Annual Review of Control, Robotics, and

Autonomous Systems
An Overview of Soft Robotics
Oncay Yasa,∗ Yasunori Toshimitsu,∗ Mike Y. Michelis,∗
Lewis S. Jones, Miriam Filippi, Thomas Buchner,
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

and Robert K. Katzschmann

Soft Robotics Laboratory, Institute of Robotics and Intelligent Systems, Department of
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Mechanical and Process Engineering, ETH Zurich, Zurich, Switzerland; email: [email protected]

Annu. Rev. Control Robot. Auton. Syst. 2023. 6:1–29 Keywords

First published as a Review in Advance on
actuation, design, sensing, simulation, modeling, control, machine learning,
November 14, 2022
manipulation, locomotion, robotic interaction
The Annual Review of Control, Robotics, and
Autonomous Systems is online at Abstract
control.annualreviews.org
Soft robots’ flexibility and compliance give them the potential to outperform
https://doi.org/10.1146/annurev-control-062322-
100607 traditional rigid-bodied robots while performing multiple tasks in unexpect-
edly changing environments and conditions. However, soft robots are yet
Copyright © 2023 by the author(s). This work is
licensed under a Creative Commons Attribution 4.0 to reveal their full potential; nature is still far more advanced in several ar-
International License, which permits unrestricted eas, such as locomotion and manipulation. To understand what limits their
use, distribution, and reproduction in any medium,
performance and hinders their transition from laboratory to real-world con-
provided the original author and source are credited.
See credit lines of images or other third-party ditions, future studies should focus on understanding the principles behind
material in this article for license information. the design and operation of soft robots. Such studies should also consider the
∗
These authors contributed equally to this article major challenges with regard to complex materials, accurate modeling, ad-
vanced control, and intelligent behaviors. As a starting point for such studies,
this review provides a current overview of the field by examining the working
mechanisms of advanced actuation and sensing modalities, modeling tech-
niques, control strategies, and learning architectures for soft robots. Next,
we summarize how these approaches can be applied to create sophisticated
soft robots and examine their application areas. Finally, we provide future
perspectives on what key challenges should be tackled first to advance soft
robotics to truly add value to our society.

1
 1. INTRODUCTION
Unlike traditional rigid-bodied robots, works on soft robots take nature as an inspiration to create
capable, compliant machines that can perform multiple complex tasks concurrently (1). Traditional
rigid-bodied robots are commonly programmed to efficiently perform a single task in well-defined
settings; however, they often have a limited ability to adapt to unexpectedly changing situations
due to the rigidity of their building materials (2). Because soft robots are made mostly of materials
whose elasticities are comparable to those of soft biological systems, they can move with high
degrees of freedom (DOFs), interact with their surroundings in a compliant and safe manner, and
adapt quickly to suddenly changing situations (3).
Soft robots have already presented their potential to change our daily lives. For example,
combining soft robotic mechanisms with traditional rigid-bodied robots has found use in dex-
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

terous object manipulation tasks within industrial settings (4–6). In addition, soft robots have
explored fragile deep-sea animals to help us better understand our world (7–9). Moreover, there
have been investigations into using minimally invasive soft robots for medical purposes, such as
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

endoscopy (10), surgery (11), and drug delivery (12) in delicate locations within the body.
Despite these promising developments, soft robots still have not been able to fully demon-
strate their hoped-for potential to revolutionize our daily lives. In this context, future research
efforts should ideally concentrate on three aspects: (a) engineering new materials that appropri-
ately interface soft and rigid robotic components and allow for the creation of multifunctional
soft robots that possess lifelike properties, such as self-healing, sensing, and growth; (b) better
understanding the principles behind soft robots’ dynamics with computational tools specifically
developed for them; and (c) enhancing existing soft robots’ performance by using control and
learning architectures that are more suitable for them.
In this review, we provide an overview of the state of the art in soft robotics and discuss steps that
could be followed to realize future soft robots. We first examine the working mechanisms of the
five most predominantly used soft robotic actuation modalities. Then, we introduce the modeling
and simulation tools that have been recently developed to make the underlying physics of soft
robots tractable. After that, we present various control approaches and controller architectures
that have been specifically created to control soft robotic platforms. Next, we examine the sensor
technologies that can be used to create more proprioceptive soft robots. Finally, we discuss the
common application areas of soft robots and conclude our review with our perspective on required
future advances in the field of soft robotics.

2. SOFT ROBOTIC ACTUATION MODALITIES

Actuators are the artificial muscles of soft robots that allow them to deform and move. The se-
lected actuation modality greatly influences soft robots’ design, fabrication, modeling, control,
and performance. Therefore, we start our review by analyzing the working principles of the
most promising soft robotic actuation modalities (fluidic, electrostatic/electrochemical, thermal,
magnetic, and biohybrid) and discuss their influence on soft robots’ design and fabrication.

2.1. Fluidic Actuation

Fluidic actuation is one of the most commonly utilized modalities to controllably deform soft
robots. It achieves motion through the spatiotemporal control of the fluidic pressure inside a
soft robot’s internal inflatable cavities (i.e., channels and chambers). Soft robots that are actuated
with this modality are composed mainly of elastomeric materials, and they have one or more
inextensible reinforcement layers, such as fibers, fabrics, or stiffer materials, around parts of their
internal inflatable cavities. These reinforcements constrain the direction of deformation to achieve

2 Yasa et al.
 a Fluidic actuation b Electrostatic actuation
Inflatable cavity ΔP = 0 i ii
V Opposing compliant V Electrode V
electrodes + –

+ –
Constraint layer
+ –
Dielectric elastomer
Air or liquid ΔP > 0
P
V Expansion
P P
P
Contraction
P
P
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

Contraction Liquid dielectric

material
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

c Electrochemical actuation d Thermal actuation e Magnetic actuation

V Electrode V Two-way SMA T < TSME Soft elastomer Bext = 0
Anode
+
– + –
+ –
– + Cathode +–
– – Custom-oriented
+ –
+ + – V magnetic domains
– – + –
–+ +– Joule heating T > TSME Bext > 0
+–
+ – + –
– +
+

Ion-conducting Swelling
polymer Contraction Realignment to Bext
V

Figure 1
The working mechanisms of the most promising soft robotic actuation modalities. (a) Fluidic actuation. The spatiotemporal control of
the fluidic pressure P inside soft robots’ internal inflatable cavities leads to actuation. (b) Electrostatic actuation. Attractive electrostatic
forces acting on two compliant opposing electrodes squeeze the sandwiched solid dielectric materials and lead to deformation (subpanel
i), and attractive electrostatic forces coupled with hydraulic pressure lead to actuation (subpanel ii). (c) Electrochemical actuation. The
migration of free cations toward the cathode causes a swelling effect and creates a stress gradient that leads to deformation. (d) Thermal
actuation. Joule heating above a certain temperature TSME causes a configuration change in SMAs, which then causes actuation.
(e) Magnetic actuation. Magnetic materials are entrapped in soft, deformable structures. An external magnetic field Bext causes a
spatiotemporal alignment of the entrapped magnets toward that field, leading to the structure’s deformation. Abbreviations: SMA,
shape-memory alloy; SME, shape-memory effect; V, voltage.

the desired deformation and motion. Pumps pressurize and depressurize the internal inflatable
cavities with fluids, such as air or water, and the internal fluidic stress generated by this process
deforms the robot’s body (Figure 1a). The deformation rate can be tuned by controlling the
amplitude and duration of the fluidic pressure in each cavity.
The inflatable cavities and surrounding structure can have different designs and segment mor-
phologies, such as cylindrical (13), pleated (14), or ribbed (15), depending on the intended robotic
operations (16). Their design determines a soft robot’s actuation speed. The actuation speed
can be enhanced by reducing the amount of fluid that is needed to inflate the cavities, allow-
ing for a rapid internal pressure change (14). Moreover, each segment’s morphology has its own
unique advantages and disadvantages that affect the robot’s performance and operational life-
time (16). Inextensible reinforcement layers, also known as constraint layers, can be created by
simply increasing the thickness of elastic materials or by using stiff materials or inflexible fabrics.

www.annualreviews.org • An Overview of Soft Robotics 3

 Most fluidic soft robots are pressurized by rotary or reciprocating pumps, where the volumetric
displacement is driven by rotating gears, screws, plungers, or diaphragms (17).
The main advantages of fluidic actuation are that it generates a wide range of motion, including
bending, twisting, and elongation, and can achieve high forces. Furthermore, continuous power
is not needed to keep the robots in their actuated state, since they can stay in that state simply by
closing their valves. The main disadvantages of fluidic actuation are that it is prone to puncture, is
not energy efficient, and is difficult to accurately control due to the nonlinear material response
and high susceptibility to failure under large payloads.

2.2. Electrostatic/Electrochemical Actuation

Electrostatic/electrochemical actuation happens through the direct translation of electrical en-
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

ergy to soft body deformation (and heat). Several different types of soft robots operate with these
actuation modalities. The designs of each actuator type resemble capacitors—two opposing elec-
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

trodes separated by a dielectric (in the case of electrostatic actuation) or an ion-conducting (in the
case of electrochemical actuation) material.

2.2.1. Dielectric elastomer actuators. Dielectric elastomer actuators (DEAs) work with at-
tractive electrostatic forces acting on two opposing stretchable electrodes, which are separated by
a compliant dielectric layer of solid material (18). DEAs can be powered by stimulating the oppos-
ing electrodes (reported voltages in the range of 0.3–27 kV); the voltage depends on the design,
material properties, and intended actuation strains (19, 20). When they are actuated, the oppos-
ing electrodes attract each other and squeeze the solid dielectric material layer through Maxwell
stress (Figure 1b, subpanel i). Dielectric material layers squeeze orthogonally, and they extend
parallel to the opposing stretchable electrodes under electrical fields due to their incompressible
nature (21). DEAs can present different deformation patterns, such as contraction, elongation,
and twisting, according to their arrangements (22). It is easy to control DEAs because there is a
direct connection between the applied voltage and the magnitude of deformation; moreover, the
elastomeric forces from the dielectric materials reset the DEAs’ shape when the actuation voltage
is lowered.
Most DEAs are manufactured using spray or spin coating and blade or slot-die casting (23).
There is a general drive in this field to lower the required actuation voltage due to limitations in
high-voltage power electronics and potential safety concerns. The required voltage input for ac-
tuation can be decreased by improving the properties of the dielectric materials (ϵr ) or by reducing
the thickness of the dielectric layers. For example, prestretching the solid dielectric material layers
is a common technique to improve the actuator’s performance (22). However, this improvement
introduces rigid components into otherwise fully soft robots.

2.2.2. Hydraulically amplified electrostatic actuators. Electrostatic actuation with hydraulic

amplification is another type of electrostatic actuation that couples electrostatic and hydraulic
forces together to achieve muscle-like performance (24). In general, the actuator units of these
soft robots are created by partially covering both sides of an elastomeric deformable shell with
conductive ink that acts as electrodes. The shell is then filled with liquid dielectric materials, such
as oils. The presence of liquid dielectric materials between the opposing electrodes generates local
hydraulic pressure inside the actuator upon electrical stimulation in the kilovolt range (Figure 1b,
subpanel ii). Unlike DEAs, the generated hydraulic pressure, coupled with Maxwell stress, advan-
tageously causes the actuators to deform extensively. Another potential advantage of using liquid
dielectric materials in a hydraulically amplified design is that the actuator is able to immediately
recover from dielectric breakdown events if a suitable material system is used (25).

4 Yasa et al.
 Several hydraulically amplified electrostatic actuator designs have been reported in the lit-
erature. Donut-shaped hydraulically amplified self-healing elastomer (HASEL) actuator designs
present a pull-in transition when they are actuated and appear in a toroid shape. Similar planar
designs present in-plane expansion when they are actuated, and their operation principle resem-
bles that of DEAs (24). Peano-HASEL actuator designs are a series of rectangular pouches and
present in-plane contraction when they are actuated (26). Another type of actuator that works
at much lower voltages is hydraulically amplified zipping electrostatic actuators for haptics (27);
liquid-amplified zipping actuators for micro-aerial vehicles (28) work on a similar principle. Al-
though most of these actuators are created by using molding (24) and heat-sealing (26) methods,
3D multimaterial printing platforms can also be used to fabricate them (29).

2.2.3. Ionic polymer–metal composite actuators. Ionic polymer–metal composite (IPMC)

Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

actuation is used to manufacture soft electrochemical actuators that are composed of an ion-
conducting polyelectrolyte layer sandwiched between two parallel-aligned flexible opposing
electrodes. Unlike electrostatic actuators, such as DEAs and hydraulically amplified electrostatic
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

actuators, IPMC actuators work at low stimulation voltages, from 1 to 5 V (30). Inside the poly-
electrolyte layer are polymers with a carbon backbone that stabilize the negatively charged ions
(anions) at their positions, and the positively charged ions (cations) can freely move inside the
polymer network. When an electric field is applied to the electrodes, the free cations migrate to-
ward the cathode and cause a swelling effect around the cathode (Figure 1c). This swelling effect,
which is due to the inhomogeneous distribution of ions inside the polyelectrolyte layer during
electrical stimulation, creates a stress gradient and causes the actuators to deform (31).
IPMC actuators respond more slowly to the applied electric fields than DEAs and HASEL ac-
tuators, since their deformation is dependent on the migration speed of cations toward the cathode.
In addition, the water content of the polyelectrolyte layer is of paramount importance in the ac-
tuators’ performance, and most of these actuators can therefore only work in moist environments
or underwater. Finally, IPMC actuators can be fabricated using different techniques, such as hot
pressing (32), sputtering, or electroless plating (33).

2.3. Thermal Actuation

Shape-memory alloys (SMAs) are the most commonly used temperature-responsive metallic
composite materials in artificial muscles for soft robots. These materials present two essential
characteristics: the shape-memory effect and the superelastic effect (34). The shape-memory ef-
fect arises when SMAs are heated above their transformation temperature; due to this effect,
even if the SMAs are permanently deformed at low temperatures, they return to their original
(pre-deformed) states when they reach a temperature that is higher than their transformation
temperature. SMAs can exist in two different phases and demonstrate different shape-memory
effects, the most common of which are one-way shape memory and two-way shape memory.
Once returned to their original states, one-way SMAs cannot return to their previous deformed
states when the temperature decreases below their transformation temperatures. However, de-
pending on their surrounding temperatures, two-way SMAs can exist in their deformed (cold)
and pre-deformed (heated) states. Although the superelastic effect seems to demonstrate behavior
similar to the shape-memory effect, it represents the isothermal recovery of SMAs above their
transformation temperatures following a mechanical loading and unloading cycle.
In soft robotics, SMAs are used as structural elements for actuation inside robots’ elastomeric
matrices. These materials deform and thus actuate soft robots due to their shape changes, which
are induced by Joule heating (Figure 1d). Some SMAs can deform by up to 8% when they are
in a straight shape; however, when they are designed in a spiral shape, their deformation can

www.annualreviews.org • An Overview of Soft Robotics 5

 exceed 300% (35). Due to the necessity of electrical stimulation, most SMA-based soft robots
that have been reported in the literature are tethered (36), and they operate underwater to en-
hance actuation speed and performance by increasing the cooling rate of SMAs (37, 38). However,
high-performance thermal robots that are actuated with SMAs can be manufactured with thermo-
conductive elastomers. These elastomers make the SMAs cool down and return to their natural
states more quickly (39).
The major drawback of SMAs is their reliance on temperature changes for actuation, which
decreases their energy efficiencies. Their bandwidth is also limited by the thermal cycle, so me-
chanical fatigue can occur following a few actuation cycles. Moreover, these materials present
nonlinearity and wide hysteresis, both of which cause control difficulties. On the other hand, they
have a high work density, can deform under low voltage stimulation, and operate cleanly and
silently, which decreases the adverse effects of toxic degradation products and constant noise on
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

living beings both underwater and on land. SMA actuators can be fabricated in a similar way to flu-
idic actuators, but their active components are pre-deformed SMAs rather than internal inflatable
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

cavities.

2.4. Magnetic Actuation

Magnetic actuation commonly arises from the spatiotemporal alignment of magnetic materi-
als (which are trapped in soft, deformable structures) under the influence of external magnetic
fields (40). This actuation principle allows for the creation of untethered, compliant soft robots
that can potentially perform safe operations inside delicate unstructured environments, such as the
brain and inner body cavities (41, 42). Soft robots that use this actuation modality are composed
mainly of microscopic ferromagnetic materials that are integrated into soft elastomeric matrices.
The spatial alignment of the magnetic domains of hard ferromagnets during the fabrication pro-
cess is the key to achieving the desired soft body deformation and motion (43, 44) (Figure 1e). The
magnetic domains of materials can be aligned by first molding magnetic pre-polymer solutions
into pre-defined complex structures and then applying strong external magnetic fields above their
magnetic saturation points (12, 45). Alternatively, one can directly align the magnetic domains us-
ing external magnets while 3D printing a soft robot (46, 47). Moreover, the domains of magnetic
materials can be reprogrammed even after fabrication by first erasing the magnetic memory of ma-
terials through heating above their Curie temperatures (48, 49) and then remagnetizing according
to the new desired domain arrangements.
The safe penetration of magnetic fields into the human body is one of the main advantages of
this actuation modality. This property allows the development of compliant robots that operate
inside the human body in an untethered fashion. However, setups with electromagnets or per-
manent magnets generate inhomogeneous magnetic fields that complicate the control of these
soft robots. In addition, even though these robots are untethered, large stationary electromag-
netic setups are needed for their powering and control. The necessity for such large and complex
setups that can produce homogeneous magnetic fields in a tiny volume limits the scaling up of
magnetic soft robots. Finally, controlling these magnetic robots becomes difficult when there are
metals in close proximity, as the robots are attracted to them. Therefore, their application has been
concentrated mainly in endoluminal procedures.

2.5. Biohybrid Actuation

Integrating the advanced functions of cells into soft robots could allow us to leverage the perfor-
mance and features of living materials. In soft robotics, biohybrid actuation can be achieved by
integrating contractile eukaryotic cells, such as skeletal muscle cells and cardiomyocytes, into soft

6 Yasa et al.
 robot designs (50). These cells are intrinsically soft and compliant, and when they are combined
with biodegradable materials, they provide an eco-compatible and sustainable source of engineer-
ing materials to build soft robots (51). Soft robots that are made of living materials also carry their
own fuel source (nutrients in the surrounding medium) and display a higher power-to-weight ra-
tio than other actuation technologies (52). Moreover, living cells can proliferate to reconstitute
lost parts of their assemblies, and therefore, soft robots that are built with living materials possess
efficient self-healing functions (53).
Contractile cells of either mammalian (54) or nonmammalian (55) origin can be used to create
biohybrid soft robots. These soft robots can be fabricated by directly extracting functional con-
tractile tissue from living organisms, known as the top-down approach, or by engineering tissues
in cell culture with primary cells or cell lines, known as the bottom-up approach (50).
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

2.5.1. Mammalian origin. Primary cardiomyocytes and skeletal muscle cells are the most com-
monly used cell types of mammalian origin for biohybrid actuation. Cardiomyocytes are widely
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

extracted from neonatal organisms and then are functionally integrated into soft, deformable ma-
terials to create thin-film soft robots (56–58). However, their involuntary contraction drastically
limits the controllability of biohybrid soft robots. To overcome this shortcoming, genetic engi-
neering strategies (i.e., optogenetic engineering) have been used to control the contraction of
these cells via light stimulation (59, 60).
Unlike adult cardiomyocytes, skeletal muscle cells easily grow into 3D assemblies on a sub-
millimeter scale. This property of skeletal muscle cells allows for the fabrication of 3D biohybrid
actuator components rather than 2.5D thin films (61, 62). In addition, skeletal muscle cells con-
tract only when they are excited by an external stimulus, which eliminates the control issues
that arise from cardiomyocytes. However, the restrictive requirement of electrical stimuli to
generate contractile forces for locomotion caused researchers to engineer optogenetic skeletal
muscle cells to power the robots with external light pulses (63). Skeletal muscle–based soft robots
can even locomote when optogenetically engineered nerve cells are integrated into the robot’s
body (64). However, the low incorporation efficiency of nerves with muscle bundles complicates
the fabrication of nerve-controlled muscle-based soft robots.

2.5.2. Nonmammalian origin. Self-contractile tissues of nonmammalian origin, such as frog

tissue (65) and insect tissue (66), can also be used to fabricate robust biohybrid soft robots. These
tissues can tolerate a wide range of environmental fluctuations and can survive in extreme pH
levels and temperature conditions without losing their contractility and function (67). In addition,
soft robots that are built with these tissues can operate at an ambient temperature without any
additional carbon dioxide and oxygen supply. Compared with mammalian tissue, the use of non-
mammalian tissues in biohybrid soft robots is advantageous for the translation of biohybrid soft
robots to real-world conditions. However, because most of these tissues still need to be directly
extracted from living animals, this fabrication approach raises ethical issues (50).

3. MODELING AND SIMULATION

Conventional modeling and simulation solutions for traditional rigid-bodied robots do not ful-
fill the requirements to characterize soft robots’ behavior. Therefore, we need modeling and
simulation solutions that have been specifically developed to accurately represent soft materials’
responses to external stimuli. This section introduces different methods that are used to model
and simulate soft robots, with the goal of applying them to the optimization and control of soft
robots. Figure 2 provides an overview of the methods for modeling soft robots.

www.annualreviews.org • An Overview of Soft Robotics 7

 3.1. Numerical Mesh-Based Modeling
In soft robotics, the most accurate modeling approach is based on simulating the continuum me-
chanics that govern the robots’ deformation due to internal and external forces during actuation.
In the numerical mesh-based modeling approach, soft robots are defined by their material prop-
erties, such as Young’s modulus and Poisson’s ratio, and the representation of their geometry. In
general, soft robots’ geometry is discretized with a mesh of simple shapes known as elements.
Examples include triangles/rectangles and tetrahedrons/hexahedrons in 2D and 3D spaces, re-
spectively (71–73). Other shapes, such as regular uniform grids or particles, can also be applied
to represent the robots’ geometry (74). The discretization of the robots’ geometry also allows the
equations to be discretized for every element and solved numerically. A sufficiently fine mesh is
essential for accurately representing soft robots’ behaviors in a simulation framework based on
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

a finite element method (FEM). In this regard, the Simulation Open Framework Architecture
(SOFA) (75) has been successfully applied to model different soft robots (68, 71, 76). Due to its
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

a FEM model b PCS model

z Bending (ux,uy) Shearing (vx,vy)
Design Meshing Solving x s=L
y

p(t,s), R(t,s)

s Twisting (uz) Stretching (vz)

x

z
y s=0

c PCC model

μi
ôi {Si}

Li μi Θi /2
i)
ñi Θ
ñi
b(
{Si}
i

a i(Θ
ôi–1

i)
{Si–1}
{Si–1} ni
{Si–1}
Θi /2
Θi Θi /2

ñi–1 Øi ñi–1 Li/Θi

d Learning-based models
States xt

States xt

xt+1 Latent states Øt Øt+1 xt+1

(Caption appears on following page)

8 Yasa et al.
 Figure 2 (Figure appears on preceding page)
An overview of the modeling strategies used in soft robotics. (a) A FEM simulation based on numerical mesh-based modeling. In this
type of simulation, soft robots are defined by their material properties and represented by meshes before solving the underlying,
discretized physical equations. The FEM model shown here is of a soft continuum arm with selectively pressurized cavities. Panel
adapted from Reference 68 with permission from IEEE. (b) A PCS model based on the Cosserat rod theory, which represents the
deformation as a series of rods and can be used to model different deformation states of soft robots. The PCS model has four modes of
deformation: bending, twisting, stretching, and shearing. In this panel, s describes the position along the Cosserat rod (going from 0 to
L), and t describes time. At any point on the rod at any given time, we can describe its deformation using a frame represented by a
translation p(t, s) and a rotation R(t, s). We can apply different external loads, such as torques in the x and y direction (ux , uy ) or forces in
the x and y direction (v x , v y ), causing bending or shearing, respectively, or torques in the z direction (uz ) or forces in the z direction (v z ),
causing twisting or stretching, respectively. Panel adapted from Reference 69 with permission from IEEE. (c) A PCC model, which is a
further simplification of the PCS model that explains only the bending behaviors of soft robots. The middle subpanel depicts an
augmented rigid arm model, which is used to compute the PCC dynamics. The right subpanel describes the geometry for computing
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

the position of the center of mass for the augmented rigid arm model. Every segment i has a local frame Si , a length Li , an angle of
curvature i , and an orientation i of the curvature plane with respect to the plane described by ñi−1 and õi−1 ; the center of mass μi has
a distance to origin bi and a distance to the center of curvature ai . Panel adapted from Reference 70 with permission from IEEE.
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

(d) Learning-based models that predict the dynamic behavior of (often) nonlinear systems either in the original state space or in a
latent/hidden representation of the states. Depending on the objective, this latent space can be low-dimensional if we want to reduce
the complexity of the input to only its core information (similar to reduced-order modeling), or it can be high-dimensional if we want
to linearize the dynamics, as seen in the Koopman operator. The design of the latent space is a trade-off between the complexity of the
dynamics and the complexity of the state. Abbreviations: FEM, finite element method; PCC, piecewise constant curvature; PCS,
piecewise constant strain.

open-source nature, it has gained popularity in the soft robotics community over other traditional
FEM frameworks, such as COMSOL Multiphysics (77) and Abaqus (78).
It is crucial to model how soft robots interact with their surroundings using a simulation frame-
work that includes contact dynamics. To this end, several efficient and accurate time-stepping
methods for elastodynamics have been proposed for soft body collisions (73, 79). These methods
solve the dynamic equations using constraints that represent the collisions. Whereas Incremental
Potential Contact (IPC) solves the constrained minimization problem for the dynamic equations
that use a contact-aware Newton-type solver (73), others have approached the task by directly
solving the mixed linear complementarity problem, which arises from the constrained optimiza-
tion (79). These results were originally introduced in the computer graphics field, but applying
them to modeling extreme deformations in real-world scenarios shows promise in the applicability
of this method to soft robotics.
In recent years, it has also been desirable to use differentiable frameworks for simulation.
These frameworks allow us to obtain the gradients of an objective (e.g., how high a soft robot
can jump in 10 seconds of simulation) with respect to input variables, such as the control signal
or length/height of the robot. Differentiable Projective Dynamics (DiffPD) provides an efficient
simulation framework, based on projective dynamics, whereby nonlinear internal forces are ap-
proximated by quadratic energy terms (72). This approximation allows the solver to precompute
part of its solution, which speeds up the process compared with FEM frameworks. DiffPD has
shown its control and design co-optimization capabilities in the context of simulated underwater
swimmers (DiffAqua) (80). DiffAqua uses simplified hydrodynamics in simulations, which only
coarsely model fluid flow for real-world scenarios. Thus, a fast multiphysical simulation frame-
work with fully simulated solid and fluid dynamics will eventually render these methods more
practical. Since full numerical fluid simulations are prohibitively computationally expensive, an
initial step toward this multiphysical system has been taken by using a deep learning surrogate
model for fluids in combination with fluid–structure interaction (81).
ChainQueen is another approach to simulating deformable materials using a material point
method that ensures differentiability and physical accuracy in real-world experiments (74).

www.annualreviews.org • An Overview of Soft Robotics 9

 Particle-based simulation frameworks are often used in computer graphics because they are
extremely efficient, but they lack physical accuracy. The material point method combines
the efficiency of particle-based methods with grid/FEM-based accuracy by projecting between the
particle and grid (and vice versa) to apply forces onto the particles and solve the equations on the
grid. This latter step is achieved by aggregating values from the particles to fixed grid nodes.
Unfortunately, transferring dynamics from simulation to reality is often not accurate. This
inaccuracy can be referred to as a sim-to-real gap, and it has been analyzed for a few soft structures
in DiffPD (82, 83). This sim-to-real gap must be closed to achieve accurate physical models and
optimize the control and design of soft robots.
Full FEM simulations are, in practice, feasible only for simple soft robots, i.e., meshes with
less than approximately 1,000 elements. The control and optimization of more complex struc-
tures are often computationally too expensive to simulate. Therefore, we need to simplify complex
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

FEM simulations. One approach that builds on FEM is reduced-order modeling (68) (Figure 2a).
Reduced-order modeling, in essence, truncates the state by orthogonally decomposing the state
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

trajectory data from a snapshot experiment. In this snapshot, the states that best describe the mo-
tion are kept, which limits the generalizability of reduced-order modeling to how the experiment
is defined.

3.2. Simplified Parameterization of Deformation

Further simplifications are often dependent on the morphology of soft robots. A Cosserat rod
model is a type of deformation model that considers bending, twisting, shearing, and stretching
for discrete line segments along the continuum soft body (84–87) (Figure 2b). Thus, a kinematic
deformation model of a Cosserat rod model can also be referred to as a piecewise constant strain
(PCS) model. As the term Cosserat rod model suggests, soft structures are approximated using
long cylindrical segments or rods. The continuum mechanics for these segments or rods can be
simplified since their symmetries are known; hence, the accuracy of this model depends greatly on
the soft robotic shape in question. The PCS model can be simplified even further by considering
only the bending of each segment. Piecewise constant curvature (PCC) (70, 88) (Figure 2c) is one
of the simplest models used to model continuum soft robots. It assumes that, within each discrete
segment along the continuous body, the bending curvature is constant and can be described by
two parameters. Although PCC and PCS are sometimes introduced as opposing concepts, the
PCC model can, in fact, be considered a subset of the deformation modes of the PCS model
used by the Cosserat rod theory. The user decides how to separate the whole length of the robot
into discrete PCC sections, but, in many cases, a single PCC section is assigned to each actuated
segment (70, 89, 90). This assignment has the benefit that the model is fully actuated, enabling
trajectory tracking in a configuration space (70, 89, 90). However, a finer discretization of the PCC
sections may improve model accuracy (91).
Other deformation models are also used to understand soft robots’ behavior. For instance, a
modeling framework for continuum soft arms based on spline interpolation between nodes allows
the modeling of static forces, such as elasticity, gravity, and actuation forces (92). In addition,
polynomial curvature models (93) and Pythagorean hodograph curves (94) are used to simulate
bending in soft robotic arms. The aforementioned deformation models have various degrees of
fidelity. It should be noted that the choice of parameterization for deformation is usually a means
to an end—the most important aspect of a soft robot’s performance is how accurately it can be
controlled, and this depends on not only the deformation model but also the controller framework.
Thus, it is important to choose a model that sufficiently describes the movement of the designated
robot design. For example, the soft continuum proprioceptive arm (SoPrA) is a manipulator that
contains an inextensible fabric along its center axis, which limits its ability to stretch and shear;
10 Yasa et al.
 therefore, the PCC model can be applied (91). In other cases, it may be advisable to also add a
stretching deformation mode to the model. Even if a simpler model is less accurate, its inaccuracy
can be compensated for by using an effective feedback or adaptive controller (95).

3.3. Learning-Based Modeling

Unlike conventional modeling strategies, machine learning methods can also be used to represent
soft robots (Figure 2d). For instance, neural networks can be trained in a supervised manner
by using experimental data to accurately understand the dynamics of soft robots in a simulation
environment. In most cases, the state space of soft robots is rather low-dimensional (e.g., bending
angles or motion marker locations) (96, 97). In an extreme case, we could use high-dimensional
FEM simulation data to train a neural network (98), which could then be used during the testing
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

phase as a black-box forward model for fast inference. However, this is a computationally expensive
approach, and although it shows potential for modeling structural deformations, such a method
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

has not yet been applied to soft robots in practice. Furthermore, these supervised methods cannot
be generalized to unknown scenarios, since they must be supervised and require new data to be
collected from scratch after every significant modification in the soft robot’s design. On the other
hand, this efficient forward method can be applied to the closed-loop control of soft robots (96),
and it performs impressively when it is combined with model predictive control (MPC) (97).
A low-dimensional representation for a soft continuum arm’s dynamics has also been discov-
ered with an autoencoder neural network architecture. Soter et al. (99) trained this architecture by
using the soft robot’s movement data, which were acquired with an RGB camera. In this work, the
authors used a latent representation to generate a map between a proprioceptive bend sensor’s data
and the low-dimensional representation of the robot. Eventually, it was possible to reconstruct the
robot’s motion by using only the sensor’s data. This work demonstrates that the complex dynamics
of soft robots can be modeled in a simple state space by using these encoding techniques.
Finally, the dynamics of soft robots can also be modeled with dynamic mode decomposition
using the Koopman operator (100). This method involves linearizing the nonlinear system dy-
namics of soft robots in a high-dimensional latent state space so that the linear dynamics can be
used for control strategies. This concept has been extended to model and control stochastic sys-
tem dynamics using a deep stochastic Koopman operator (DeSKO) (101) with a stability-assured
control method for soft robots.

4. CONTROL STRATEGIES
Like modeling, the control of soft robots is challenging due to their continuum nature. Con-
ventional control schemes that are used for controlling rigid-bodied robots cannot be directly
applied to soft robots because they assume that discrete joints are located along a chain of rigid
links. Therefore, we need precise controllers that have been developed specifically for dynamic
soft robots with their high DOFs. Figure 3 provides an overview of the methods for controlling
soft robots.

4.1. Model-Based Control

In this control approach, a model is first created with a priori knowledge about the structure of soft
robots; then, the model is used to formulate a controller. This class of controllers contrasts with
data-driven control approaches, which generate appropriate control policies directly from actual
motion data. Because model-based control has been extensively researched in conventional rigid-
bodied robotics (103), there are several existing controller architectures that can be applied to soft

www.annualreviews.org • An Overview of Soft Robotics 11

 a Model-based proportional– b Koopman-based control c Soft actor–critic model for
derivative control reinforcement learning
xd ut xt
Soft robot Feedback controller
–Dc J(q)q˙ + Kc (xd – x)
Sampled Training
x Replay batch Learns policy πФ to maximize
DeSKO controller μt buffer
Eτ~P,π [∑∞
t=0
γ trt]
K
J T(q)
πФ
State-feedback controller
q,q̇ Action sampling
Nominal system
μ̂t st + 1
Samples an action from policy
G(q) μˆ t + 1 = Aμˆ t + Bct (st,at,rt,st+1)
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

at ~ πФ (a|st)
Εxt = Cμˆ t
|

J+T(q)(Kq + Dq)
˙
ct at
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

MPC controller
˙ –
(q,q)(I J+ (q)J(q))q̇ Sensor Actuator

Figure 3
An overview of the control strategies used in soft robotics. (a) Model-based proportional–derivative control of a soft continuum arm in
a task space. This approach enables Cartesian impedance control to the desired end-effector position xd by attaching a virtual spring
and damper (with stiffness Kc and damping Dc ) to the end effector. Panel adapted from Reference 90 with permission from SAGE
Publications. (b) Koopman-based control using the DeSKO model. The nonlinear dynamics can be described by a linear model using
the Koopman operator K, enabling linear control methods to be applied. Panel adapted with permission from Reference 101. (c) An
adapted soft actor–critic model for reinforcement learning on soft robots. The controller learns a policy π φ to maximize the objective
function. This policy is based on the robot’s history of state s, action a, and reward r. Panel adapted from Reference 102 with permission
from IEEE. Abbreviations: DeSKO, deep stochastic Koopman operator; MPC, model predictive control.

robotics. However, these architectures need some modifications due to the particular properties
of soft robots, which complicate their direct application.
The first example of these modifications is that soft robots are composed of compliant materials
that are not assumed in conventional models and controllers for rigid-bodied robots. Therefore,
additional elements, such as linear (70) or nonlinear (104) elasticity, must be introduced to ad-
dress the compliant nature of soft robots. Furthermore, soft robots have infinite DOFs due to
their continuous deformation capability, and this capability must be approximated with discrete
models, such as the PCC (70) or PCS (84) model. Depending on the discretization of the deforma-
tion model, the configuration space can have more DOFs than the number of actuators, making
them underactuated systems. In this case, underactuated control techniques, such as optimal con-
trol, must be used to generate control inputs. In addition, the actuation space of soft robots is
usually force based (e.g., air pressure for fluidic actuation, or voltage input for electrostatic actua-
tion) rather than position based. Whereas low-level firmware for rigid robots often accepts joint
position as input from the controller, the control signal for soft robots must typically be sent in
the force space.

4.1.1. Kinematic model-based control. Kinematic model-based control considers the posture
of soft robots without any force. Therefore, this approach only requires a simple model, and it
can perform only position-based or velocity/acceleration-based control. This property makes it
less effective for interactive tasks, such as manipulation. Nonetheless, it is effective at achieving
demonstrative behaviors; for example, a kinematics-based controller that uses a simplified Jacobian
has been used to control the tip position of a soft continuum arm and applied to tasks such as
opening doors and performing stick-shift operations (105). This observation suggests that, for

12 Yasa et al.
 soft robots, a relatively simple kinematics-only low-level controller, paired with an effective high-
level planner that defines the specific task, can be as performant as controllers that are based on
complex models. Inverse kinematics, which is the process of calculating the pose that is required
to achieve a certain task (e.g., the tip position), presents additional challenges for soft robots with
high DOFs (although a closed-form solution can be found in some limited cases) (106).

4.1.2. Dynamic model-based control. Dynamic model-based control considers the dynamics
of soft robots; it starts with creating a manipulator equation (i.e., the equation of motion for a
link chain) and proceeds by applying a controller based on this equation. This control approach
considers physical forces such as inertia or gravity, and thus, it can achieve more accurate control
than its kinematic counterparts, and it is more versatile in achieving force-based interactions with
a robot’s surroundings.
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

In the augmented rigid body model approach, the manipulator equation is calculated by using
a rigid link model that emulates the structure of soft robots (70, 89–91). This augmented rigid
body model is a virtual link structure that matches the kinematic structure of the PCC model at
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

both ends of the PCC section. It also matches the center of mass’s position, thereby approximating
the dynamic properties of soft robots. The dynamics of this model can be computed with standard
robotics libraries and then converted to a robot’s configuration space. This strategy has been ap-
plied to dynamic feedback control for soft robotic arms in 2D (89, 90) and 3D (70) configurations
(Figure 3a). Lagrangian dynamics can also be used to obtain the dynamics of the manipulator.
In this method, the potential energy V and kinetic energy T of soft robots are computed analyti-
cally, after which the Lagrangian L = T − V is used in the Lagrangian equation of motion. This
strategy results in a more computationally efficient calculation of the dynamics compared with the
augmented rigid body method (95).
Rather than focusing on improving the trajectory accuracy of soft robots in isolated experi-
mental environments, recent research on dynamic model-based soft robotic control has started
to focus on applications to real-world tasks for soft robots by incorporating concepts that were
developed for rigid arm control. Operational space control is an area of research in which clas-
sic robot control has recently been applied to soft robots. Operational space control is important
for robots to achieve specific tasks, and it has been researched from both a theoretical (107) and
practical (108, 109) perspective. For example, it has been shown that even underactuated systems
that have an actuation space with the same DOFs as the task can achieve operational space con-
trol (107). Recently, a soft arm was able to achieve practical and dynamic movements, such as
drawing and throwing, by using operational space control (109).
It is often difficult to obtain an accurate model parameter value for soft robots due to their
nonlinear material properties or unexpected deformations, which occur due to environmental
interactions. Therefore, adaptive control is another area of research in which classic robotic con-
cepts have been applied to soft robots. Adaptive controllers based on the dynamics model from
Lagrangian dynamics (95) and the augmented rigid body model (110) have been proposed for soft
robots. Controllers that exploit a priori knowledge about the uncertainty of robots’ movements
and adapt to specific changes, such as adapting to orientation errors, have also been proposed
(111).

4.2. Data-Driven Control

The model-based control approaches that were introduced in the previous section use a priori
knowledge about a robot’s structure to construct an analytical model. By contrast, data-driven
control approaches use motion data from an actual robot’s movement or its simulated model to
train the controller. Data-driven control can be effective at capturing individual differences in, or

www.annualreviews.org • An Overview of Soft Robotics 13

 the nonlinear properties of, soft robots that are hard to model analytically, although data must be
collected for training.
The Koopman operator linearizes the nonlinear dynamics of soft robots for modeling and sim-
ulation. It can also be used to successfully apply linear control theory to nonlinear systems (100).
For example, the use of robust MPC with a Koopman operator enhances soft robots’ performance
and stability against disturbances (101) (Figure 3b). Learned models have also been used in com-
bination with classic MPC for soft continuum arms (96). The MPC is based on a dynamic model,
learned from motor babbling, and it enables a soft arm to reach statically unreachable poses by
swinging upward, greatly increasing its workspace.
As a counterpart to traditional control methods, reinforcement learning has also been applied
to a fluidic soft arm for position control (112) (Figure 3c). Deep Q-learning, which is a model-
free reinforcement learning approach, is used for controlling this soft arm. The reinforcement
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

learning agent is trained in simulation using the Cosserat rod theory for the model, and then the
learned position control policy is applied to the soft arm. The control policy is robust against
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

external disturbances such as loading, even though no external loads were applied during training.
A powerful reinforcement learning framework has been created by using a modified version of
soft actor–critic (102). This framework has been fully trained on experimental data, and it can be
applied to the position control of a complex soft robot actuated by vibrations and pneumatics.

5. SENSING STRATEGIES AND SENSOR TECHNOLOGIES

The sensorization of soft robots is a long-standing challenge that must be overcome to com-
plete autonomous tasks. Since most conventional sensors have been designed for rigid structures,
they are difficult to integrate into the compliant, flexible bodies of soft robots. Soft robots the-
oretically have infinite DOFs, so sensors for measuring their continuous posture changes must
have sufficient motion range, accuracy, and resolution. Such sensors, when combined with a priori
knowledge, can then properly reconstruct the soft robots’ overall structure. In this section, we first
discuss the sensing strategies (i.e., proprioception and tactile sensing) that can be applied to soft
robots to strengthen their modeling, control, and safety. We then examine the available sensor
technologies that can be used for proprioceptive and tactile sensing (Figure 4).

5.1. Proprioception
Proprioception is the sense of self-movement and body position. In soft robotics, proprioception
provides information about the current state of a soft robot, which is crucial for proper feed-
back control. Motion capture systems are often used to measure soft robots’ posture, especially
in research works that deal with modeling and control (70, 90, 95, 96, 100, 101). In this method,
reflective markers are placed along soft robots’ bodies and detected by infrared cameras, enabling
real-time ground-truth measurements of robots’ posture. However, this method requires an exter-
nal motion capture setup and does not work under occlusion, limiting its applicability to interactive
tasks in the real world. For these reasons, various efforts have been made to embed soft robots with
internal proprioceptive sensors so that their posture can be measured without motion capture
systems (91, 113).

5.2. Tactile Sensing

While proprioception refers to being aware of one’s own movement, tactile sensing involves
gathering information about the external environment through the sense of touch. One of the
advantages of soft robots is their ability to conduct compliant interactions with the environment,

14 Yasa et al.
 a Resistive flex sensor b Resistive contact sensor c Conductive ink sensor

Sensor Sensor
Actuator

Conductive ink
Contact
Resistance pressure
change
R = pl/A
Body
deformation

d e
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

Capacitive sensors Fiber Bragg grating sensor

i Self-capacitance mode ii Mutual-capacitance mode Incident
light
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Sensor
Electric Electric
field lines field lines

Reflected
Electrode TX RX light
electrode electrode

Reflected light
after deformation

f Magnetic sensor
Hall effect Magnetic Measures ΦB Magnetic field’s
sensor field lines change direction change

Soft body Permanent Body

magnet deformation

Figure 4
An overview of the sensor technologies used in soft robotics. (a) A resistive flex sensor, which detects soft robots’ deformation states
through changes in their resistance due to bending. (b) A resistive contact sensor, which detects the physical interaction of soft robots
with their surroundings through changes in their resistance due to force and pressure. (c) A conductive ink sensor, which detects soft
robots’ deformation state through changes in their resistance due to stretching. The amount of stretch that conductive ink sensors are
subject to can be measured using Pouillet’s law, which correlates the resistivity with the cross-sectional area and length of the resistive
material. (d) Capacitive sensors, which detect the interaction of soft robots with their surroundings through changes in their generated
electric fields. When an object approaches a capacitive sensor in self-capacitance mode (subpanel i), it drains the generated electric field
by the sensor; when an object approaches a capacitive sensor in mutual-capacitance mode (subpanel ii), the shielding effect reduces the
electric field that is received by the RX electrode. (e) A fiber Bragg grating sensor, which detects soft robots’ deformation state through
changes in the spectrum profile of the reflected light within the optical fiber. ( f ) A magnetic sensor based on the Hall effect. This
sensor detects soft robots’ deformation state by measuring magnetic flux changes arising from the displacement of a permanent magnet.
Abbreviations: RX, receiver; TX, transmitter.

so it is crucial for a soft robot to be able to sense when it comes into contact with something.
Methods to sense contact include resistive sensors (114), capacitive sensors (113, 115), and mag-
netic sensors (116). Most of these technologies require an array of sensors, which are placed across
a robot’s surface, to gather detailed contact sensing data. Another method that has been recently

www.annualreviews.org • An Overview of Soft Robotics 15

 developed uses a camera inside a bubble gripper with a textured internal surface. This method
allows the gripper’s deformation, which is caused by contact, to be observed (117).

5.3. Sensor Technologies

Different sensor technologies can be used to enable previously described proprioceptive and tactile
sensing abilities in soft robotics. These sensor technologies have different performance metrics
in object detection, deformation type, nonlinearity, and hysteresis. Therefore, evaluating these
metrics and choosing the most appropriate sensor technology for the envisioned soft robot’s design
is crucial.

5.3.1. Resistive and piezoresistive sensors. The resistance of elastic conductive materials
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

changes depending on their deformation states. The correlation between the changes in the
resistance and the deformation state of these materials allows soft robots’ actuation states to
be identified. This technology allows soft robots to gain information about their surroundings
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

through contact-based interactions (Figure 4a). These sensor technologies offer low-cost solu-
tions for proprioceptive soft robots (68, 118–120). However, their readout suffers from hysteresis,
and therefore measurements of dynamic movements may not be accurate (119). 3D-printing tech-
nologies can be used to print resistive sensors together with soft robots’ bodies, enabling an
integrated sensor–actuator that allows for complex sensor placements (118, 121). Moreover, resis-
tive sensors can be configured to detect contact instead of self-induced deformations (Figure 4b).
Thus, a soft gripper with resistive flex and touch sensors is able to haptically identify objects that
it is holding (114). Piezoresistive sensors work according to the piezoelectric effect; the materials’
resistivity changes according to the applied strain or pressure. However, the nonlinear readouts
and hysteresis properties of piezoresistive sensors require a trained recurrent neural network to
convert the raw measurements into pose estimation results (122).

5.3.2. Conductive liquid sensors. Conductive liquids, such as liquid metals (123) or ionic liq-
uids (124), can be used inside the internal deformable cavities of soft robots to measure changes
in the resistance according to the soft robots’ deformation states. Once these changes in the re-
sistance have been measured, the robots’ deformation states can be reconstructed using Pouillet’s
law (123) (Figure 4c). These sensors are prone to hysteresis, similar to resistive sensors. Nonethe-
less, complex sensor arrangements can be created by introducing channels for the conductive inks
during the fabrication process (125), which increases the freedom for sensor placement. An op-
timization method can be used to generate the most effective sensor route to sense deformation
and demonstrate its sensing capability on a physical sensing element (123).

5.3.3. Capacitive sensors. Capacitive sensors can be used to proprioceptively measure soft
robots’ poses and exteroceptively measure their contact with their surroundings. They can op-
erate in two different modes for haptic sensing: self-capacitance mode and mutual-capacitance
mode (113). The self-capacitance mode applies an alternating current to a single electrode and
detects nearby objects that drain the generated electric fields (Figure 4d, subpanel i). The mutual-
capacitance mode requires a pair of electrodes to send and receive signals; in this mode, an
approaching conductive object causes a shielding effect, while physical contact that brings the
electrodes closer together causes a coupling effect (113, 115) (Figure 4d, subpanel ii). It is dif-
ficult to detect nonconductive objects with capacitive sensors. However, the coupling effect for
the mutual-capacitive mode depends on physical deformation and thus can potentially be used to
sense nonconductive objects. A numerical optimization method has been used to merge measure-
ments from capacitive and pressure sensors using a FEM model in the SOFA simulator to predict
the position and magnitude of the contact force applied to soft robots (113).

16 Yasa et al.
 5.3.4. Optical fiber sensors. Optical fiber sensors that are based on the fiber Bragg grating
pattern are promising to measure the continuous deformation of soft robots. Inside these sensors,
optical fibers are engraved with a regular pattern at the point of measurement that is designed to
return a particular spectral frequency response (126). When the section with the grating is strained,
the frequency response changes, making it possible to measure the strain at that point (Figure 4e).
By introducing these patterns along the length of the fiber and bundling multiple fibers together,
one can reconstruct the sensor array’s continuous geometry. Different grating position and sensor
placement configurations have been proposed for effectively sensing the continuous curvature, as
well as several algorithms to reconstruct a soft robot’s deformation states (127–130). The fibers
are thin and require little space to be integrated into soft robots’ bodies. However, the readout
terminal that measures the frequency response tends to be bulky, prohibiting mobile applications.
Though fiber Bragg grating sensing systems are commercially available, the sensors and termi-
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

nals are expensive for widespread application. However, the sensing element’s thin form factor and
the curvature readout’s accuracy make them a viable solution for critical applications, such as min-
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

imally invasive surgery (130), when the high-cost and large-readout equipment can be afforded in
return for better outcomes.

5.3.5. Magnetic sensors. Magnetic sensors use the Hall effect for sensing (Figure 4 f ). Unlike
resistive sensors, magnetic sensors are low cost and not prone to hysteresis. For instance, a flexible
printed circuit board with an integrated magnetic sensor and small magnet can be embedded in
the neutral midplane of a planar soft robot; when the robot is flexed, the magnet moves relative to
the sensor, resulting in a measurable change of magnetic field (119). Although these sensors offer
hysteresis-free measurements and compact form factors, they are vulnerable to the external effects
of metal or magnets, and they cross-talk with other nearby magnetic sensors unless they have been
carefully shielded (131). Due to their precision and ability to work at a distance, catheters with
magnetic sensing can be used to locate their tips in vivo and therefore reduce procedure times for
patients (132, 133).

6. APPLICATION AREAS
Soft robots enable complex operations that require navigation in unstructured environments
and safe interaction with delicate objects. They have already presented their potential in assist-
ing search-and-rescue missions (134), exploring underwater habitats (135), manipulating fragile
objects in industrial settings (136), and executing noninvasive medical operations (137). In this sec-
tion, we examine different soft robots that are used for manipulation, exploration, or healthcare
applications.

6.1. Manipulation
Most soft robots can leverage their natural softness and intrinsic compliance without extensive
sensorization (to measure deformation, contact, or force) to manipulate fragile objects or interact
with their surroundings. Manipulation with soft robots can be achieved in three distinct ways: by
controlling actuation, adhesion, or stiffness (138) (Figure 5a). Most soft robotic actuation modal-
ities can be applied to manipulate objects by controlling the actuation of soft robots (7, 24, 91,
145). However, when soft robots operate based on controlled adhesion, they should have special
structural elements (e.g., gecko-inspired pillars) that can interact with the target object (146) or
generate attractive physical forces (e.g., electroadhesion forces) between their surface and the tar-
get (139). Additionally, when soft robots manipulate objects by changing their stiffness, they can
be fabricated using different actuation modalities and have different configurations. In this group

www.annualreviews.org • An Overview of Soft Robotics 17

 a Manipulation b Exploration
i ii i ii

c Healthcare
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

i ii iii
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Figure 5
Common soft robotic applications. (a) Manipulation. Soft robots can be used to manipulate delicate objects due to their softness and
compliance. In addition, their manipulation performance can be enhanced by strengthening their interaction with the target object.
This improved interaction can be achieved by controlling their adhesion to the target object’s surface (subpanel i) or by tuning their
stiffness according to the target object’s weight (subpanel ii). Subpanel i reproduced with permission from Reference 139; copyright
2015 John Wiley and Sons. Subpanel ii reproduced with permission from Reference 140; copyright 2019 John Wiley and Sons.
(b) Exploration. Soft robots allow us to explore our world due to their ability to safely interact with their surroundings, both underwater
(subpanel i) and on land (subpanel ii), without creating external stress on living beings. Subpanel i reproduced with permission from
Reference 135; copyright 2018 The American Association for the Advancement of Science. Subpanel ii reproduced with permission
from Reference 141; copyright 2021 The American Association for the Advancement of Science. (c) Healthcare. Soft robots can be
successfully used in several different medical areas, including prosthetics (subpanel i), catheters (142) (subpanel ii), and drug delivery
(subpanel iii), due to their mechanical resemblance to natural tissues and high-DOF operation capabilities. Subpanel i reproduced with
permission from Reference 143; copyright 2021 Springer Nature. Subpanel ii reproduced with permission from Wyss Zurich (Project
Nanoflex); copyright 2021 Wyss Zurich. Subpanel iii reproduced with permission from Reference 144; copyright 2021 John Wiley and
Sons. Abbreviation: DOF, degree of freedom.

of soft robots, the ones that work with a jamming mechanism are the most promising for manipu-
lating high-load objects with a minimal applied force (136). In general, soft robots have been used
to gently manipulate delicate, lightweight objects. Nonetheless, most soft robots cannot manipu-
late heavy objects. This issue can be tackled by designing soft robots that can tune their stiffness
according to the weights of their target objects (140). With this strategy, it is possible to enhance
soft robots’ performance and widen their application areas.

6.2. Exploration
Soft robots, especially ones that have bioinspired designs (135), facilitate the exploration of our
world by monitoring or interacting with living beings in their surroundings. Their biomimicry al-
lows them to reach and follow their living surveillance targets without causing them apparent stress
or threatening them (135) (Figure 5b, subpanel i). In addition, these robots can be designed to
camouflage themselves by changing their colors according to their surrounding environment (147,
148). Moreover, they can work under extreme environmental conditions, e.g., with large tem-
perature variations (149, 150) or extreme pressures (151), without any structural modification.

18 Yasa et al.
 Furthermore, their intrinsic compliance ensures that they can navigate with a limited control
input in unpredictable environments (e.g., debris) during search-and-rescue missions (134). In
general, the tethered nature of most soft robots limits their use for exploration applications. As a
result, several research works have focused on creating soft robots that can be remotely operated
(Figure 5b, subpanel ii). In this regard, batteries are often integrated into soft robots as an energy
source (135). Apart from batteries, monopropellant fuels, which can be supplied through microflu-
idic logic gates, are also used to power these robots (152). Fuel cells served as energy storage and
fluidic actuator media in a swimming soft robot (153).

6.3. Healthcare
The compliant nature of soft robots enables their use in several healthcare applications, including
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

rehabilitation, minimally invasive surgery, and targeted drug delivery. In rehabilitation, soft robots
provide safe cooperation with patients and assist them during recovery. Soft robots’ inherent soft-
ness and compliance allow them to follow motion paths that are kinematically similar to humans’.
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Their softness also allows them to conform to patients’ bodies and prevent the application of
nonphysiological loads that could damage the musculoskeletal system during rehabilitation.
While soft robots have been developed for rehabilitation applications for most major joints
of the body, most efforts have been devoted to restoring hand function. Soft hand devices are
often shaped like gloves, supporting finger bending to execute rehabilitation exercises or assis-
tive tasks (143) (Figure 5c, subpanel i). These gloves have been driven by different mechanisms,
including fluidic, shape-memory, and electrostatic actuators.
In minimally invasive surgery, soft robots can improve the dexterity, flexibility, and maneuver-
ability of surgical tools, thereby reducing the size and number of incisions that are required to
perform an operation (Figure 5c, subpanel ii). In general, minimally invasive surgeries involve the
use of two or three semirigid tools, assisted by a rigid endoscope for vision; however, the use of
multiple access ports leads to challenges in tool encumbrance and triangulation. Soft robots can be
introduced through a single incision or natural orifice and navigate around organs and anatomic
structures to reach the surgical target site. Soft surgical robots that use various actuation mecha-
nisms have been demonstrated in proof-of-concept studies, including fluidic (154), magnetic (142,
155), and shape-memory-driven (156) robots. These robots have also been applied to several ex
vivo models and in preclinical animal trials.
In drug delivery, soft robots can be used to deliver therapies to remote inner parts of the brain,
liver, or pancreas by avoiding their accumulation in nontargeted body regions and decreasing
their off-target distribution (144, 157). Reaching these regions requires navigating through the
vasculature or hollow organs, which cannot be easily achieved with wired surgical tools. Soft robots
can be small enough to deliver biological agents on demand in both tethered and untethered
fashions. These soft robots are manipulated through external stimuli, such as magnetic fields,
ultrasound, or temperature (Figure 5c, subpanel iii).

7. CONCLUSION AND FUTURE PERSPECTIVES

The great potential of soft robotics to revolutionize our daily lives is already becoming evident
from the performance of the existing soft robots in delicate, unstructured environments. However,
soft robots cannot yet fully replicate the complex functions that can be observed in nature, such
as the smooth, high-DOF movement of an elephant trunk, the adaptation of cells to different en-
vironmental conditions, and the self-healing of tissues following an injury. Therefore, soft robots
have not yet revealed their full potential to surpass the performance of existing rigid-bodied robots
and to present the complex behaviors that can be seen in living beings.

www.annualreviews.org • An Overview of Soft Robotics 19

 a Dedicated controllers and stable b Operation mode and lifetime c Eco-friendliness and sustainability
interfaces
Stable interface Light Electricity

Soft actuator

Semirigid Contraction
sensor

Rigid
camera Self- Muscular Exploration
healing layer

Dedicated controller
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

Myotube
Progenitor cell
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Nucleus Biodegradation

Figure 6
Potential areas of focus for future studies to increase understanding of the principles behind the design and operation of existing soft
robots and tackle issues that limit their performance under real-world conditions. This figure highlights three key issues. (a) Dedicated
controllers and stable interfaces. Soft robots’ performance should be improved with new stable mechanical interfaces between the soft
and rigid components and dedicated control architectures for precise human–robot interactions. (b) Operation mode and lifetime. The
robots’ operational lifetime should be enhanced by using, for example, self-healing, energy-efficient biological constructs; such
constructs will need to be highly integrated and require new types of operation modes, such as optical or electrical cell stimulation.
(c) Eco-friendliness and sustainability. Soft robots should be eco-friendly and operate in close proximity to living beings without
creating a negative impact on them. The robots should also be degradable without leaving toxic by-products in the environment.

To advance soft robotics, future studies should focus more on understanding the principles
behind the design and operation of the existing soft robots (158) and tackle the major issues that
limit their performance and translation from laboratories to real-world conditions. In this regard,
future soft robots’ performance should be improved by developing special controllers and stable
interfaces between their soft (e.g., actuators) and rigid (e.g., sensors) components (159). In addi-
tion, soft robots should be compact, integrated platforms, and their operational lifetime should
be enhanced by using durable, self-healing, elastic materials as structural building elements (160).
Only in this way can they compete with their rigid-bodied counterparts in a realistic work environ-
ment. Unlike traditional rigid-bodied robots, soft robots’ entire structural components, including
their electronics and powering units, should eventually be eco-friendly and made from sustainable
materials to eliminate their negative impact on the environment (161) (Figure 6). In the following
subsections, we discuss possible research efforts that can be made to tackle these major issues so
that highly performant, integrated, and sustainable soft robots can be created in the future.

7.1. Dedicated Controllers and Stable Interfaces

The performance of existing soft robots can be significantly enhanced by developing dedicated
power and control systems for them. Unlike traditional rigid-bodied robots, soft robots suffer
from not having power controllers, models, or control architectures that have been customized
for the actuation and sensing modalities that are needed to obtain and track the desired soft body
deformation. Rather than continuously iterating the existing soft robots’ actuator or sensor de-
sign to enhance their performance, future research efforts should try to concentrate on efficiently

20 Yasa et al.
 powering these designs and modeling their dynamics and relevant multiphysics. Suitable and fast
models can then be used to create specialized controllers and learning architectures. Soft robotics
will move forward faster when these architectures have been developed specifically for soft robots,
since this development will lead to more predictable and safer soft robots that are capable of
reliably operating near living beings.
Interfacing between soft and rigid components is a trade-off that affects both the performance
and the operational lifetime of soft robots. The added interfaces provide successful information
and energy transfer between soft robots’ components, such as actuators, controllers, sensors, and
power units. Additionally, added rigid components strengthen soft robots’ mechanical stability
during actuation and operation. Unlike in traditional rigid-bodied robotics, there are no well-
characterized, standardized, and stable interfaces in soft robotics that can be readily used to create
soft robots. For this reason, it is not yet possible to systematically design and mass manufacture
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

soft robots that can exceed the performance of both traditional rigid-bodied robots and living
creatures. For instance, better interfacing would allow us to properly integrate sensors through-
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

out soft robots’ bodies for proprioception and tactile sensing. Dense proprioception and tactile
sensing would lead to a more accurate perception of soft robots’ deformation states, improved
control, and smoother interaction with their surroundings. Such a capability would promote the
safety of soft robots’ interaction with their surroundings. As another example, stable interfacing
would advance biohybrid soft robotics by increasing the performance of tissue-engineered muscle
constructs (145). Tendon-like structures could be used to connect soft, living muscle constructs
with the semirigid components of soft robots to obtain a better mechanical advantage and
deformation rate.

7.2. Operation Mode and Lifetime

The term compactness refers to how small we can manufacture an integrated platform. Soft robots’
compactness determines their operation areas and modes, i.e., tethered or untethered. For exam-
ple, most fluidic soft robots that have been developed for manipulation tasks are bulky and tethered
because they require fluidic pressure for actuation. Commercial electronic components have al-
ready allowed researchers to manufacture untethered fluidic soft robots that can locomote both
on land (149) and underwater (135). By contrast, soft robots that are powered with high-voltage
electricity are mostly tethered and not compact due to the lack of small electronics that are nec-
essary for high voltage generation. Fortunately, recent research efforts have focused on designing
and manufacturing centimeter-scale, battery-powered high-voltage power suppliers that can en-
able the creation of untethered electrostatic soft robots (162). As a final example, even though
biohybrid soft robots usually require an electrical input for actuation, recent biohybrid soft robots
have been able to operate in an untethered fashion by using optogenetically engineered muscle
cells that react to light stimulation (40, 59) or by integrating optogenetically engineered nerve
cells that can control muscle contraction (64).
The operational lifetime is another important point that should be considered when fabricat-
ing soft robots. Soft robots that operate in an untethered fashion have a limited lifetime due to
the power storage capacity of existing battery technologies. Therefore, alternative green energy
sources, such as solar energy (163), should be considered when designing soft robots that will
perform operations outside the laboratory in the future. In addition to powering limitations, soft
robots are usually prone to being damaged and ruptured due to their soft materials’ properties.
For example, fluidic soft robots are no longer able to operate when a sharp or pointy object hits
them. In addition, when the uncontrolled internal fluidic pressure increases, their soft bodies can
rupture. In these cases, it would be advantageous for the robots to be built using self-healing elastic

www.annualreviews.org • An Overview of Soft Robotics 21

 materials as structural building elements because robots manufactured from such materials would
be able to recover their functions and continue to operate (164, 165). As another example, most
electrostatic soft robots stop working after several thousand cycles due to dielectric breakdown.
The ones that operate with liquid dielectric materials can survive after many dielectric breakdowns
due to the self-healing ability of liquid dielectric materials. For example, donut HASEL robots can
operate for more than 1 million cycles while lifting a 150-g object at a 15% strain (24). Finally,
most biohybrid robots lose their functionalities after a couple of days due to fatigue and cell dam-
age, which results from an insufficient supply of nutrients and oxygen. However, the operational
lifetime of these robots could be increased by integrating vessel-like structures into their designs
to supply proper nutrients and oxygen. In addition, cells of nonmammalian origin that can survive
under ambient room conditions without any carbon dioxide supply and are resistant to extreme
environmental variations could potentially be used as actuators in soft robotics (66, 166).
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

7.3. Eco-Friendliness and Sustainability

Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Most existing soft robots have a negative impact on our ecosystem due to their building materials
and power sources. In the future, new strategies should be developed to reduce the carbon foot-
print of soft robots in order to minimize their negative impact on the environment after disposal.
Therefore, future soft robots should be sustainable. For example, they could be manufactured
from recyclable plastic materials. In addition, biodegradable natural elastic materials could be
used for fabrication, since these materials do not turn into environmentally toxic side products
after degradation (161). Moreover, we should consider recycling electronic components, or we
could fabricate robots that are able to operate without any electronics (141). In addition, we could
use natural sources, such as solar energy (163) and fuel cells (153), to power our robots. Finally,
we should consider silent actuation modalities, such as electrostatic and biohybrid ones, when de-
signing soft robots for environmental operations. The constant noise of machines is a problem
for living beings everywhere—both in the oceans and on land. For instance, humans who are con-
stantly exposed to noise can become stressed or depressed or even develop tinnitus (167). In this
regard, we should overhaul our approach to enhancing the eco-friendliness of future soft robots
by considering their materials’ properties and their often overlooked and fixable adverse effects
while operating next to living beings.

DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.

AUTHOR CONTRIBUTIONS
R.K.K. and O.Y. conceived the original concept and manuscript outline, wrote the manuscript,
generated figures, and edited the text. Y.T. contributed to Sections 4 and 5, generated figures,
and edited the text. M.Y.M. contributed to Section 3, generated figures, and edited the text. L.S.J.
contributed to Section 5.3. M.F. contributed to Section 2.5. T.B. contributed to Section 2.2.

LITERATURE CITED
1. Kim S, Laschi C, Trimmer B. 2013. Soft robotics: a bioinspired evolution in robotics. Trends Biotechnol.
31:287–94
2. Rus D, Tolley MT. 2015. Design, fabrication and control of soft robots. Nature 521:467–75
3. Majidi C. 2019. Soft-matter engineering for soft robotics. Adv. Mater. Technol. 4:1800477

22 Yasa et al.
 4. Amend J, Cheng N, Fakhouri S, Culley B. 2016. Soft robotics commercialization: jamming grippers
from research to product. Soft Robot. 3:213–22
5. Jørgensen TB, Jensen SHN, Aanæs H, Hansen NW, Krüger N. 2019. An adaptive robotic system for
doing pick and place operations with deformable objects. J. Intell. Robot. Syst. 94:81–100
6. Wang Z, Kanegae R, Hirai S. 2021. Circular shell gripper for handling food products. Soft Robot. 8:542–54
7. Galloway KC, Becker KP, Phillips B, Kirby J, Licht S, et al. 2016. Soft robotic grippers for biological
sampling on deep reefs. Soft Robot. 3:23–33
8. Kurumaya S, Phillips BT, Becker KP, Rosen MH, Gruber DF, et al. 2018. A modular soft robotic wrist
for underwater manipulation. Soft Robot. 5:399–409
9. Vogt DM, Becker KP, Phillips BT, Graule MA, Rotjan RD, et al. 2018. Shipboard design and fabrication
of custom 3D-printed soft robotic manipulators for the investigation of delicate deep-sea organisms.
PLOS ONE 13:e0200386
10. Bernth JE, Arezzo A, Liu H. 2017. A novel robotic meshworm with segment-bending anchoring for
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

colonoscopy. IEEE Robot. Autom. Lett. 2:1718–24

11. Abidi H, Gerboni G, Brancadoro M, Fras J, Diodato A, et al. 2018. Highly dexterous 2-module soft
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

robot for intra-organ navigation in minimally invasive surgery. Int. J. Med. Robot. 14:e1875
12. Hu W, Lum GZ, Mastrangeli M, Sitti M. 2018. Small-scale soft-bodied robot with multimodal
locomotion. Nature 554:81–85
13. Martinez RV, Branch JL, Fish CR, Jin L, Shepherd RF, et al. 2013. Robotic tentacles with
three-dimensional mobility based on flexible elastomers. Adv. Mater. 25:205–12
14. Mosadegh B, Polygerinos P, Keplinger C, Wennstedt S, Shepherd RF, et al. 2014. Pneumatic networks
for soft robotics that actuate rapidly. Adv. Funct. Mater. 24:2163–70
15. Onal CD, Rus D. 2013. Autonomous undulatory serpentine locomotion utilizing body dynamics of a
fluidic soft robot. Bioinspir. Biomim. 8:026003
16. Marchese AD, Katzschmann RK, Rus D. 2015. A recipe for soft fluidic elastomer robots. Soft Robot.
2:7–25
17. Polygerinos P, Correll N, Morin SA, Mosadegh B, Onal CD, et al. 2017. Soft robotics: review of
fluid-driven intrinsically soft devices; manufacturing, sensing, control, and applications in human-robot
interaction. Adv. Eng. Mater. 19:1700016
18. Pelrine RE, Kornbluh RD, Joseph JP. 1998. Electrostriction of polymer dielectrics with compliant
electrodes as a means of actuation. Sens. Actuators A 64:77–85
19. Perju E, Ko YS, Dünki SJ, Opris DM. 2020. Increased electromechanical sensitivity of polysiloxane
elastomers by chemical modification with thioacetic groups. Mater. Des. 186:108319
20. Huang J, Lu T, Zhu J, Clarke DR, Suo Z. 2012. Large, uni-directional actuation in dielectric elastomers
achieved by fiber stiffening. Appl. Phys. Lett. 100:211901
21. Gupta U, Qin L, Wang Y, Godaba H, Zhu J. 2019. Soft robots based on dielectric elastomer actuators:
a review. Smart Mater. Struct. 28:103002
22. Youn JH, Jeong SM, Hwang G, Kim H, Hyeon K, et al. 2020. Dielectric elastomer actuator for soft
robotics applications and challenges. Appl. Sci. 10:640
23. Iacob M, Verma A, Buchner T, Sheima Y, Katzschmann R, Opris DM. 2021. Slot-die coating of an
on-the-shelf polymer with increased dielectric permittivity for stack actuators. ACS Appl. Polym. Mater.
4:150–57
24. Acome E, Mitchell SK, Morrissey T, Emmett M, Benjamin C, et al. 2018. Hydraulically amplified self-
healing electrostatic actuators with muscle-like performance. Science 359:61–65
25. Rothemund P, Kellaris N, Mitchell SK, Acome E, Keplinger C. 2021. HASEL artificial muscles for a
new generation of lifelike robots—recent progress and future opportunities. Adv. Mater. 33:2003375
26. Kellaris N, Gopaluni Venkata V, Smith GM, Mitchell SK, Keplinger C. 2018. Peano-HASEL actuators:
muscle-mimetic, electrohydraulic transducers that linearly contract on activation. Sci. Robot. 3:eaar3276
27. Leroy E, Hinchet R, Shea H. 2020. Multimode hydraulically amplified electrostatic actuators for
wearable haptics. Adv. Mater. 32:2002564
28. Helps T, Romero C, Taghavi M, Conn AT, Rossiter J. 2022. Liquid-amplified zipping actuators for
micro-air vehicles with transmission-free flapping. Sci. Robot. 7:eabi8189

www.annualreviews.org • An Overview of Soft Robotics 23

 29. O’Neill MR, Acome E, Bakarich S, Mitchell SK, Timko J, et al. 2020. Rapid 3D printing of
electrohydraulic (HASEL) tentacle actuators. Adv. Funct. Mater. 30:2005244
30. Wang J, Gao D, Lee PS. 2021. Recent progress in artificial muscles for interactive soft robotics. Adv.
Mater. 33:2003088
31. Yang Y, Wu Y, Li C, Yang X, Chen W. 2020. Flexible actuators for soft robotics. Adv. Intell. Syst.
2:1900077
32. Lee SJ, Han MJ, Kim SJ, Jho JY, Lee HY, Kim YH. 2006. A new fabrication method for IPMC actuators
and application to artificial fingers. Smart Mater. Struct. 15:1217
33. Mousavi MSS, Alaei A, Hasani M, Kolahdouz M, Manteghi F, Ataei F. 2018. Fabrication of ionic polymer
metal composite for bio-actuation application: sputtering and electroless plating methods. Mater. Res.
Express 6:035312
34. Cianchetti M. 2013. Fundamentals on the use of shape memory alloys in soft robotics. In Interdisciplinary
Mechatronics, ed. MK Habib, JP Davim, pp. 227–54. Hoboken, NJ: Wiley
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

35. Laschi C, Cianchetti M, Mazzolai B, Margheri L, Follador M, Dario P. 2012. Soft robot arm inspired
by the octopus. Adv. Robot. 26:709–27
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

36. Lin HT, Leisk GG, Trimmer B. 2011. GoQBot: a caterpillar-inspired soft-bodied rolling robot. Bioinspir.
Biomim. 6:026007
37. Villanueva A, Smith C, Priya S. 2011. A biomimetic robotic jellyfish (Robojelly) actuated by shape
memory alloy composite actuators. Bioinspir. Biomim. 6:036004
38. Kim HJ, Song SH, Ahn SH. 2012. A turtle-like swimming robot using a smart soft composite (SSC)
structure. Smart Mater. Struct. 22:014007
39. Huang X, Kumar K, Jawed MK, Mohammadi Nasab A, Ye Z, et al. 2019. Highly dynamic shape memory
alloy actuator for fast moving soft robots. Adv. Mater. Technol. 4:1800540
40. Abbott JJ, Diller E, Petruska AJ. 2020. Magnetic methods in robotics. Annu. Rev. Control Robot. Auton.
Syst. 3:57–90
41. Kim Y, Parada GA, Liu S, Zhao X. 2019. Ferromagnetic soft continuum robots. Sci. Robot. 4:eaax7329
42. Kim Y, Zhao X. 2022. Magnetic soft materials and robots. Chem. Rev. 122:5317–64
43. Huang HW, Sakar MS, Petruska AJ, Pané S, Nelson BJ. 2016. Soft micromachines with programmable
motility and morphology. Nat. Commun. 7:12263
44. Lum GZ, Ye Z, Dong X, Marvi H, Erin O, et al. 2016. Shape-programmable magnetic soft matter. PNAS
113:6007–15
45. Ren Z, Hu W, Dong X, Sitti M. 2019. Multi-functional soft-bodied jellyfish-like swimming. Nat.
Commun. 10:2703
46. Kim Y, Yuk H, Zhao R, Chester SA, Zhao X. 2018. Printing ferromagnetic domains for untethered
fast-transforming soft materials. Nature 558:274–79
47. Xu T, Zhang J, Salehizadeh M, Onaizah O, Diller E. 2019. Millimeter-scale flexible robots with
programmable three-dimensional magnetization and motions. Sci. Robot. 4:eaav4494
48. Li M, Wang Y, Chen A, Naidu A, Napier BS, et al. 2018. Flexible magnetic composites for
light-controlled actuation and interfaces. PNAS 115:8119–24
49. Alapan Y, Karacakol AC, Guzelhan SN, Isik I, Sitti M. 2020. Reprogrammable shape morphing of
magnetic soft machines. Sci. Adv. 6:eabc6414
50. Ricotti L, Trimmer B, Feinberg AW, Raman R, Parker KK, et al. 2017. Biohybrid actuators for robotics:
a review of devices actuated by living cells. Sci. Robot. 2:eaaq0495
51. Mazzolai B, Laschi C. 2020. A vision for future bioinspired and biohybrid robots. Sci. Robot. 5:eaba6893
52. Trimmer BA. 2020. Metal or muscle? The future of biologically inspired robots. Sci. Robot. 5:eaba6149
53. Raman R, Grant L, Seo Y, Cvetkovic C, Gapinske M, et al. 2017. Damage, healing, and remodeling in
optogenetic skeletal muscle bioactuators. Adv. Healthc. Mater. 6:1700030
54. Xi J, Schmidt JJ, Montemagno CD. 2005. Self-assembled microdevices driven by muscle. Nat. Mater.
4:180–84
55. Akiyama Y, Hoshino T, Iwabuchi K, Morishima K. 2012. Room temperature operable autonomously
moving bio-microrobot powered by insect dorsal vessel tissue. PLOS ONE 7:e38274
56. Feinberg AW, Feigel A, Shevkoplyas SS, Sheehy S, Whitesides GM, Parker KK. 2007. Muscular thin
films for building actuators and powering devices. Science 317:1366–70

24 Yasa et al.
 57. Nawroth JC, Lee H, Feinberg AW, Ripplinger CM, McCain ML, et al. 2012. A tissue-engineered
jellyfish with biomimetic propulsion. Nat. Biotechnol. 30:792–97
58. Williams BJ, Anand SV, Rajagopalan J, Saif MTA. 2014. A self-propelled biohybrid swimmer at low
Reynolds number. Nat. Commun. 5:3081
59. Park SJ, Gazzola M, Park KS, Park S, Di Santo V, et al. 2016. Phototactic guidance of a tissue-engineered
soft-robotic ray. Science 353:158–62
60. Lee KY, Park SJ, Matthews DG, Kim SL, Marquez CA, et al. 2022. An autonomously swimming
biohybrid fish designed with human cardiac biophysics. Science 375:639–47
61. Cvetkovic C, Raman R, Chan V, Williams BJ, Tolish M, et al. 2014. Three-dimensionally printed
biological machines powered by skeletal muscle. PNAS 111:10125–30
62. Guix M, Mestre R, Patiño T, De Corato M, Fuentes J, et al. 2021. Biohybrid soft robots with self-
stimulating skeletons. Sci. Robot. 6:eabe7577
63. Raman R, Cvetkovic C, Uzel SG, Platt RJ, Sengupta P, et al. 2016. Optogenetic skeletal muscle-powered
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

adaptive biological machines. PNAS 113:3497–502

64. Aydin O, Zhang X, Nuethong S, Pagan-Diaz GJ, Bashir R, et al. 2019. Neuromuscular actuation of
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

biohybrid motile bots. PNAS 116:19841–47

65. Herr H, Dennis RG. 2004. A swimming robot actuated by living muscle tissue. J. NeuroEng. Rehabil. 1:6
66. Uesugi K, Shimizu K, Akiyama Y, Hoshino T, Iwabuchi K, Morishima K. 2016. Contractile performance
and controllability of insect muscle-powered bioactuator with different stimulation strategies for soft
robotics. Soft Robot. 3:13–22
67. Baryshyan A, Domigan L, Hunt B, Trimmer B, Kaplan D. 2014. Self-assembled insect muscle
bioactuators with long term function under a range of environmental conditions. RSC Adv. 4:39962–68
68. Katzschmann RK, Thieffry M, Goury O, Kruszewski A, Guerra TM, et al. 2019. Dynamically closed-
loop controlled soft robotic arm using a reduced order finite element model with state observer. In 2019
2nd IEEE International Conference on Soft Robotics (RoboSoft), pp. 717–24. Piscataway, NJ: IEEE
69. Doroudchi A, Berman S. 2021. Configuration tracking for soft continuum robotic arms using inverse
dynamic control of a Cosserat rod model. In 2019 2nd IEEE International Conference on Soft Robotics
(RoboSoft), pp. 207–14. Piscataway, NJ: IEEE
70. Katzschmann RK, Della Santina C, Toshimitsu Y, Bicchi A, Rus D. 2019. Dynamic motion control of
multi-segment soft robots using piecewise constant curvature matched with an augmented rigid body
model. In 2019 2nd IEEE International Conference on Soft Robotics (RoboSoft), pp. 454–61. Piscataway, NJ:
IEEE
71. Coevoet E, Morales-Bieze T, Largilliere F, Zhang Z, Thieffry M, et al. 2017. Software toolkit for
modeling, simulation, and control of soft robots. Adv. Robot. 31:1208–24
72. Du T, Wu K, Ma P, Wah S, Spielberg A, et al. 2021. DiffPD: Differentiable Projective Dynamics. ACM
Trans. Graph. 41:13
73. Li M, Ferguson Z, Schneider T, Langlois TR, Zorin D, et al. 2020. Incremental potential contact:
intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39:49
74. Hu Y, Liu J, Spielberg A, Tenenbaum JB, Freeman WT, et al. 2019. ChainQueen: a real-time differ-
entiable physical simulator for soft robotics. In 2019 International Conference on Robotics and Automation
(ICRA), pp. 6265–71. Piscataway, NJ: IEEE
75. Allard J, Cotin S, Faure F, Bensoussan PJ, Poyer F, et al. 2007. SOFA—an open source framework for
medical simulation. In Medicine Meets Virtual Reality 15, ed. JD Westwood, RS Haluck, HM Hoffman,
GT Mogel, R Phillips, et al., pp. 13–18. Amsterdam: IOS
76. Goury O, Carrez B, Duriez C. 2021. Real-time simulation for control of soft robots with self-collisions
using model order reduction for contact forces. IEEE Robot. Autom. Lett. 6:3752–59
77. COMSOL. 2021. COMSOL Multiphysics® simulation software. COMSOL. https://www.comsol.
com/comsol-multiphysics
78. Smith M. 2009. ABAQUS/Standard User’s Manual, Version 6.9. Johnston, RI: Dassault Syst. Simulia
79. Verschoor M, Jalba AC. 2019. Efficient and accurate collision response for elastically deformable models.
ACM Trans. Graph. 38:17
80. Ma P, Du T, Zhang JZ, Wu K, Spielberg A, et al. 2021. DiffAqua: a differentiable computational design
pipeline for soft underwater swimmers with shape interpolation. ACM Trans. Graph. 40:132

www.annualreviews.org • An Overview of Soft Robotics 25

 81. Nava E, Zhang JZ, Michelis MY, Du T, Ma P, et al. 2022. Fast aquatic swimmer optimization with
differentiable projective dynamics and neural network hydrodynamic models. In Proceedings of the 39th
International Conference on Machine Learning, ed. K Chaudhuri, S Jegelka, L Song, C Szepesvari, G Niu,
S Sabato, pp. 16413–27. Proc. Mach. Learn. Res. 162. N.p.: PMLR
82. Zhang JZ, Zhang Y, Ma P, Nava E, Du T, et al. 2021. Learning material parameters and hydrodynamics
of soft robotic fish via differentiable simulation. arXiv:2109.14855 [cs.RO]
83. Dubied M, Michelis MY, Spielberg A, Katzschmann RK. 2022. Sim-to-real for soft robots using
differentiable FEM: recipes for meshing, damping, and actuation. IEEE Robot. Autom. Lett. 7:5015–22
84. Naughton N, Sun J, Tekinalp A, Parthasarathy T, Chowdhary G, Gazzola M. 2021. Elastica: a compliant
mechanics environment for soft robotic control. IEEE Robot. Autom. Lett. 6:3389–96
85. Renda F, Boyer F, Dias J, Seneviratne L. 2018. Discrete Cosserat approach for multisection soft
manipulator dynamics. IEEE Trans. Robot. 34:1518–33
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

86. Till J, Aloi V, Rucker C. 2019. Real-time dynamics of soft and continuum robots based on Cosserat rod
models. Int. J. Robot. Res. 38:723–46
87. Grazioso S, Di Gironimo G, Siciliano B. 2019. A geometrically exact model for soft continuum robots:
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

the finite element deformation space formulation. Soft Robot. 6:790–811

88. Della Santina C, Bicchi A, Rus D. 2020. On an improved state parametrization for soft robots with
piecewise constant curvature and its use in model based control. IEEE Robot. Autom. Lett. 5:1001–8
89. Della Santina C, Katzschmann RK, Biechi A, Rus D. 2018. Dynamic control of soft robots interact-
ing with the environment. In 2018 IEEE International Conference on Soft Robotics (RoboSoft), pp. 46–53.
Piscataway, NJ: IEEE
90. Della Santina C, Katzschmann RK, Bicchi A, Rus D. 2020. Model-based dynamic feedback control of a
planar soft robot: trajectory tracking and interaction with the environment. Int. J. Robot. Res. 39:490–513
91. Toshimitsu Y, Wong KW, Buchner T, Katzschmann R. 2021. SoPrA: fabrication & dynamical modeling
of a scalable soft continuum robotic arm with integrated proprioceptive sensing. In 2021 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS), pp. 653–60. Piscataway, NJ: IEEE
92. Huang X, Zou J, Gu G. 2021. Kinematic modeling and control of variable curvature soft continuum
robots. IEEE/ASME Trans. Mechatron. 26:3175–85
93. Della Santina C, Rus D. 2019. Control oriented modeling of soft robots: the polynomial curvature case.
IEEE Robot. Autom. Lett. 5:290–98
94. Singh I, Amara Y, Melingui A, Mani Pathak P, Merzouki R. 2018. Modeling of continuum manipulators
using Pythagorean hodograph curves. Soft Robot. 5:425–42
95. Kazemipour A, Fischer O, Toshimitsu Y, Wong KW, Katzschmann RK. 2022. Adaptive dynamic sliding
mode control of soft continuum manipulators. In 2022 International Conference on Robotics and Automation
(ICRA), pp. 3259–65. Piscataway, NJ: IEEE
96. Thuruthel TG, Falotico E, Renda F, Laschi C. 2019. Model-based reinforcement learning for closed-
loop dynamic control of soft robotic manipulators. IEEE Trans. Robot. 35:124–34
97. Gillespie MT, Best CM, Townsend EC, Wingate D, Killpack MD. 2018. Learning nonlinear dynamic
models of soft robots for model predictive control with neural networks. In 2018 IEEE International
Conference on Soft Robotics (RoboSoft), pp. 39–45. Piscataway, NJ: IEEE
98. Pfaff T, Fortunato M, Sanchez-Gonzalez A, Battaglia P. 2021. Learning mesh-based simulation with
graph networks. In 2021 International Conference on Learning Representations. La Jolla, CA: Int. Conf.
Learn. Represent. https://openreview.net/pdf?id=roNqYL0_XP
99. Soter G, Conn A, Hauser H, Rossiter J. 2018. Bodily aware soft robots: integration of propriocep-
tive and exteroceptive sensors. In 2018 IEEE International Conference on Robotics and Automation (ICRA),
pp. 2448–53. Piscataway, NJ: IEEE
100. Bruder D, Fu X, Gillespie RB, Remy CD, Vasudevan R. 2020. Data-driven control of soft robots using
Koopman operator theory. IEEE Trans. Robot. 37:948–61
101. Han M, Euler-Rolle J, Katzschmann RK. 2022. DeSKO: stability-assured robust control with a deep
stochastic Koopman operator. In 2022 International Conference on Learning Representations. La Jolla, CA:
Int. Conf. Learn. Represent. https://openreview.net/pdf?id=hniLRD_XCA

26 Yasa et al.
 102. Kim JI, Hong M, Lee K, Kim D, Park YL, Oh S. 2020. Learning to walk a tripod mobile robot using
nonlinear soft vibration actuators with entropy adaptive reinforcement learning. IEEE Robot. Autom. Lett.
5:2317–24
103. Zhang B, Liu P. 2021. Model-based and model-free robot control: a review. In RiTA 2020: Proceedings of
the 8th International Conference on Robot Intelligence Technology and Applications, ed. E Chew, APPA Majeed,
P Liu, J Platts, H Myung, et al., pp. 45–55. Singapore: Springer
104. Caasenbrood BJ, Pogromsky AY, Nijmeijer H. 2020. Dynamic modeling of hyper-elastic soft robots
using spatial curves. IFAC-PapersOnLine 53(2):9238–43
105. Jiang H, Wang Z, Jin Y, Chen X, Li P, et al. 2021. Hierarchical control of soft manipulators towards
unstructured interactions. Int. J. Robot. Res. 40:411–34
106. Garriga-Casanovas A, Rodriguez y Baena F. 2019. Kinematics of continuum robots with constant
curvature bending and extension capabilities. J. Mech. Robot. 11:011010
107. Della Santina C, Pallottino L, Rus D, Bicchi A. 2019. Exact task execution in highly under-actuated soft
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

limbs: an operational space based approach. IEEE Robot. Autom. Lett. 4:2508–15
108. Wang H, Yang B, Liu Y, Chen W, Liang X, Pfeifer R. 2016. Visual servoing of soft robot manipulator
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

in constrained environments with an adaptive controller. IEEE/ASME Trans. Mechatron. 22:41–50

109. Fischer O, Toshimitsu Y, Kazemipour A, Katzschmann RK. 2023. Dynamic task space control enables
soft manipulators to perform real-world tasks. Adv. Intell. Syst. 5:2200024
110. Trumić M, Della Santina C, Jovanović K, Fagiolini A. 2021. Adaptive control of soft robots based on an
enhanced 3D augmented rigid robot matching. In 2021 American Control Conference (ACC), pp. 4991–96.
Piscataway, NJ: IEEE
111. Verghese M, Richter F, Gunn A, Weissbrod P, Yip M. 2022. Model-free visual control for continuum
robot manipulators via orientation adaptation. In Robotics Research: The 19th International Symposium
ISRR, ed. T Asfour, E Yoshida, J Park, H Christensen, O Khatib, pp. 959–70. Cham, Switz.: Springer
112. Satheeshbabu S, Uppalapati NK, Chowdhary G, Krishnan G. 2019. Open loop position control of
soft continuum arm using deep reinforcement learning. In 2019 International Conference on Robotics and
Automation (ICRA), pp. 5133–39. Piscataway, NJ: IEEE
113. Navarro SE, Nagels S, Alagi H, Faller LM, Goury O, et al. 2020. A model-based sensor fusion approach
for force and shape estimation in soft robotics. IEEE Robot. Autom. Lett. 5:5621–28
114. Homberg BS, Katzschmann RK, Dogar MR, Rus D. 2019. Robust proprioceptive grasping with a soft
robot hand. Auton. Robots 43:681–96
115. Teyssier M, Parilusyan B, Roudaut A, Steimle J. 2021. Human-like artificial skin sensor for physi-
cal human-robot interaction. In 2021 IEEE International Conference on Robotics and Automation (ICRA),
pp. 3626–33. Piscataway, NJ: IEEE
116. Mirzanejad H, Agheli M. 2019. Soft force sensor made of magnetic powder blended with silicone rubber.
Sens. Actuators A 293:108–18
117. Kuppuswamy N, Alspach A, Uttamchandani A, Creasey S, Ikeda T, Tedrake R. 2020. Soft-bubble grip-
pers for robust and perceptive manipulation. In 2020 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), pp. 9917–24. Piscataway, NJ: IEEE
118. Shih B, Christianson C, Gillespie K, Lee S, Mayeda J, et al. 2019. Design considerations for 3D printed,
soft, multimaterial resistive sensors for soft robotics. Front. Robot. AI 6:30
119. Ozel S, Skorina EH, Luo M, Tao W, Chen F, et al. 2016. A composite soft bending actuation module with
integrated curvature sensing. In 2016 IEEE International Conference on Robotics and Automation (ICRA),
pp. 4963–68. Piscataway, NJ: IEEE
120. Nguyen PH, Sridar S, Zhang W, Polygerinos P. 2017. Design and control of a 3-chambered fiber
reinforced soft actuator with off-the-shelf stretch sensors. Int. J. Intell. Robot. Appl. 1:342–51
121. Georgopoulou A, Clemens F. 2022. Pellet-based fused deposition modeling for the development of soft
compliant robotic grippers with integrated sensing elements. Flex. Print. Electron. 7:025010
122. Truby RL, Della Santina C, Rus D. 2020. Distributed proprioception of 3D configuration in soft,
sensorized robots via deep learning. IEEE Robot. Autom. Lett. 5:3299–306
123. Tapia J, Knoop E, Mutn M, Otaduy MA, Bächer M. 2020. MakeSense: automated sensor design for
proprioceptive soft robots. Soft Robot. 7:332–45

www.annualreviews.org • An Overview of Soft Robotics 27

 124. Chossat JB, Park YL, Wood RJ, Duchaine V. 2013. A soft strain sensor based on ionic and metal liquids.
IEEE Sens. J. 13:3405–14
125. Truby RL, Wehner M, Grosskopf AK, Vogt DM, Uzel SG, et al. 2018. Soft somatosensitive actuators
via embedded 3D printing. Adv. Mater. 30:1706383
126. Wang H, Zhang R, Chen W, Liang X, Pfeifer R. 2016. Shape detection algorithm for soft manipulator
based on fiber Bragg gratings. IEEE/ASME Trans. Mechatron. 21:2977–82
127. Zhang Z, Wang X, Wang S, Meng D, Liang B. 2018. Shape detection and reconstruction of soft robotic
arm based on fiber Bragg grating sensor array. In 2018 IEEE International Conference on Robotics and
Biomimetics (ROBIO), pp. 978–83. Piscataway, NJ: IEEE
128. Sefati S, Murphy RJ, Alambeigi F, Pozin M, Iordachita I, et al. 2018. FBG-based control of a continuum
manipulator interacting with obstacles. In 2018 IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS), pp. 6477–83. Piscataway, NJ: IEEE
129. Zhuang W, Sun G, Li H, Lou X, Dong M, Zhu L. 2018. FBG based shape sensing of a silicone octopus
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

tentacle model for soft robotics. Optik 165:7–15

130. Monet F, Sefati S, Lorre P, Poiffaut A, Kadoury S, et al. 2020. High-resolution optical fiber shape sens-
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

ing of continuum robots: a comparative study. In 2020 IEEE International Conference on Robotics and
Automation (ICRA), pp. 8877–83. Piscataway, NJ: IEEE
131. Walck G, Perdereau V. 2014. Software compensation of magnetic crosstalk on Hall-effect-based ro-
tary encoders close together. In 2014 IEEE International Conference on Robotics and Automation (ICRA),
pp. 4906–11. Piscataway, NJ: IEEE
132. Khaykin Y, Oosthuizen R, Zarnett L, Wulffhart ZA, Whaley B, et al. 2011. CARTO-guided vs. NavX-
guided pulmonary vein antrum isolation and pulmonary vein antrum isolation performed without 3-D
mapping: effect of the 3-D mapping system on procedure duration and fluoroscopy time. J. Interv. Card.
Electrophysiol. 30:233–40
133. Khaykin Y, Zarnett L, Friedlander D, Wulffhart ZA, Whaley B, et al. 2012. Point-by-point pulmonary
vein antrum isolation guided by intracardiac echocardiography and 3D mapping and duty-cycled mul-
tipolar AF ablation: effect of multipolar ablation on procedure duration and fluoroscopy time. J. Interv.
Card. Electrophysiol. 34:303–10
134. der Maur PA, Djambazi B, Haberthür Y, Hörmann P, Kübler A, et al. 2021. RoBoa: construction
and evaluation of a steerable vine robot for search and rescue applications. In 2021 IEEE International
Conference on Soft Robotics (RoboSoft), pp. 15–20. Piscataway, NJ: IEEE
135. Katzschmann RK, DelPreto J, MacCurdy R, Rus D. 2018. Exploration of underwater life with an
acoustically controlled soft robotic fish. Sci. Robot. 3:eaar3449
136. Brown E, Rodenberg N, Amend J, Mozeika A, Steltz E, et al. 2010. Universal robotic gripper based on
the jamming of granular material. PNAS 107:18809–14
137. Gorissen B, De Volder M, Reynaerts D. 2018. Chip-on-tip endoscope incorporating a soft robotic
pneumatic bending microactuator. Biomed. Microdevices 20:73
138. Shintake J, Cacucciolo V, Floreano D, Shea H. 2018. Soft robotic grippers. Adv. Mater. 30:1707035
139. Shintake J, Rosset S, Schubert B, Floreano D, Shea H. 2016. Versatile soft grippers with intrinsic
electroadhesion based on multifunctional polymer actuators. Adv. Mater. 28:231–38
140. Zhang YF, Zhang N, Hingorani H, Ding N, Wang D, et al. 2019. Fast-response, stiffness-tunable soft
actuator by hybrid multimaterial 3D printing. Adv. Funct. Mater. 29:1806698
141. Drotman D, Jadhav S, Sharp D, Chan C, Tolley MT. 2021. Electronics-free pneumatic circuits for
controlling soft-legged robots. Sci. Robot. 6:eaay2627
142. Chautems C, Tonazzini A, Boehler Q, Jeong SH, Floreano D, Nelson BJ. 2020. Magnetic continuum
device with variable stiffness for minimally invasive surgery. Adv. Intell. Syst. 2:1900086
143. Gu G, Zhang N, Xu H, Lin S, Yu Y, et al. 2021. A soft neuroprosthetic hand providing simultaneous myo-
electric control and tactile feedback. Nat. Biomed. Eng. https://doi.org/10.1038/s41551-021-00767-0
144. Fusco S, Sakar MS, Kennedy S, Peters C, Bottani R, et al. 2014. An integrated microrobotic platform
for on-demand, targeted therapeutic interventions. Adv. Mater. 26:952–57
145. Morimoto Y, Onoe H, Takeuchi S. 2018. Biohybrid robot powered by an antagonistic pair of skeletal
muscle tissues. Sci. Robot. 3:eaat4440

28 Yasa et al.
 146. Song S, Sitti M. 2014. Soft grippers using micro-fibrillar adhesives for transfer printing. Adv. Mater.
26:4901–6
147. Morin SA, Shepherd RF, Kwok SW, Stokes AA, Nemiroski A, Whitesides GM. 2012. Camouflage and
display for soft machines. Science 337:828–32
148. Won P, Kim KK, Kim H, Park JJ, Ha I, et al. 2021. Transparent soft actuators/sensors and camouflage
skins for imperceptible soft robotics. Adv. Mater. 33:2002397
149. Tolley MT, Shepherd RF, Galloway KC, Wood RJ, Whitesides GM, et al. 2014. A resilient, untethered
soft robot. Soft Robot. 1:213–23
150. Li T, Li G, Liang Y, Cheng T, Dai J, et al. 2017. Fast-moving soft electronic fish. Sci. Adv. 3:e1602045
151. Li G, Chen X, Zhou F, Liang Y, Xiao Y, et al. 2021. Self-powered soft robot in the Mariana Trench.
Nature 591:66–71
152. Wehner M, Truby RL, Fitzgerald DJ, Mosadegh B, Whitesides GM, et al. 2016. An integrated design
and fabrication strategy for entirely soft, autonomous robots. Nature 536:451–55
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

153. Aubin CA, Choudhury S, Jerch R, Archer LA, Pikul JH, Shepherd RF. 2019. Electrolytic vascular systems
for energy-dense robots. Nature 571:51–57
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

154. Ranzani T, Cianchetti M, Gerboni G, De Falco I, Menciassi A. 2016. A soft modular manipulator for
minimally invasive surgery: design and characterization of a single module. IEEE Trans. Robot. 32:187–
200
155. Lussi J, Mattmann M, Sevim S, Grigis F, De Marco C, et al. 2021. A submillimeter continuous variable
stiffness catheter for compliance control. Adv. Sci. 8:2101290
156. Kobayashi T, Matsunaga T, Haga Y. 2016. Active bending electric endoscope using shape memory alloy
wires. In New Trends in Medical and Service Robots, ed. H Bleuler, M Bouri, F Mondada, D Pisla, A Rodic,
P Helmer, pp. 131–39. Cham, Switz.: Springer
157. Mair LO, Adam G, Chowdhury S, Davis A, Arifin DR, et al. 2021. Soft capsule magnetic millirobots for
region-specific drug delivery in the central nervous system. Front. Robot. AI 26:226
158. Hawkes EW, Majidi C, Tolley MT. 2021. Hard questions for soft robotics. Sci. Robot. 6:eabg6049
159. Rothemund P, Kim Y, Heisser RH, Zhao X, Shepherd RF, Keplinger C. 2021. Shaping the future of
robotics through materials innovation. Nat. Mater. 20:1582–87
160. Terryn S, Langenbach J, Roels E, Brancart J, Bakkali-Hassani C, et al. 2021. A review on self-healing
polymers for soft robotics. Mater. Today 47:187–205
161. Hartmann F, Baumgartner M, Kaltenbrunner M. 2021. Becoming sustainable, the new frontier in soft
robotics. Adv. Mater. 33:2004413
162. Mitchell SK, Martin T, Keplinger C. 2022. A pocket-sized ten-channel high voltage power supply for
soft electrostatic actuators. Adv. Mater. Technol. 7:2101469
163. Mirvakili SM, Leroy A, Sim D, Wang EN. 2021. Solar-driven soft robots. Adv. Sci. 8:2004235
164. Terryn S, Brancart J, Lefeber D, Van Assche G, Vanderborght B. 2017. Self-healing soft pneumatic
robots. Sci. Robot. 2:eaan4268
165. Pena-Francesch A, Jung H, Demirel MC, Sitti M. 2020. Biosynthetic self-healing materials for soft
machines. Nat. Mater. 19:1230–35
166. Dong X, Kheiri S, Lu Y, Xu Z, Zhen M, Liu X. 2021. Toward a living soft microrobot through
optogenetic locomotion control of Caenorhabditis elegans. Sci. Robot. 6:eabe3950
167. Rauschecker JP, Leaver AM, Mühlau M. 2010. Tuning out the noise: limbic-auditory interactions in
tinnitus. Neuron 66:819–26

www.annualreviews.org • An Overview of Soft Robotics 29

AS06_TOC ARjats.cls January 25, 2023 10:59

Annual Review of
Control, Robotics,
and Autonomous
Contents Systems

Volume 6, 2023

An Overview of Soft Robotics

Oncay Yasa, Yasunori Toshimitsu, Mike Y. Michelis, Lewis S. Jones,
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

Miriam Filippi, Thomas Buchner, and Robert K. Katzschmann p p p p p p p p p p p p p p p p p p p p p p p p p p 1

Soft Actuators and Robots Enabled by Additive Manufacturing
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

Dong Wang, Jinqiang Wang, Zequn Shen, Chengru Jiang, Jiang Zou,
Le Dong, Nicholas X. Fang, and Guoying Gu p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p31
Adaptive Control and Intersections with Reinforcement Learning
Anuradha M. Annaswamy p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p65
On the Timescales of Embodied Intelligence for Autonomous
Adaptive Systems
Fumiya Iida and Fabio Giardina p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p95
Toward a Theoretical Foundation of Policy Optimization for
Learning Control Policies
Bin Hu, Kaiqing Zhang, Na Li, Mehran Mesbahi, Maryam Fazel,
and Tamer Başar p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 123
Sequential Monte Carlo: A Unified Review
Adrian G. Wills and Thomas B. Schön p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 159
Construction Robotics: From Automation to Collaboration
Stefana Parascho p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 183
Embodied Communication: How Robots and People Communicate
Through Physical Interaction
Aleksandra Kalinowska, Patrick M. Pilarski, and Todd D. Murphey p p p p p p p p p p p p p p p p p p p p p 205
The Many Facets of Information in Networked Estimation
and Control
Massimo Franceschetti, Mohammad Javad Khojasteh, and Moe Z. Win p p p p p p p p p p p p p p p p p 233
Crowd Dynamics: Modeling and Control of Multiagent Systems
Xiaoqian Gong, Michael Herty, Benedetto Piccoli, and Giuseppe Visconti p p p p p p p p p p p p p p p p 261
Noise in Biomolecular Systems: Modeling, Analysis,
and Control Implications
Corentin Briat and Mustafa Khammash p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 283
AS06_TOC ARjats.cls January 25, 2023 10:59

Exploiting Liquid Surface Tension in Microrobotics

Antoine Barbot, Francisco Ortiz, Aude Bolopion, Michaël Gauthier,
and Pierre Lambert p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 313
Spacecraft-Mounted Robotics
Panagiotis Tsiotras, Matthew King-Smith, and Lorenzo Ticozzi p p p p p p p p p p p p p p p p p p p p p p p p p 335
Grasp Learning: Models, Methods, and Performance
Robert Platt p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 363
Control of Multicarrier Energy Systems from Buildings to Networks
Roy S. Smith, Varsha Behrunani, and John Lygeros p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 391
Annu. Rev. Control Robot. Auton. Syst. 2023.6:1-29. Downloaded from www.annualreviews.org

Control of Low-Inertia Power Systems

Florian Dörfler and Dominic Groß p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 415
Access provided by 189.28.85.70 on 10/21/23. See copyright for approved use.

How the CYBATHLON Competition Has Advanced

Assistive Technologies
Lukas Jaeger, Roberto de Souza Baptista, Chiara Basla, Patricia Capsi-Morales,
Yong Kuk Kim, Shuro Nakajima, Cristina Piazza, Michael Sommerhalder,
Luca Tonin, Giacomo Valle, Robert Riener, and Roland Sigrist p p p p p p p p p p p p p p p p p p p p p p p p 447
Into the Robotic Depths: Analysis and Insights from the DARPA
Subterranean Challenge
Timothy H. Chung, Viktor Orekhov, and Angela Maio p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 477

Errata
An online log of corrections to Annual Review of Control, Robotics, and Autonomous
Systems articles may be found at http://www.annualreviews.org/errata/control