. Frazelle, and I. Walker, “A comparison of constant
on certain kinematic simplifications that are not consistent with curvature forward kinematics for multisection continuum manipulators,”
continuum mechanics. In particular, PCC remains mechanically in Proc. IEEE Int. Conf. Robot. Comp., 2018, pp. 217–223.
[9] D. Kim et al., “Review of machine learning methods in soft robotics,”
consistent only for specific actuation designs and when gravity PLoS One, vol. 16, no. 2, pp. 1–24, 2021.
is negligible. Note that although widely used, this approach [10] X. Wang, Y. Li, and K. Kwok, “A survey for machine learning-based
being a subcase of the fully consistent GVS, it should be ad- control of continuum robots,” Front. Robot. AI, vol. 8, 2021.
[11] J. Burgner-Kahrs, D. C. Rucker, and H. Choset, “Continuum robots for
vantageously replaced by this alternative approach. Regarding medical applications: A survey,” IEEE Trans. Robot., vol. 31, no. 6,
the other simplified models of Section VI, while functional pp. 1261–1280, Dec. 2015.
approaches built around a reference curve, are conceptually [12] H. B. Gilbert, “On the mathematical modeling of slender biomedical
continuum robots,” Front. Robot. AI, vol. 8, 2021.
natural and can be computationally efficient, they often struggle [13] P. Rao, Q. Peyron, S. Lilge, and J. Burgner-Kahrs, “How to model tendon-
to capture the torsional state in a consistent way. Also note that driven continuum robots and benchmark modelling performance,” Front.
in contrast to modeling approaches based on a direct material Robot. AI, vol. 7, 2021.
[14] F. Bullo and A. Lewis, Geometric Control of Mechanical Systems (Texts
discretization, the discrete rod method is fully consistent since it in Applied Mathematics 49). Berlin, Germany: Springer-Verlag, 2004.
is based from the beginning on the discrete differential geometry [15] R. Murray, Z. Li, and S. Sastry, A Mathematical Introduction to Robotic
of the Cosserat theory (and from that point of view, should Manipulation. Boca Raton, FL, USA: CRC Press, 1994.
[16] P. Villaggio, Mathematical Models for Elastic Structures. Cambridge,
be classified into the directors based approaches of Section U.K.: Cambridge Univ. Press, 1997.
III-B). Moreover, it has been developed from the start for fast [17] R. Ogden, Non-Linear Elastic Deformations. New York, NY, USA: Dover
(interactive) simulation by the computer graphics community Publication, 1999.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

18 IEEE TRANSACTIONS ON ROBOTICS

[18] F. Renda, F. Giorgio-Serchi, F. Boyer, and C. Laschi, “Modelling [43] G. Maloisel, E. Knoop, C. Schumacher, and M. Bächer, “Automated
cephalopod-inspired pulsed-jet locomotion for underwater soft robots,” routing of muscle fibers for soft robots,” IEEE Trans. Robot., vol. 37,
Bioinspiration Biomimetics, vol. 10, no. 5, 2015, Art. no. 055005. no. 3, pp. 996–1008, Jun. 2021.
[19] F. Boyer and F. Renda, “Poincaré’s equations for cosserat media: Appli- [44] J. Tapia, E. Knoop, M. Mutný, M. A. Otaduy, and M. Bächer, “Make-
cation to shells,” J. Nonlinear Sci., vol. 27, pp. 1–44, 2017. sense: Automated sensor design for proprioceptive soft robots,” Soft
[20] F. Renda, F. Giorgio-Serchi, F. Boyer, C. Laschi, J. Dias, and L. Senevi- Robot., vol. 7, no. 3, pp. 332–345, 2020.
ratne, “A unified multi-soft-body dynamic model for underwater soft [45] F. Renda, F. Boyer, J. Dias, and L. Seneviratne, “Discrete cosserat ap-
robots,” Int. J. Robot. Res., vol. 37, no. 6, pp. 648–666, 2018. proach for multisection soft manipulator dynamics,” IEEE Trans. Robot.,
[21] E. Cosserat and F. Cosserat, Theorie Des Corps Deformables. Cambridge, vol. 34, no. 6, pp. 1518–1533, Dec. 2018.
U.K.: Cambridge Univ. Press, 1909. [46] J. Burgess, “Bending stiffness in a simulation of undersea cable deploy-
[22] F. Boyer, M. Porez, and W. Khalil, “Macro-continuous computed torque ment,” Int. J. Offshore Polar Eng., vol. 3, no. 3, pp. 197–204, 1993.
algorithm for a three-dimensional EEL-like robot,” IEEE Trans. Robot., [47] A. A. Tjavaras, Q. Zhu, Y. Liu, M. S. Triantafyllou, and D. K. P. Yue,
vol. 22, no. 4, pp. 763–775, Aug. 2006. “The mechanics of highly-extensible cables,” J. Sound Vib., vol. 213,
[23] F. Boyer, S. Ali, and M. Porez, “Macrocontinuous dynamics for hyper- no. 4, pp. 709–737, 1998.
redundant robots: Application to kinematic locomotion bioinspired by [48] Q. Peyron, Q. Boehler, K. Rabenorosoa, P. Renaud, and N. Andreff,
elongated body animals,” IEEE Trans. Robot., vol. 28, no. 2, pp. 303–317, “Kinematic analysis of magnetic continuum robots using continuation
Apr. 2012. method and bifurcation analysis,” IEEE Robot. Autom. Lett., vol. 3, no. 4,
[24] F. Renda, M. Giorelli, M. Calisti, M. Cianchetti, and C. Laschi, “Dynamic pp. 3646–3653, Oct. 2018.
model of a multibending soft robot arm driven by cables,” IEEE Trans. [49] S. Briot and A. Goldsztejn, “Singularity conditions for continuum parallel
Robot., vol. 30, no. 5, pp. 1109–1122, Oct. 2014. robots,” IEEE Trans. Robot., vol. 38, no. 1, pp. 507–525, Feb. 2021.
[25] D. B. Camarillo, C. F. Milne, C. R. Carlson, M. R. Zinn, and [50] D. Trivedi, A. Lotfi, and C. D. Rahn, “Geometrically exact models for soft
J. K. Salisbury, “Mechanics modeling of tendon-driven continuum robotic manipulators,” IEEE Trans. Robot., vol. 24, no. 4, pp. 773–780,
manipulators,” IEEE Trans. Robot., vol. 24, no. 6, pp. 1262–1273, Aug. 2008.
Dec. 2008. [51] Y. Sun, “Modelling and simulation of low-tension oceanic cable/body
[26] D. Rucker and R. Webster, “Statics and dynamics of continuum robots deployment,” Ph.D. dissertation, Univ. Connecticut, Stamford, CT, USA,
with general tendon routing and external loading,” IEEE Trans. Robot., 1996.
vol. 27, no. 6, pp. 1033–1044, Dec. 2011. [52] J. Till, V. Aloi, and D. Rucker, “Real-time dynamics of soft and continuum
[27] F. Renda, M. Cianchetti, H. Abidi, J. Dias, and L. Seneviratne, “Screw- robots based on cosserat rod models,” Int. J. Robot. Res., vol. 38, no. 6,
based modeling of soft manipulators with tendon and fluidic actuation,” pp. 723–746, 2019.
J. Mech. Robot., vol. 9, 2017, Art. no. 041012. [53] C. Gatti-Bono and N. Perkins, “Physical numerical modelling of the
[28] F. Boyer, V. Lebastard, F. Candelier, and F. Renda, “Dynamics of contin- dynamic behavior of fly line,” J. Sound Vib., vol. 255, no. 3, pp. 555–577,
uum and soft robots: A strain parameterization based approach,” IEEE 2002.
Trans. Robot., vol. 37, no. 3, pp. 847–863, Jun. 2021. [54] F. Janabi-Sharifi, A. Jalali, and I. Walker, “Cosserat rod-based dynamic
[29] E. Reissner, “On one-dimensional finite-strain beam theory: The plane modeling of tendon-driven continuum robots: A tutorial,” IEEE Access,
problem,” J. Appl. Math. Phys., vol. 23, no. 5, pp. 795–804, 1972. vol. 9, pp. 68703–68719, 2021.
[30] A. A. Shabana, “Continuum-based geometry/analysis approach for flex- [55] C. B. Black, J. Till, and D. C. Rucker, “Parallel continuum robots:
ible and soft robotic systems,” Soft Robot., vol. 5, no. 5, pp. 613–621, Modeling, analysis, and actuation-based force sensing,” IEEE Trans.
2018. Robot., vol. 34, no. 1, pp. 29–47, Feb. 2018.
[31] M. Xavier, A. Fleming, and Y. Yong, “Finite element modeling of soft [56] D. Rucker, R. Webster, G. Chirikjian, and N. Cowan, “Equilibrium
fluidic actuators: Overview and recent developments,” Adv. Intell. Syst., conformations of concentric-tube continuum robots,” Int. J. Robot. Res.,
vol. 3, no. 2, 2021, Art. no. 2000187. vol. 29, no. 10, pp. 1263–1280, 2010.
[32] P. Moseley, J. M. Florez, H. A. Sonar, G. Agarwal, W. Curtin, and [57] J. Till et al., “A dynamic model for concentric tube robots,” IEEE Trans.
J. Paik, “Modeling, design, and development of soft pneumatic actuators Robot., vol. 36, no. 6, pp. 1704–1718, Dec. 2020.
with finite element method,” Adv. Eng. Mater., vol. 18, no. 6, pp. 978–988, [58] R. Featherstone, Rigid Body Dynamics Algorithms. New York, NY, USA:
2016. Springer2008.
[33] C. Tawk and G. Alici, “Finite element modeling in the design process of [59] F. Boyer, M. Porez, A. Leroyer, and M. Visonneau, “Fast dynamics of
3D printed pneumatic soft actuators and sensors,” Robotics, vol. 9, no. 3, an EEL-like robot-comparisons with Navier-Stokes simulations,” IEEE
2020, Art. no. 52. Trans. Robot., vol. 24, no. 6, pp. 1274–1288, Dec. 2008.
[34] J. Allard et al., “Sofa—An open source framework for medical simula- [60] F. Renda, M. Cianchetti, M. Giorelli, A. Arienti, and C. Laschi, “A
tion,” MMVR 15-Med. Meets Virtual Reality, vol. 125, pp. 13–18, 2007. 3D steady-state model of a tendon-driven continuum soft manipulator
[35] F. Faure et al., “Sofa: A multi-model framework for interactive physi- inspired by the octopus arm,” Bioinspiration Biomimetics, vol. 7, no. 2,
cal simulation,” in Studies in Mechanobiology, Tissue Engineering and 2012, Art. no. 025006.
Biomaterials. Berlin, Germany: Springer, 2012. [61] A. L. Orekhov and N. Simaan, “Solving cosserat rod models via colloca-
[36] E. Coevoet et al., “Software toolkit for modeling, simulation, and control tion and the magnus expansion,” in Proc. IEEE Int. Conf. Intell. Robots
of soft robots,” Adv. Robot., vol. 31, no. 22, pp. 1208–1224, 2017. Syst., 2020, pp. 8653–8660.
[37] G. Zheng, O. Goury, M. Thieffry, A. Kruszewski, and C. Duriez, “Con- [62] J. Simo and L. Vu-Quoc, “On the dynamics in space of rods undergo-
trollability pre-verification of silicone soft robots based on finite-element ing large motions—A geometrically exact approach,” Comput. Methods
method,” in Proc. IEEE Int. Conf. Robot. Autom., 2019, pp. 7395–7400. Appl. Mech. Eng., vol. 66, no. 2, pp. 125–161, 1988.
[38] C. Duriez, “Control of elastic soft robots based on real-time finite [63] J. Simo and D. Fox, “On a stress resultant geometrically exact shell model.
element method,” in Proc. IEEE Int. Conf. Robot. Autom., 2013, Part I: Formulation and optimal parametrization,” Comput. Methods Appl.
pp. 3982–3987. Mech. Eng., vol. 72, no. 3, pp. 267–304, 1989.
[39] O. Goury and C. Duriez, “Fast, generic, and reliable control and simu- [64] A. Cardona and M. Geradin, “A beam finite element non-linear theory
lation of soft robots using model order reduction,” IEEE Trans. Robot., with finite rotations,” Int. J. Numer. Methods Eng., vol. 26, no. 11,
vol. 34, no. 6, pp. 1565–1576, Dec. 2018. pp. 2403–2438, 1988.
[40] O. Goury, B. Carrez, and C. Duriez, “Real-time simulation for control [65] F. Boyer and D. Primault, “Finite element of slender beams in finite
of soft robots with self-collisions using model order reduction for con- transformations: A geometrically exact approach,” Int. J. Numer. Meth-
tact forces,” IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 3752–3759, ods. Eng., vol. 59, no. 5, pp. 669–702, 2004.
Apr. 2021. [66] V. Sonneville, A. Cardona, and O. Brüls, “Geometrically exact beam
[41] Y. Adagolodjo, F. Renda, and C. Duriez, “Coupling numerical deformable finite element formulated on the special euclidean group SE(3),” Comput.
models in global and reduced coordinates for the simulation of the direct Methods Appl. Mech. Eng., vol. 268, pp. 451–474, 2014.
and the inverse kinematics of soft robots,” IEEE Robot. Autom. Lett., [67] “ LMS samtech samcef Mecano,”. [Online]. Available: https://www.
vol. 6, no. 2, pp. 3910–3917, Apr. 2021. siemens.com/plm/lms
[42] M. Bächer, E. Knoop, and C. Schumacher, “Design and control of soft [68] A. Shabana, “Definition of the slopes and the finite element absolute nodal
robots using differentiable simulation,” Curr. Robot. Rep., vol. 2, no. 2, coordinate formulation,” Multibody Syst. Dyn., vol. 1, no. 3, pp. 339–348,
pp. 211–221, 2021. 1997.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 19

[69] C. Meier, A. Popp, and W. A. Wall, “Geometrically exact finite el- [94] G. S. Chirikjian and J. W. Burdick, “A modal approach to hyper-
ement formulations for slender beams: Kirchhoff–love theory versus redundant manipulator kinematics,” IEEE Trans. Robot. Autom., vol. 10,
Simo–Reissner theory,” Arch. Comput. Methods Eng., vol. 26, no. 1, no. 3, pp. 343–354, Jun. 1994.
pp. 163–243, 2019. [95] G. S. Chirikjian, “Hyper-redundant manipulator dynamics: A continuum
[70] İ. Tunay, “Spatial continuum models of rods undergoing large defor- approximation,” Adv. Robot., vol. 9, no. 3, pp. 217–243, 1994.
mation and inflation,” IEEE Trans. Robot., vol. 29, no. 2, pp. 297–307, [96] M. Ivanescu, M. C. Florescu, N. Popescu, and D. Popescu, “Coil function
Apr. 2013. control problem for a hyperredundant robot,” in Proc. IEEE/ASME Int.
[71] S. M. H. Sadati, S. E. Naghibi, I. D. Walker, K. Althoefer, and Conf. Adv. Intell. Mechatron., 2007, pp. 1–6.
T. Nanayakkara, “Control space reduction and real-time accurate model- [97] I. Singh, Y. Amara, A. Melingui, P. Mani Pathak, and R. Merzouki,
ing of continuum manipulators using Ritz and Ritz–Galerkin methods,” “Modeling of continuum manipulators using pythagorean hodograph
IEEE Robot. Autom. Lett., vol. 3, no. 1, pp. 328–335, Jan. 2018. curves,” Soft Robot., vol. 5, no. 4, pp. 425–442, 2018.
[72] P. Češarek, M. Saje, and D. Zupan, “Dynamics of flexible beams: Finite- [98] M. Wiese, K. Rüstmann, and A. Raatz, “Kinematic modeling of a soft
element formulation based on interpolation of strain measures,” Finite pneumatic actuator using cubic hermite splines,” in Proc. IEEE Int. Conf.
Elem. Anal. Des., vol. 72, pp. 47–63, 2013. Intell. Robots Syst., 2019, pp. 7176–7182.
[73] S. Grazioso, G. Di Gironimo, and B. Siciliano, “A geometrically exact [99] P. Rao, Q. Peyron, and J. Burgner-Kahrs, “Shape representation and
model for soft continuum robots: The finite element deformation space modeling of tendon-driven continuum robots using Euler arc splines,”
formulation,” Soft Robot., vol. 6, no. 6, pp. 790–811, 2019. IEEE Robot. Autom. Lett., vol. 7, no. 3, pp. 8114–8121, Jul. 2022.
[74] F. Renda, C. Armanini, V. Lebastard, F. Candelier, and F. Boyer, “A geo- [100] K. M. Digumarti, B. Trimmer, A. T. Conn, and J. Rossiter, “Quantify-
metric variable-strain approach for static modeling of soft manipulators ing dynamic shapes in soft morphologies,” Soft Robot., vol. 6, no. 6,
with tendon and fluidic actuation,” IEEE Robot. Autom. Lett., vol. 5, no. 3, pp. 733–744, 2019.
pp. 4006–4013, Jul. 2020. [101] I. Gravagne, C. Rahn, and I. Walker, “Large deflection dynamics and
[75] A. T. Mathew, I. M. B. Hmida, C. Armanini, F. Boyer, and F. Renda, control for planar continuum robots,” IEEE/ASME Trans. Mechatron.,
“SoRoSim: A MATLAB toolbox for hybrid rigid-soft robots based on vol. 8, no. 2, pp. 299–307, Jun. 2003.
the geometric variable-strain approach,” IEEE Robot. Autom. Mag., to be [102] B. Jones and I. Walker, “Kinematics for multisection continuum robots,”
published, doi: :10.1109/MRA.2022.3202488. IEEE Trans. Robot., vol. 22, no. 1, pp. 43–55, Feb. 2006.
[76] F. Renda, C. Messer, C. Rucker, and F. Boyer, “A sliding-rod variable- [103] I. Walker, “Continuous backbone “continuum” robot manipulators,” Int.
strain model for concentric tube robots,” IEEE Robot. Autom. Lett., vol. 6, Scholarly Res. Notices, vol. 2013, 2013.
no. 2, pp. 3451–3458, Apr. 2021. [104] C. Escande, T. Chettibi, R. Merzouki, V. Coelen, and P. M. Pathak, “Kine-
[77] F. Boyer, V. Lebastard, F. Candelier, and F. Renda, “Extended Hamilton’s matic calibration of a multisection bionic manipulator,” IEEE/ASME
principle applied to geometrically exact Kirchhoff sliding rods,” J. Sound Trans. Mechatron., vol. 20, no. 2, pp. 663–674, Apr. 2015.
Vib., vol. 516, 2022, Art. no. 116511. [105] O. Lakhal, A. Melingui, and R. Merzouki, “Hybrid approach for modeling
[78] G. F. Carrier, “The spaghetti problem,” Amer. Math. Monthly, vol. 56, and solving of kinematics of a compact bionic handling assistant manip-
no. 1, pp. 669–672, 1949. ulator,” IEEE/ASME Trans. Mechatron., vol. 21, no. 3, pp. 1326–1335,
[79] F. Renda and L. Seneviratne, “A geometric and unified approach for Jun. 2016.
modeling soft-rigid multi-body systems with lumped and distributed [106] T. Mahl, A. Hildebrandt, and O. Sawodny, “A variable curvature contin-
degrees of freedom,” in Proc. IEEE Int. Conf. Robot. Autom., 2018, uum kinematics for kinematic control of the bionic handling assistant,”
pp. 1567–1574. IEEE Trans. Robot., vol. 30, no. 4, pp. 935–949, Aug. 2014.
[80] C. Armanini, I. Hussain, M. Z. Iqbal, D. Gan, D. Prattichizzo, and [107] V. Falkenhahn, T. Mahl, A. Hildebrandt, R. Neumann, and O. Sawodny,
F. Renda, “Discrete cosserat approach for closed-chain soft robots: “Dynamic modeling of bellows-actuated continuum robots using the
Application to the fin-ray finger,” IEEE Trans. Robot., vol. 37, no. 6, Euler–Lagrange formalism,” IEEE Trans. Robot., vol. 31, no. 6,
pp. 2083–2098, Dec. 2021. pp. 1483–1496, Dec. 2015.
[81] B. Caasenbrood, A. Pogromsky, and H. Nijmeijer, “Control-oriented [108] I. S. Godage, E. Guglielmino, D. T. Branson, G. A. Medrano-Cerda, and
models for hyperelastic soft robots through differential geometry of D. G. Caldwell, “Novel modal approach for kinematics of multisection
curves,” Soft Robot.,2022. continuum arms,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2011,
[82] C. D. Santina and D. Rus, “Control oriented modeling of soft robots: pp. 1093–1098.
The polynomial curvature case,” IEEE Robot. Autom. Lett., vol. 5, no. 2, [109] I. Godage, G. Medrano-Cerda, D. Branson, E. Guglielmino, and
pp. 290–298, Apr. 2020. D. Caldwell, “Modal kinematics for multisection continuum arms,”
[83] C. D. Santina, “The soft inverted pendulum with affine curvature,” in Bioinspiration Biomimetics, vol. 10, no. 2, 2015, Art. no. 035002.
Proc. IEEE 59th Conf. Decis. Control, 2020, pp. 4135–4142. [110] I. S. Godage, D. T. Branson, E. Guglielmino, G. A. Medrano-Cerda, and
[84] L. U. Odhner and A. M. Dollar, “The smooth curvature model: An D. G. Caldwell, “Dynamics for biomimetic continuum arms: A
efficient representation of Euler–Bernoulli flexures as robot joints,” IEEE modal approach,” in Proc. IEEE Int. Conf. Robot. Biomim., 2011,
Trans. Robot., vol. 28, no. 4, pp. 761–772, Aug. 2012. pp. 104–109.
[85] D. Bigoni, Nonlinear Solid Mechanics: Bifurcation Theory and Material [111] I. Godage, D. Branson, E. Guglielmino, G. Medrano-Cerda, and
Instability. Cambridge, U.K.: Cambridge Univ. Press, 2012. D. Caldwell, “Shape function-based kinematics and dynamics for vari-
[86] C. Armanini, F. Dal Corso, D. Misseroni, and D. Bigoni, “From the elas- able length continuum robotic arms,” in Proc. IEEE Int. Conf. Robot.
tica compass to the elastica catapult: An essay on the mechanics of soft Autom., 2011, pp. 452–457.
robot arm,” Proc. Roy. Soc. A, vol. 473, no. 2198, 2017, Art. no. 20160870. [112] I. Godage, G. Medrano-Cerda, D. Branson, E. Guglielmino, and
[87] K. Xu and N. Simaan, “Analytic formulation for kinematics, statics D. Caldwell, “Dynamics for variable length multisection continuum
and shape restoration of multi-backbone continuum robots via elliptic arms,” Int. J. Robot. Res., vol. 35, no. 6, pp. 695–722, 2016.
integrals,” ASME J. Mech. Robot., vol. 2, 2009, Art. no. 11006. [113] R. Webster, J. Kim, N. Cowan, G. Chirikjian, and A. Okamura, “Nonholo-
[88] K. Xu and N. Simaan, “Intrinsic wrench estimation and its performance nomic modeling of needle steering,” Int. J. Robot. Res., vol. 25, no. 5/6,
index for multisegment continuum robots,” IEEE Trans. Robot., vol. 26, pp. 509–525, 2006.
no. 3, pp. 555–561, Jun. 2010. [114] R. Webster, J. Romano, and N. Cowan, “Mechanics of precurved-tube
[89] P. Polygerinos et al., “Modeling of soft fiber-reinforced bending actua- continuum robots,” IEEE Trans. Robot., vol. 25, no. 1, pp. 67–78,
tors,” IEEE Trans. Robot., vol. 31, no. 3, pp. 778–789, Jun. 2015. Feb. 2009.
[90] V. Cacucciolo, F. Renda, E. Poccia, C. Laschi, and M. Cianchetti, [115] P. Sears and P. Dupont, “A steerable needle technology using curved
“Modelling the nonlinear response of fibre-reinforced bending fluidic concentric tubes,” in Proc. IEEE Int. Conf. Intell. Robots Syst., 2006,
actuators,” Smart Mater. Struct., vol. 25, no. 10, 2016, Art. no. 105020. pp. 2850–2856.
[91] G. Olson, R. Hatton, J. Adams, and Y. Mengüç, “An Euler–Bernoulli [116] T. F. Allen, L. Rupert, T. R. Duggan, G. Hein, and K. Albert, “Closed-form
beam model for soft robot arms bent through self-stress and external non-singular constant-curvature continuum manipulator kinematics,” in
loads,” Int. J. Solids. Struct., vol. 207, pp. 113–131, 2020. Proc. IEEE Int. Conf. Soft Robot., 2020, pp. 410–416.
[92] S. Hirose, “Biologically inspired robots: Snake-like locomotors and [117] C. Della Santina, A. Bicchi, and D. Rus, “On an improved state
manipulators,” Robotica, vol. 12, no. 3, pp. 282–282, 1994. parametrization for soft robots with piecewise constant curvature and
[93] S. Hirose and H. Yamada, “Snake-like robots,” IEEE Robot. Autom. Mag., its use in model based control,” IEEE Robot. Autom. Lett., vol. 5, no. 2,
vol. 16, no. 1, pp. 88–98, Mar. 2009. pp. 1001–1008, Apr. 2020.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

20 IEEE TRANSACTIONS ON ROBOTICS

[118] D. B. Camarillo, C. R. Carlson, and J. K. Salisbury, “Configuration [143] D. Braganza, D. M. Dawson, I. D. Walker, and N. Nath, “A neural network
tracking for continuum manipulators with coupled tendon drive,” IEEE controller for continuum robots,” IEEE Trans. Robot., vol. 23, no. 6,
Trans. Robot., vol. 25, no. 4, pp. 798–808, Aug. 2009. pp. 1270–1277, Dec. 2007.
[119] S. M. H. Sadati et al., “A geometry deformation model for braided [144] A. Melingui, C. Escande, N. Benoudjit, R. Merzouki, and J. Mbede,
continuum manipulators,” Front. Robot. AI, vol. 4, 2017. “Qualitative approach for forward kinematic modeling of a compact
[120] W. Rone and P. Ben-Tzvi, “Continuum robot dynamics utilizing the prin- bionic handling assistant trunk,” IFAC Proc. Volumes, vol. 47, no. 3,
ciple of virtual power,” IEEE Trans. Robot., vol. 30, no. 1, pp. 275–287, pp. 9353–9358, 2014.
Feb. 2014. [145] A. Melingui, O. Lakhal, B. Daachi, J. B. Mbede, and R. Mer-
[121] W. S. Rone and P. Ben-Tzvi, “Mechanics modeling of multisegment zouki, “Adaptive neural network control of a compact bionic handling
rod-driven continuum robots,” J. Mech. Robot., vol. 6, no. 4, 2014, arm,” IEEE/ASME Trans. Mechatron., vol. 20, no. 6, pp. 2862–2875,
Art. no. 041006. Dec. 2015.
[122] P. E. Dupont, J. Lock, B. Itkowitz, and E. Butler, “Design and control of [146] M. Giorelli, F. Renda, M. Calisti, A. Arienti, G. Ferri, and C. Laschi,
concentric-tube robots,” IEEE Trans. Robot., vol. 26, no. 2, pp. 209–225, “Neural network and jacobian method for solving the inverse statics of a
Apr. 2010. cable-driven soft arm with nonconstant curvature,” IEEE Trans. Robot.,
[123] R. Xu, S. F. Atashzar, and R. V. Patel, “Kinematic instability in vol. 31, no. 4, pp. 823–834, Aug. 2015.
concentric-tube robots: Modeling and analysis,” in Proc. IEEE Int. Conf. [147] M. Giorelli, F. Renda, M. Calisti, A. Arienti, G. Ferri, and C. Laschi,
Biomed. Robot. Biomechatron., 2014, pp. 163–168. “Learning the inverse kinetics of an octopus-like manipulator in three-
[124] R. Hendrick, H. B. Gilbert, and R. J. Webster, “Designing snap-free dimensional space,” Bioinspiration Biomimetics, vol. 10, no. 3, 2015,
concentric tube robots: A local bifurcation approach,” in Proc. IEEE Int. Art. no. 035006.
Conf. Robot. Autom., 2015, pp. 2256–2263. [148] T. Thuruthel, E. Falotico, M. Manti, A. Pratesi, M. Cianchetti, and
[125] H. Habibi, C. Yang, I. S. Godage, R. Kang, I. D. Walker, and D. T. C. Laschi, “Learning closed loop kinematic controllers for continuum
Branson, “A Lumped-Mass Model for Large Deformation Continuum manipulators in unstructured environments,” Soft Robot., vol. 4, no. 3,
Surfaces Actuated by Continuum Robotic Arms,” J. Mech. Robot., pp. 285–296, 2017.
vol. 12, no. 1, 2019, Art. no. 011014. [149] T. Thuruthel, E. Falotico, F. Renda, and C. Laschi, “Learning dy-
[126] J. Jung, R. S. Penning, N. J. Ferrier, and M. R. Zinn, “A modeling namic models for open loop predictive control of soft robotic
approach for continuum robotic manipulators: Effects of nonlinear in- manipulators,” Bioinspiration Biomimetics, vol. 12, no. 6, 2017,
ternal device friction,” in Proc. IEEE Int. Conf. Intell. Robots Syst., 2011, Art. no. 066003.
pp. 5139–5146. [150] T. G. Thuruthel, E. Falotico, F. Renda, and C. Laschi, “Model-
[127] J. Austin, R. Corrales-Fatou, S. Wyetzner, and H. Lipson, “Titan: A based reinforcement learning for closed-loop dynamic control of soft
parallel asynchronous library for multi-agent and soft-body robotics robotic manipulators,” IEEE Trans. Robot., vol. 35, no. 1, pp. 124–134,
using Nvidia Cuda,” in Proc. IEEE Int. Conf. Robot. Autom., 2020, Feb. 2019.
pp. 7754–7760. [151] M. Rolf and J. J. Steil, “Efficient exploratory learning of inverse kinemat-
[128] M. Kutzer, S. Segreti, C. Brown, M. Armand, R. H. Taylor, and ics on a bionic elephant trunk,” IEEE Trans. Neural. Netw. Learn. Syst.,
S. Mears, “Design of a new cable-driven manipulator with a large open vol. 25, no. 6, pp. 1147–1160, Jun. 2014.
lumen: Preliminary applications in the minimally-invasive removal of [152] D. Bruder, X. Fu, R. B. Gillespie, C. D. Remy, and R.
osteolysis,” in Proc. IEEE Int. Conf. Robot. Autom., 2011, pp. 2913–2920. Vasudevan, “Data-driven control of soft robots using Koopman op-
[129] M. Khoshnam and R. V. Patel, “A pseudo-rigid-body 3R model for a erator theory,” IEEE Trans. Robot., vol. 37, no. 3, pp. 948–961,
steerable ablation catheter,” in Proc. IEEE Int. Conf. Robot. Autom., 2013, Jun. 2021.
pp. 4427–4432. [153] H. Arbabi and I. Mezic, “Ergodic theory, dynamic mode decomposition,
[130] V. Venkiteswaran, J. Sikorski, and S. Misra, “Shape and contact force and computation of spectral properties of the Koopman operator,” SIAM
estimation of continuum manipulators using pseudo rigid body models,” J. Appl. Dyn. Syst., vol. 16, no. 4, pp. 2096–2126, 2017.
Mech. Mach. Theory, vol. 139, pp. 34–45, 2019. [154] M. Thieffry, A. Kruszewski, C. Duriez, and T. Guerra, “Control design
[131] Y. Yu, L. L. Howell, C. Lusk, Y. Yue, and M. He, “Dynamic modeling of for soft robots based on reduced-order model,” IEEE Robot. Autom. Lett.,
compliant mechanisms based on the pseudo-rigid-body model,” J. Mech. vol. 4, no. 1, pp. 25–32, Jan. 2019.
Des., vol. 127, no. 4, pp. 760–765, 2005. [155] J. Chenevier, D. González, J. V. Aguado, F. Chinesta, and E. Cueto,
[132] H. Mochiyama and T. Suzuki, “Dynamical modelling of a hyper-flexible “Reduced-order modeling of soft robots,” PLoS One, vol. 13, no. 2,
manipulator,” in Proc. SICE Annu. Conf., 2002, vol. 3, pp. 1505–1510. pp. 1–15, 2018.
[133] Y. Ganji and F. Janabi-Sharifi, “Catheter kinematics for intracardiac nav- [156] S. Sadati et al., “TMTDyn: A MATLAB package for modeling and
igation,” IEEE Trans Biomed Eng., vol. 56, no. 3, pp. 621–32, Mar. 2009. control of hybrid rigid–continuum robots based on discretized lumped
[134] M. Sfakiotakis, A. Kazakidi, A. Chatzidaki, T. Evdaimon, and D. P. systems and reduced-order models,” Int. J. Robot. Res., vol. 40, no. 1,
Tsakiris, “Multi-arm robotic swimming with octopus-inspired compliant pp. 296–347, 2021.
web,” in Proc. IEEE Int. Conf. Intell. Robots Syst., 2014, pp. 302–308. [157] M. A. Graule, C. B. Teeple, T. P. McCarthy, G. R. Kim, R. C. St Louis, and
[135] M. Bergou, M. Wardetzky, S. Robinson, E. Audoly, and B. Grinspun, R. J. Wood, “Somo: Fast and accurate simulations of continuum robots
“Discrete elastic rods,” ACM Trans. Graph., vol. 27, no. 3, pp. 1–12, in complex environments,” in Proc. IEEE Int. Conf. Intell. Robots Syst.,
2008. 2021, pp. 3934–3941.
[136] M. Bergou, B. Audoly, E. Vouga, M. Wardetzky, and E. Grinspun, [158] Y. Hu et al., “Chainqueen: A real-time differentiable physical simula-
“Discrete viscous threads,” ACM Trans. Graph., vol. 29, no. 4, pp. 1–10 tor for soft robotics,” in Proc. IEEE Int. Conf. Robot. Autom., 2019,
2010. pp. 6265–6271.
[137] B. Audoly, N. Clauvelin, P.-T. Brun, M. Bergou, E. Grinspun, and [159] M. Li et al., “Incremental potential contact: Intersection-and inversion-
M. Wardetzky, “A discrete geometric approach for simulating the dy- free, large-deformation dynamics,” ACM Trans. Graph., vol. 39, no. 4,
namics of thin viscous threads,” J. Comput. Phys., vol. 253, pp. 18–49, 2020.
2013. [160] P. Ma et al., “Diffaqua: A differentiable computational design pipeline
[138] M. Gazzola, L. H. Dudte, A. G. McCormick, and L. Mahadevan, “For- for soft underwater swimmers with shape interpolation,” ACM Trans.
ward and inverse problems in the mechanics of soft filaments,” R. Soc. Graph., vol. 40, no. 4, pp. 1–14, 2021.
Open Sci., vol. 5, no. 6, 2018, Art. no. 171628. [161] C. Della Santina, C. Duriez, and D. Rus, “Model based control of soft
[139] N. Naughton, J. Sun, A. Tekinalp, T. Parthasarathy, G. Chowdhary, and robots: A survey of the state of the art and open challenges,” 2021,
M. Gazzola, “Elastica: A compliant mechanics environment for soft arXiv:2110.01358.
robotic control,” IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 3389–3396, [162] F. Renda, C. Armanini, A. Mathew, and F. Boyer, “Geometrically-exact
Apr. 2021. inverse kinematic control of soft manipulators with general threadlike ac-
[140] N. Goldberg et al., “On planar discrete elastic rod models for the loco- tuators’ routing,” IEEE Robot. Autom. Lett., vol. 7, no. 3, pp. 7311–7318,
motion of soft robots,” Soft Robot., vol. 6, no. 5, pp. 595–610, 2019. Jul. 2022.
[141] W. Huang, X. Huang, C. Majidi, and M. K. Jawed, “Dynamic simulation [163] C. Armanini, A. A. Alshehhi, A. T. Mathew, I. B. Hmida, C. Stefanini,
of articulated soft robots,” Nature Commun., vol. 11, 2020, Art. no. 2233. and F. Renda, “Model-based design optimization of underwater flagellate
[142] V. Megaro, J. Zehnder, M. Bächer, S. Coros, M. Gross, and propellers,” IEEE Robot. Autom. Lett., vol. 7, no. 4, pp. 10089–10096,
B. Thomaszewski, “A computational design tool for compliant mech- Oct. 2022.
anisms,” ACM Trans. Graph., vol. 36, no. 4, 2017, Art. no. 82.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ARMANINI et al.: SOFT ROBOTS MODELING: A STRUCTURED OVERVIEW 21

Costanza Armanini (Member, IEEE) received the Christian Duriez (Senior Member, IEEE) received
B.Sc. and M.Sc. degrees in civil engineering and the Ph.D. degree in robotics from the University of
the Ph.D. degree in mechanical engineering from the Evry and CEA, Évry, France, in 2004.
University of Trento, Trento, Italy, in 2011, 2014, and He joined INRIA, Le Chesnay-Rocquencourt,
2018, respectively. France, in 2006 and he was promoted as the Directeur
She is currently a Postdoctoral Fellow with the De- de Recherche in 2014 and is currently the Head of
partment of Mechanical Engineering, Khalifa Univer- DEFROST Team. He is also currently the Research
sity, Abu Dhabi, UAE. Her research interests include Director with Inria Lille–Nord Europe, Lille, France.
continuum mechanics and structural instabilities for He participated to the creation of the open-source
the development of novel technologies, with a special SOFA framework. His research interests include soft
focus on soft robotics. robot models and control, fast finite-element methods,
simulation of contact response, and new algorithms for haptics.

Frédéric Boyer received the Diploma in mechani-

cal engineering from the Institut Nationale Polytech- Federico Renda (Member, IEEE) received the B.Sc.
nique de Grenoble, Grenoble, France, in 1991, the and M.Sc. degrees in biomedical engineering from
Master of Research degree in mechanics from the the University of Pisa, Pisa, Italy, in 2007 and 2009,
University of Grenoble, Grenoble, France, in 1991, respectively, and the Ph.D. degree in biorobotics
and the Ph.D. degree in robotics from the University from the Biorobotics Institute, Scuola Superiore
of Paris VI, Paris, France, in 1994. Sant’Anna, Pisa, in 2014.
He is currently a Professor with the Department of He is currently an Associate Professor with the De-
Automatic Control, IMT-Atlantique, Nantes, France, partment of Mechanical Engineering, Khalifa Univer-
where he is a member of the Robotics Team, Labora- sity, Abu Dhabi, UAE. His research interests include
toire des Sciences du Numérique de Nantes (LS2N), dynamic modeling and control of soft and continuum
Nantes, France. His research interests include structural dynamics, geometric robots using principles of geometric mechanics.
mechanics, and biorobotics.

Anup Teejo Mathew received the Ph.D. degree in

mechanical engineering from the National University
of Singapore, Singapore, in 2019.
He is currently a Postdoctoral Fellow with Khalifa
University, Abu Dhabi, UAE. He has authored or
coauthored several papers on topics ranging, such as
soft robotics modeling, electroactive polymers, and
energy harvesting. His current research interests in-
clude theoretical modeling of soft robotics and design
and development of underwater soft robotics.