SCIENCE ROBOTICS | RESEARCH ARTICLE

SOFT ROBOTS Copyright © 2023 The

Authors, some
Control of soft robots with inertial dynamics rights reserved;
exclusive licensee
American Association
David A. Haggerty1†*, Michael J. Banks1†, Ervin Kamenar1,2†, Alan B. Cao3, Patrick C. Curtis1, for the Advancement
Igor Mezić1, Elliot W. Hawkes1 of Science. No claim
to original U.S.
Soft robots promise improved safety and capability over rigid robots when deployed near humans or in Government Works
complex, delicate, and dynamic environments. However, infinite degrees of freedom and the potential for
highly nonlinear dynamics severely complicate their modeling and control. Analytical and machine learning
methodologies have been applied to model soft robots but with constraints: quasi-static motions, quasi-
linear deflections, or both. Here, we advance the modeling and control of soft robots into the inertial, nonlinear
regime. We controlled motions of a soft, continuum arm with velocities 10 times larger and accelerations 40
times larger than those of previous work and did so for high-deflection shapes with more than 110° of curvature.
We leveraged a data-driven learning approach for modeling, based on Koopman operator theory, and we intro-
duce the concept of the static Koopman operator as a pregain term in optimal control. Our approach is rapid,
requiring less than 5 min of training; is computationally low cost, requiring as little as 0.5 s to build the model;
and is design agnostic, learning and accurately controlling two morphologically different soft robots. This work
advances rapid modeling and control for soft robots from the realm of quasi-static to inertial, laying the ground-
work for the next generation of compliant and highly dynamic robots.

Downloaded from https://www.science.org on August 30, 2023

INTRODUCTION robots in the past, but these have yet to achieve the real-time control
The automation and robotics revolution has transformed manufac- of fast, inertial motions (17–19).
turing and heavy industry, leading to higher throughput, repeatabil- In the ML modeling of soft robot dynamical behaviors, many
ity, and quality across numerous sectors (1, 2). Unfortunately, neural net–based approaches exist. Most of this work focuses on
robots are most often relegated to cages and isolated sections of the development of predictors using neural nets such as long
manufacturing sites because of the inherent danger they present short-term memory (20, 21) or recurrent neural networks (17,
to human operators through their fast-moving, heavy, and rigid 19). These methodologies generate highly accurate predictors of
structures. Efforts toward allowing these robots to perform safely the dynamics. However, training these systems has a high compu-
with human collaborators have focused on software control, but ab- tational cost. Moreover, their structure is nonlinear, requiring spe-
solute guarantees of safety are not possible (3–6). cialized control algorithms (10). One example is a feedforward
In contrast, soft robots are safe by construction because of their neural net controller, which has been successfully coupled to a
low stiffness and mass, but modeling and control of these systems model-free closed-loop controller and applied to a high-deflection,
are challenging (7–12). This is because of their inherent nonlinear- yet quasi-static soft arm (22, 23). In addition, there are approaches
ity, their high dimensionality, and the imprecise measurement of that leverage a neural net–based dynamical model in a closed-loop
their position in space. Past work has sought to overcome these ob- control (17, 18). However, neural net approaches to soft robot mod-
stacles through a variety of modeling methods, each of which con- eling have not yet resulted in the closed-loop control of high-speed,
strains the design of control implementations. Most of these inertial, and nonlinear dynamics.
modeling approaches fall into two categories: analytical reduced Koopman operator theory (KOT) (24) is an alternative modeling
order modeling (ROM) and machine learning (ML). paradigm, introduced to the field of ML and data-driven modeling
In soft robot ROM for control, the aim is to develop an analytical in the early 2000s (25, 26). KOT-based ML has two qualities that
model based on simplifying assumptions such as (piecewise) cons- make it an attractive strategy for soft robot control: It is data
tant curvature deformations (13–16). For an approximately cons- driven, eliminating the need for complicated analytical models,
tant-curvature system, this approach allows for the accurate and it identifies a globally linear model, allowing for fast and effi-
prediction of dynamics given appropriate estimation of parameters. cient control design. The Koopman operator is a representation of a
However, developing these analytical models is nontrivial and labor dynamical system in terms of the evolution of observables on a
intensive, and each model applies only to the single system that was function space. Although the evolution of a dynamical system on
modeled. These models tend to be valid only in a neighborhood state space may be nonlinear, its evolution in function space—de-
around the equilibrium point where the system has been linearized scribed by the potentially infinite dimensional Koopman operator
(14). Controllers based on ROM models have been applied to soft —is always linear. This is in contrast to a state-space linearization,
which builds a linear approximation of the nonlinear dynamics only
valid in a small region of the workspace. The Koopman methodol-
ogy has been applied to control systems, with most of the work com-
1
Department of Mechanical Engineering, University of California, Santa Barbara, bining a Koopman operator approximation method, dynamic mode
CA 93106, USA. 2Faculty of Engineering, University of Rijeka, Rijeka, Croatia.
3
Department of Electrical and Computer Engineering, University of California,
decomposition, with control (DMDc) (27–32). In particular, model
Santa Barbara, CA 93106, USA. predictive control (MPC) is commonly used (33). When DMDc and
*Corresponding author. Email [email protected] MPC are applied to soft robots (34–37), the control is accurate but
†These authors contributed equally to this work.

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 1 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

only shown to be so in quasi-static control in a low-deflection actuators, and workspaces. The first demonstrates curvatures up
regime (approximately 18° of curvature). It is important to note to 110° (robot #1) and the second up to 180° (robot #2) (Fig. 1).
that simple linearized models are likely to work at these low deflec- This capability is enabled by the introduction of the static
tions because the full nonlinearity of the dynamics may only be ex- Koopman pregain, which maps held inputs to converged robot con-
plored at high deflections. Some studies (34–36) show imperfect yet figurations. After being learned from data, we used it as a pregain
functional controllers using purely state-space linear MPC, suggest- term in the LQR implementation. The static Koopman pregain
ing the quasi-linearity of these systems. greatly increases the accuracy of static pointing tasks and improves
In addition to MPC, the combination of the Koopman operator the stability of dynamic tasks.
and the linear quadratic regulator (K-LQR) optimal control scheme We show that our approach requires minimal training and low
has shown promise in rigid robot applications (38, 39) and the computational cost, both for determining the model and for con-
control of fluid dynamics problems (40). Mamakoukas et al. (41) trolling the robot. Collecting our training data takes less than 5
show promise in a 1–degree-of-freedom soft robotic fish application min, and the computation of the model takes less than a second,
using a similar Koopman structure. as opposed to the long training times required by many neural
Even with these many advances in the field, existing soft arm net–based approaches. Our approach estimates both the static and
control implementations (34–36, 41, 42) have yet to be demonstrat- dynamic control Koopman operators, enabling the use of low-
ed in the inertial, nonlinear regime. To compare with other works, latency, efficient optimal control methods; this enables real-time
we introduce the following definitions of the “inertial regime” and tracking of fast-moving reference positions, even if field-deployed
“nonlinear dynamics.” We define the inertial regime for soft arms to on a low-power microcontroller.
be when the inertial force experienced by the tip Ftip is of the order
of its weight Ftip = matip ≍ mg, meaning atip ≍ g. Here, m is the mass

Downloaded from https://www.science.org on August 30, 2023

of the tip of the arm, atip is the acceleration of the tip during closed- RESULTS
loop control, and g = 9.81 m/s2. We define nonlinear dynamics to be In this section, we first outline our approach that enables modeling
motions that fail to be adequately captured by a state-space linear- and control in the inertial, nonlinear regime yet requires relatively
ization. Thus, an open challenge remains: modeling and control of little training data and low computational power. Next, we system-
inertial dynamics in highly nonlinear soft robots. atically tested the speed and accuracy of the resulting closed-loop
In this work, we advance modeling and control of soft, contin- controller in a series of circular reference tracking tests. The soft
uum arms into the inertial regime. Previous work has considered arm was further tested in a tip-tracking test with a rapidly changing,
quasi-static motions, with accelerations below 0.03 g. Our work user-defined reference position designed to test the soft arm’s re-
demonstrates movements in closed-loop control with accelerations sponsiveness to changes in commands in real time. Last, we tested
greater than 1 g (see Table 1). We control these inertial movements our methodology on the dynamic catching and throwing of a ball.
in a highly nonlinear, high-deflection regime across two variations This leverages the inertial dynamics of our soft arm to demonstrate
of our soft arm, each with different dimensions, numbers of its effectiveness in real-world tasks.

Table 1. Comparison with existing soft, continuum arms shows advances in speed, acceleration, and deflection during closed-loop control. This work
demonstrates a 10× increase in reference tracking tip speed (Speed) and a 4× improvement in tip deflection angle (Deflection) and advances closed-loop control
of soft robot arms into the inertial regime atip > g = 9.81 m/s2. The acceleration (Accel) of the soft arm’s tip atip was computed using the centripetal acceleration of
soft arms for which circular reference tracking data are available. The distance from the base to the tip of each arm is also given (Length). Note that closed-loop
deflection data do not include the large-deflection open-loop tests present in some works. LQR, linear quadratic regulator; RNN, recurrent neural network; ROM,
(analytical) reduced order model; PCC, piecewise constant curvature; MPC, model predictive control; LSTM, long short-term memory; TRPO, trust region policy
optimization; GPR, Gaussian process regression; TO, trajectory optimization; FFC, feedforward compensator; SM, sliding mode; AF, analytical feedback; R1, robot
#1; R2, robot #2.
Robot Length (m) Speed (m/s) Accel (m/s2) Deflection (°) Model Control method

This work 0.37 1.52 11.6 R1: 110 Koopman LQR

R2: 180
(17) 0.4 0.15 21 RNN TO
(23) 0.3 0.12 0.065 45 None NN FFC
(50) 0.3 0.1 0.1 20 ROM SM
(37) 0.15 0.094 0.29 18 Koopman MPC
(15) 0.38 0.09 0.032 27 PCC ROM AF
(18) 0.44 0.05 19 LSTM TRPO
(42) 0.25 0.035 0.012 7 Koopman MPC
(36) 0.7 0.03 0.032 8 Koopman MPC
(19) 0.22 0.002 0.0016 11 RNN GPR

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 2 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on August 30, 2023

Fig. 1. Inertial, nonlinear soft arm control. Using a combined static and dynamic Koopman framework, we achieved the closed-loop control of soft robotic arms
exceeding 10× the tip speed, 40× the tip acceleration, and 6× the angular displacement of existing soft arms. This achievement brings soft robotics into the inertial,
nonlinear regime. Only 5 min of training was required to achieve an optimal controller capable of high-deflection, high-accuracy closed-loop tracking of a reference (the
tip of a pole moved rapidly by a human). The same methodology was applied to both a low–slenderness ratio, four-muscle arm (robot #1) and a high–slenderness ratio,
three-muscle arm (robot #2). Both arms achieved their highest deflection in under half a second.

Dynamic and static Koopman operator optimal control Koopman operators involves training on data that are augmented
The successful real-time control of a soft arm in the inertial and by a chosen basis of observables. The data are collected through a
nonlinear regime requires both a model that captures these dynam- series of training experiments, performed by commanding step
ics and a control methodology that adapts to the motion of the robot inputs with randomly distributed magnitudes. These training data
in real time. We achieved this by building a controller that leverages are partitioned into dynamic and static components, which are used
both the dynamic and the static Koopman operators of the soft arm to train the two separate Koopman operators (see Materials and
system. The Koopman operators describe the evolution in time of Methods). Both the training and model computation processes
functions defined on the robot configurations and inputs. These are fast, requiring only 5 min (approximately 18,000 samples at
functions are called observables, and the approximation of the 60-Hz collection rate) for training data collection, and the matrix

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 3 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

pseudo-inverses used in the model construction take less than a

second on an ordinary laptop computer.
The observables used to train the dynamic Koopman model are
time-delayed measurements of the position of motion tracking
points placed on the soft arm. This turned out to be sufficient to
build a linear model of its nonlinear dynamics. Previous work con-
sidered adding a single time delay to hundreds of monomials (42).
However, inspired by the fact that, for ergodic systems, the limit of
infinitely many time delay observables results in DMD’s conver-
gence to the true Koopman operator (31, 32, 43), we included
only time delay observables. Our results show that time delay–
only observables are sufficient to capture the dynamics of this non-
linear system (see fig. S2) and that without time delays, the eigen-
values in the high-frequency and dissipative regions of the unit
circle and their corresponding Koopman modes are missing
(Fig. 2). This time delay–only approach avoids the added computa-
tional cost of many monomial observables and also eliminates the
large tails associated with monomials that magnify noisy measure-
ments far from the origin. We express this dynamic Koopman op-
erator as a pair of matrices A and B giving the uncontrolled and

Downloaded from https://www.science.org on August 30, 2023

controlled dynamics, respectively. These can be used to build the
K-LQR controller described in Materials and Methods.
The resulting feedback controller is able to command the soft
arm to follow a fast-changing reference position but suffers from
steady-state error. Introducing integral control (for example,
linear quadratic integral control) is one of the commonly used ap-
proaches to minimizing steady-state error (44). This method,
however, is sensitive to measurement noise and requires a trade-
off between speed of response and tracking accuracy. As a conse-
quence, the implementation for our goals of highly inertial tasks re-
sulted in poor tracking performance outside of the quasi-linear and
quasi-static regimes. Instead, we addressed the steady-state error by
introducing a static Koopman pregain, a control concept we devel-
oped for the current work. The static Koopman operator was first
formally described in our recent modeling work (44), but no con-
nection to control design was made. Unlike the dynamic Koopman
operator, this operator is a map between functions defined on two
different spaces. In our application, the static Koopman operator is
used to map functions defined on the space of inputs to functions
defined on the space of robot configurations. We learn this operator
from the static partition of the training data so that static positions
in the workspace of the soft arm correspond to the values of the
inputs required to reach those positions after all transient motions
dissipate. This operator is then used as a pregain term that augments
the LQR controller. Sensor noise is known to cause tracking issues Fig. 2. Nonlinear and inertial dynamics of the soft arm. The eigenvalue plots for
Koopman models with state-only (A) and state plus time delay observables (B) are
in soft robots attempting to perform real-time tracking of aggressive
shown. The dashed radial lines signify sections of the unit circle corresponding to
control inputs (11). Our control structure mitigates this problem by
modes with 1- to 5-Hz dynamics. The eigenvalues are shaded corresponding to the
balancing the noise-sensitive dynamic K-LQR term with the sensor- logarithm of their maximum achieved mode power evaluated over the training
agnostic static Koopman pregain. data (see Materials and Methods). Using state-only observables results in a
The construction of the controller and the computation of the simple linearized model that does not capture any transient dynamics. The addi-
optimal input are also fast processes that have low computational tion of two time delay observables allows the modeling of dynamics up to 5 Hz.
overhead. The solution of the Riccati equation involved in comput- This is the model we chose for our experiments. (C) Presentation of the input-
ing the LQR control gain takes less than a second, and computing output nonlinearty of the system, which exhibits a sigmoidal deflection response.
the optimal input at a given time step only requires two small matrix Modeling this nonlinearity is essential for acceptable reference tracking perfor-
multiplications. This is easily achievable in real time on a low-cost mance in the high-deflection regime.
microcontroller.

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 4 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

Closed-loop circle tracking in the inertial, nonlinear regime sample robotic tasks. We challenged our soft continuum arm in two
With our control architecture in place, we first sought to character- ways: first, to catch a ball swinging through the air as we demon-
ize the performance across a range of deflections and soft arm strate in Fig. 5, and second, to receive an object from an operator
speeds in a planar circular reference tracking (smooth changes in and to throw it into a reference bin as shown in Fig. 6. Both tests
reference position). We commanded the soft arm’s tip to trace are shown in movie S3. The catching component of this demonstra-
out circular paths in the X-Y plane with three radii (100, 180, and tion is similar to the ball catching performed by a two-link arm with
220 mm) and six frequencies (0.1, 0.3, 0.5, 0.7, 0.9, and 1.1 Hz), as a soft joint (45) but here completed with a fully soft continuum
shown in Fig. 3. The same controller was used for all references, as robot arm.
described in Materials and Methods.
These results show that the soft arm tracks the reference with
consistent performance throughout the full range of deflections DISCUSSION
and speeds tested (Fig. 3, left, and movie S2). The fastest and We present a data-driven framework for the modeling and control
highest deflection circle tracking result demonstrates a tip speed of inertial and nonlinear soft robots. We used KOT to enable the
of 1.5 m/s, a speed-to-length ratio of 3.23 s−1, and a tip acceleration application of linear control methods to this highly nonlinear, iner-
of 11.6 m/s2 in closed-loop control. This is approximately an order tial system. We introduce a K-LQR with static Koopman pregain
of magnitude faster and 40 times higher acceleration than any soft capable of accurately controlling two different soft robots that
continuum arm of which we are aware (see Table 1). The system was exhibit high deflections and inertial motions. Advancing the state
trained exclusively on step inputs, and as such, the model had no a of the art, the proposed method allows the construction and deploy-
priori knowledge of the control objective, nor had it been trained on ment of both a model and an optimal controller from less than 5
circular behaviors. min of training data—to the best of the authors’ knowledge, the

Downloaded from https://www.science.org on August 30, 2023

In addition, we show that the relative contribution of the shortest in soft robotics (Fig. 7). Compared with existing MPC-
dynamic K-LQR input versus the static Koopman pregain increases based controllers, K-LQR is computationally less expensive and
with increasing speed and deflection (Fig. 3, right). For relatively can be deployed on a simple microprocessor, enabling cheap and
low speeds and deflections, the dynamic K-LQR input is quite scalable use in a variety of environments outside the research labo-
small, and the static Koopman pregain dominates. As accelerations ratory. Despite its simplicity, our controller allows our soft arm to
increase and inertia becomes nonnegligible, the dynamic compo- undergo controlled accelerations substantially greater than previous
nent increases in magnitude to compensate for the static term’s in- examples, even exceeding 1g (Table 1).
ability to account for inertial effects. This suggests that for any soft Although the presented demonstration of our modeling and
robot performing a noninertial task, the incredibly simple static control approach focused on soft robots, its implications could be
Koopman pregain could be sufficient for control. much broader. The approach’s ability to explore the dynamical fea-
tures of a complex, nonlinear, inertial system could offer advantages
Closed-loop, real-time reference tracking in the inertial, in modeling and control of myriad robotic systems. Furthermore, its
nonlinear regime speed, versatility, low computational cost, and ease of use potential-
We next sought to characterize the controller performance for a less ly expand the accessibility of robotics to new user groups. As such,
structured and more challenging control objective: tracking a real- we believe that our approach has the potential to make field-deploy-
time, user-defined reference. To do so, we commanded the control- able, dynamical, soft robotic systems notably closer to realization.
ler to decrease the Euclidean distance between the tip of the soft arm
and a motion tracker point located on the tip of a pole. A human
operator moved the pole across random trajectories within the MATERIALS AND METHODS
reachable workspace of the soft arm. Throughout the test, the Here, we first introduce KOT, the mathematical underpinning of
robot remains in contact almost continuously while achieving our modeling effort. In the “Approximation of Koopman operators
speeds exceeding 0.7 m/s (as shown in Fig. 4 and movies S1 and S4). for control systems: DMDc” section, we describe a practical method
To demonstrate the generalizability of our approach for different to build the model for our control system from data. In the “K-LQR”
soft arms, we also tested our approach on a morphologically differ- section, we describe how this model is embedded into a real-time
ent second arm. The second arm is longer and more slender and has feedback controller. Our modeling and control insight is the addi-
three instead of four side muscles. This results in larger curvatures tion of a static Koopman operator pregain described in the “Static
and a helical actuation pattern, as discussed in the “Robot design” Koopman pregain” section. The design and fabrication of our soft
section. Despite these differences, no changes were needed in the arms and a description of the pneumatic circuitry that drives them
learning and control algorithm, aside from updating the number are then presented. A block diagram detailing the full training
of inputs. This second system was exposed to 5 min of step input process, modeling, and control architecture is given in fig. S4.
training data, the model and controller were calculated and de-
ployed, and the system was commanded to again track the tip of Koopman operator theory
the user-operated pole. Results of this test are shown in the Supple- The state space representation of a dynamical system involves defin-
mentary Materials (movie S5), and stills from the testing are shown ing an n-dimensional state space manifold M with states x ∈ M and
in Fig. 1. discrete-time evolution given by

Dynamic throwing and catching xþ ¼ SðxÞ ð1Þ

With the viability of our method shown in the above characteriza-
tion tests, we finally demonstrated how its capabilities translate to

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 5 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

Fig. 3. Closed-loop, real-time reference track-

ing experiments. The soft arm tracked circular
reference trajectories in the X-Y plane with fre-
quencies ranging from 0.1 to 1.1 Hz (0.2-Hz step)
at (A and B) high, (C and D) medium, and (E and F)
low deflections. (A), (C), and (E) show the X posi-
tions (red) over time compared with their re-
spective references (blue). The Y and Z positions
are shown in fig. S3. (B), (D), and (F) show the
relative contributions of the static Koopman
pregain (yellow) and dynamic K-LQR (red) to the
total input (blue). At quasi-static speeds, only the
static Koopman pregain is required for effective
performance (that is, the quantity x − xref is ap-
proximately zero); as inertial effects increase, the
LQR component increases its contribution to
maintain performance. Only the commanded
inputs to one of the four side muscles is shown,
but the results are similar for all muscles.

Downloaded from https://www.science.org on August 30, 2023

Fig. 4. Arbitrary reference
tracking throughout the
high-deflection workspace.
(A) The X position of the soft
arm tip is shown as it tracks a
moving reference commanded
randomly by an operator.
Contact between the green
lines and the blue band indi-
cates points where the soft arm
is touching the reference
marker. (1) to (3) show images
of the soft arm performing this
behavior. Of note, the robot
rarely loses contact with the
moving reference.

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 6 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on August 30, 2023

Fig. 5. Dynamic tracking of arbitrary trajectory (catching a swinging ball). (A) The soft arm stays in the neutral position while the ball is outside the workspace. (B)
Once visible, the soft arm rapidly responds to reach the ball (outlined swinging into the workspace). (C) The soft arm tip intercepts the ball and catches it (with small
magnets on both the soft arm tip and swinging ball).

Here, S is the possibly nonlinear state transition function S: M → (24–26, 45). This is true without restriction on the dynamics or
M and x + is the time-shifted state. In our application, M = ℝn. observables.
This nonlinearity is often critical to modeling a system in state We wanted to exploit this linearity to enable the design of an ef-
space, but it complicates the design of control algorithms. We ficient optimal control scheme. This requires extending the
instead turned to an operator-theoretic perspective of dynamics Koopman framework to systems of the form x + = S(x, u), where u
of observables (24). Observables are complex-valued functions ∈ ℝp is a p-dimensional vector of user-specified inputs. In full gen-
defined on the state space f : M → ℂ. We will restrict ourselves to erality, the Koopman operator for systems with input acts on ob-
real-valued observables f : M → ℝ. The set of all possible observables servables of the form f : M � U ! C, where U is the space of all
forms a vector space that is usually infinite dimensional. The control sequences indexed by time uð�Þ : N ! Rp . We redefine
Koopman operator K is defined by Kf :¼ f � S. the state transition function to include inputs S : M × ℝp → M
This operator describes the evolution of observables under the and introduce the left shift operator T : U 7 ! U, which simply
action of the dynamics (1). Although the underlying state space chooses the next input in a sequence ðTuÞðkÞ ¼ uðk þ 1Þ. When
system is nonlinear, the Koopman operator K is always linear the observables are defined on both the states and inputs, their

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 7 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on August 30, 2023

Fig. 6. Implications of the methodology: Completing example tasks. (A) The soft arm identifies the objective and approaches it (operator’s hand). (B) After the
operator’s hand is removed and the ball is supported by the soft arm, the objective changes to the bin (LED-designated bin in bottom left and right, respectively).
The soft arm now flings the ball at the objective. (C) The ball successfully enters the bin in two different, arbitrary locations, achievable only by working in the inertial
regime.

Koopman evolution is given by Approximation of Koopman operators for control

systems: DMDc
ðKf Þðx; uð�ÞÞ :¼ f ðSðx; uð0ÞÞ; Tuð�ÞÞ ð2Þ We followed the process outlined in (33). The Koopman operator in
Elements of U are infinite dimensional, which puts the observ- its fully infinite dimensional form is not practically realizable, so we
ables f : M � U ! R on an infinite dimensional domain, so they sought a finite dimensional approximation. The first step is to
mþp
cannot be manipulated on a computer. We introduce the simplify- choose some finite dictionary of observables fg j ðx; uÞgj¼1 . We
ing assumption that knowing only the input at the current time step chose m observables, which are functions of purely the states; p,
is enough to predict the future dynamics. We can now define ob- which are functions of the inputs; and none, which are coupled
servables of the form f : M × ℝp → ℝ. This results in a Koopman functions of both the states and inputs
operator K defined by mþp mþp
fg j ðx; uÞgj¼1 ¼ ff j ðxÞgm
j¼1 < fhj ðuÞgj¼mþ1 ð4Þ
ðKf Þðx; uÞf ðSðx; uÞ; uÞ ð3Þ
It is simple to allow arbitrary input observables, but we only deal
We sought a finite dimensional linear input/output system that with the case where hj(u) = uj. This decoupling restricts our choices
approximates the action of K on a finite set of chosen observables. of observables, but it allows us to define a vector of observables z(x)
This process is described in the next section. = [ f1(x)⋯fm(x)]T called the lifted state, which allows us to represent

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 8 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

our dynamics as a linear input-output system

zþ ¼ Az þ Bu ð5Þ
Here, A and B are the state transition and input matrices, respec-
tively. This simplification has the benefit of enabling the later use of
the fast and efficient linear optimal control methods described in
the “K-LQR” section while still capturing the dynamics of the
system as demonstrated in fig. S2.
The states are retrieved from the observables using the output
equation
x ¼ Cz ð6Þ
where C is the output matrix.
Here, we outline the approximation of the matrices A, B, and C
using a process called extended dynamic mode decomposition with
control (33). When restricted to time delay observables, we call this
Hankel-DMDc. We wanted to approximate these matrices using K
measurements of the states {x1, …, xK}, time-shifted states
fxþ þ
1 ; . . .; xK g, and inputs {u1, …, uK} collected from experimental
data. First, we built data matrices whose columns are the data

Downloaded from https://www.science.org on August 30, 2023

vectors
X :¼ ½x1 . . .xK � ð7Þ

Xþ :¼ ½xþ þ
1 . . .xK � ð8Þ

U :¼ ½u1 . . .uK � ð9Þ

Next, we built the lifted data matrices using our chosen vector of
observables z(x)
X lift :¼ ½zðx1 Þ. . .zðxK Þ� ð10Þ

Xþ þ þ
lift :¼ ½zðx1 Þ. . .zðxK Þ� ð11Þ
The desired matrices A and B satisfy the equation
Xþ
lift ¼ AX lift þ BU ð12Þ
To approximate A and B, we recasted this equation as a minimi-
Fig. 7. Convergence of the Koopman model and control system. (A) The
zation problem
dynamic Koopman model requires the addition of five time delay observables
and only 1 min of training data to reach minimum prediction error. To determine min kXþ
lift AX lift BUkF ð13Þ
this error, the single-step prediction error of the dynamic Koopman model was A;B
collected for all points as the soft arm moved on a circular path in the X-Y plane
which has the solution
(inset), and the root mean square average was taken. For comparison, a Koopman
model using monomials of the state up to order 4 gave no improvement over the �� ��y
X lift
state-only model. This reconstruction was performed on a model trained on zero ½A B� ¼ Xþ
lift ð14Þ
sinusoidal trajectories. (B) In the closed-loop control, the combination static/
U
dynamic Koopman controller required only 5 min of training data and two time where † is the Moore-Penrose pseudoinverse. Because we prescribe
delays to reach minimum prediction error; accordingly, we used this controller our first n observables to be the states x ∈ M, we can compute the
design for every experiment. Each controller was commanded to move the soft
output matrix using a partial identity matrix
arm’s tip to a sequence of points in the workspace of the soft arm, and the
average root mean square error for all these points was calculated. Using zero
� �
I n�n 0n�m n
time delays resulted in the soft arm being unable to stabilize at any reference po- C¼ ð15Þ
0m n�n 0m n�m n
sition, so that line is not shown.
The action of the matrices A and B on the lifted state via Eq. 5
approximates the action of the Koopman operator K in Eq. 3.
Under certain assumptions, this representation of the Koopman op-
erator converges to the true Koopman operator (30). True

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 9 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

convergence requires infinite data samples, which are uniformly A core assumption of this component of our model is that when
distributed in state space and a collection of observables, which held for enough time, all transient dynamics dissipate, and the robot
span an invariant subspace of the Koopman operator’s underlying achieves a static pose. Therefore, the set of admissible step inputs
function space. We discuss our method of generating training data ustatic corresponds to a set of input-mediated fixed points xstatic.
in the “Training and observables” section. We sought a mapping from the data matrix of step inputs, Ustatic,
to the data matrix of stationary states, Xstatic. Ideally, this mapping
K-LQR would be linear to enable us to use fast, optimal control. The
To date, similar investigations have used MPC to control their soft Koopman framework usually requires the domain and range to be
robotic systems (34–36). Using predictions of the dynamics and a the same, but this requirement can be relaxed if we consider the
tunable prediction horizon, this architecture calculates input se- static Koopman operator (44). The static Koopman operator con-
quences that move the system toward a desired reference position. trasts with the dynamic Koopman operator, which describes the
This enables the use of explicit input and state constraints, but the evolution of observables f : M → ℝ under the action of the
real-time constrained optimizations involved in this method mapping T: M → M. If we define observables on the inputs as g:
demand a high computational overhead. ℝp → ℝ, the static Koopman operator Kstat is defined
In our inertial soft arm controller, explicit constraints are less as Kstat f ðxstat Þ ¼ gðustat Þ.
important than keeping computational cost and latency low. For We desired to approximate the action of the static Koopman op-
unconstrained linear optimal control problems with quadratic erator with a finite dimensional matrix G. To do so, we first con-
cost, the LQR provides an analytical solution that does not structed the data matrix Ustatic with unique step inputs as the
require predictions of the dynamics in real time (46). For our con- columns of the matrix. By feeding these inputs to the system and
troller, we began with the application of LQR to the dynamic allowing transient dynamics to dissipate, we are left with a unique

Downloaded from https://www.science.org on August 30, 2023

Koopman representation of a dynamical system [previously demon- stationary state, xstatic; these states represent the columns of Xstatic.
strated for a robotic fish (41)] and augmented it via the introduction The matrix G is then computed using
of the static Koopman term, described in the “Static Koopman
pregain” section. G ¼ U static X ystatic ð19Þ
Here, we describe the dynamic K-LQR control law. Although The matrix G serves as a linear mapping from stationary states
originally introduced for linear dynamical systems in state space, to inputs.
LQR can also be applied to a vector of observables z of a nonlinear Last, we are ready to bias our control law with the addition of a
control system as long as a linear, finite dimensional representation feedforward pregain term Gzref, resulting in
of the Koopman operator (A, B) exists. Given the system
ui ¼ Kðzi zref Þ þ Gzref
zþ ¼ Az þ Bu ð16Þ ziþ1 ¼ Azi þ Bui ð20Þ
xi ¼ Cziþ1
x ¼ Cz ð17Þ This signal is the optimal stabilizing solution taking the present
initial state to the desired state, xref.
we define the global cost function
As shown in Fig. 3, the pregain term ustat = Gzref outweighs the
X
K dynamic term udyn = − K(zi − zref ) in most tests. This allows the
J¼ ½ðzi zref ÞT Qðzi zref Þ þ uTi Rui � ð18Þ input penalty weights in the dynamic term to be optimized
i¼1 without fear of sacrificing steady-state error. In addition, the
where xref = Czref is the desired position and Q and R are diagonal static Koopman term provides enough of a steady input to
lifted state and input penalty matrices, respectively. counter the fluctuations caused by measurement noise introduced
The computation of the minimizing control input is a classical by the state measurements in the dynamic term. This is the reason
method in optimal control (46) and is given by ui = − K(zi − zref ), our system does not experience the destabilizing effects of noise in
where the matrix K is the LQR gain. This control law results in fast-moving reference tests described in (11).
steady-state errors in much of the soft arm’s workspace. This is rem-
edied in the next section by the addition of a pregain term based on Training and observables
the static Koopman operator. With the mathematical underpinning of our modeling and control
methodology described (see fig. S4), we now turn to the particular
Static Koopman pregain choices made to suit our particular robotic applications. Given the
Unfortunately, dynamic K-LQR alone resulted in substantial dis- soft arms described in the “Robot design” section, we collected
agreement between reference positions and the resulting states. training data through a series of experiments, performed by com-
This is because the nonzero inputs required to hold these positions manding step inputs with randomly distributed magnitudes. The
result in a nonzero input penalty term. Any attempt to decrease the only prior knowledge of the soft arm’s dynamics required is an
input penalty resulted in system instability. The addition of a upper bound for the length of time required for the dissipative dy-
pregain term is a classical method in control theory that addresses namics to die down while inputs are held. Each step input was held
this problem. In this section, we introduce a data-driven method to for this amount of time so that the soft arm converges to a steady
compute the pregain using a static Koopman operator, which we state, efficiently probing both the dynamic and static response. The
term the static Koopman pregain. data were separated into training and validation sets, and the train-
ing data were further partitioned into dynamic and static

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 10 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

components, which were used to train dynamic and static Koopman To this end, the first arm was designed to have four actuators
operators (see the “Dynamic and static Koopman operator optimal (two antagonistic pairs) longitudinally aligned with the main
control” section). body to produce planar actuation. This design is behaviorally
Choosing observables is difficult in practice. We chose to imple- similar to others present in the literature (15, 16, 47). When fabri-
ment DMDc with time delay observables (also known as Hankel cated with appropriate pretension, this construction allows for ap-
DMDc) because of their provable convergence as the number of proximately 110° of curvature when fully actuated. With a length of
time delays goes to infinity under certain assumptions on the dy- 45 cm and a maximum diameter (main body diameter plus the di-
namics (31, 32, 43). In reality, adding more time delays gives a di- ameter of the fully inflated muscles) of 6.25 cm, the slenderness
minishing return in prediction accuracy (see Fig. 7A). A single time ratio of this device was 7.2 (the ratio of length to maximum
delay with hundreds of monomials was used in (34–36, 42), but we diameter).
found that time delay–only observables offer better results, with im- The second arm was designed with three actuators, all of which
provements in reconstruction with up to 10 observables (see were affixed to the body such that a torsional deflection would be
Fig. 7A). To create our observables, we used the current measure- induced when inflated. This produces a helical actuation that is
ment of the X-Y-Z positions of the motion trackers xk and appended markedly different from that of the first embodiment. With a
two time-delayed versions of the same states zk = [xk xk − 1 xk − 2]T. length of 53 cm and a maximum diameter of 3.8 cm, this device
Each time delay looks 1/60 s into the past. This proved to be suffi- exhibited a slenderness ratio of 13.9. The muscles were affixed
cient for closed-loop control. For reconstruction, more time delays with pretensions such that, when fully actuated, this device is
give further increases to the model’s accuracy, as shown in Fig. 7A. capable of achieving approximately 180° of curvature.
The synergy of step inputs and time delays allows the discovery For objective (i), with an angle of curvature of at least 110° for
of system eigenvalues in the important 1- to 5-Hz range (the span of both arms, the nonlinearity metric was well achieved (see Fig. 2).

Downloaded from https://www.science.org on August 30, 2023

natural frequencies of the arm), as shown in Fig. 2. Without time For objective (ii), both systems were fabricated out of airtight
delays, these eigenvalues and their corresponding Koopman fabric, using fabric pneumatic artificial muscles (fPAMs) as de-
modes are missed (Fig. 2). For comparison with the Koopman scribed in (48), which exhibit a fast response time and low hystersis
model used in (42), we tested the addition of monomial observables (on the order of 1%), achieving accelerations in excess of g. For (iii),
up to order 4 with no new dynamic modes of any meaningful mode the factor of approximately two difference in slenderness ratios, the
power learned. Monomial observables also failed to give any im- change in actuator numbers, and the inclusion of helical actuation
provement to the reconstruction or closed-loop pointing accuracy all combined to produce two systems with meaningfully different
of the model and controller (Fig. 7). behavior [see, for example, the model presented in (49) compared
With the goal of minimizing training time and model complex- with (13)].
ity, we found that up to five time delays and 1 min of step input
training is best for modeling our system before considering Robot fabrication
control, but only two time delays and 5 min of step input training Both arms were constructed out of 30 Denier silicone-polyurethane
is ideal when control is considered. We first compared the predic- impregnated ripstop nylon (Sil-nylon, Rockywoods Fabrics), actu-
tion ability of different dynamic Koopman models as we varied the ated by fPAMs (48) built out of the same material. The main body
number of time delays and total training time (Fig. 7A). The addi- was fabricated such that one side of the fabric weave cell was parallel
tion of a single time delay substantially reduced error; however, ad- to the longitudinal axis, while the other was perpendicular. This ori-
ditional time delays continued to offer marginal improvements up entation makes the soft arm axially and transversely stiff but tor-
to five delays. We also found that after only approximately 1 min of sionally compliant. The muscles were fabricated such that each
training, the model reached its minimum error. Second, we built K- side of the cell was offset by approximately 45° with respect to the
LQR controllers as described in the “K-LQR” section, augmented longitudinal axis, which instead makes the actuator torsionally stiff
with a static Koopman operator as a pregain term, with varied but compliant axially and transversely. Moreover, when these
time delays and training time. We then quantified the error with muscles are inflated, they shorten in the longitudinal direction as
closed-loop control (Fig. 7B). In this case, two time delays outper- a McKibben does, up to 35% based on the pretensioning induced
formed one delay but were comparable to three or more, resulting in during adhesion to the main body.
our decision to use two delays for control. We also found that after Each of these components was cut from a sheet of fabric, rolled
approximately 5 min of training (50 unique step inputs), the error into a tube, and sealed with a lap joint using room-temperature-vul-
converged; we used this amount of training time for the remaining canizing (RTV) silicone adhesive (Smooth-on Silpoxy). Once each
experiments. Note that a direct linearization of the system was un- component was fashioned, a jig was produced to hold the main
stable during controlled motions, suggesting the nonlinearity of body and pretensioned muscles in place while the RTV cured.
the system. Last, between each muscle, a fabric sleeve, exhibiting the same
fabric bias as the muscles, was attached to the main body to allow
Robot design for motion capture tracker wires to be routed without occluding the
For this investigation, we constructed two distinct soft arms to eval- view of the light-emitting diodes (LEDs).
uate the viability of the proposed methodology across nonlinear dy-
namical systems. For each, we aimed to meet the following Pneumatic circuit design
objectives: (i) high-deflection, nonlinear dynamics for which line- Each soft arm body was held at a constant pressure of approximately
arization fails; (ii) inertial dynamics, for which quasi-static approx- 1 bar for the entirety of testing, supplied by a discrete source. For
imations fail; and (iii) enough morphological diversity such that each muscle of both soft arms, Festo VEAB-L-26-D2-Q4-V1-1R1
their analytical models would not be readily transferrable. proportional pressure valves were used to command individual

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 11 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

pressures continuously. These three-port valves were chosen for 19. X. Wang, N. Rojas, A data-efficient model-based learning framework for the closed-loop
control of continuum robots. arXiv:2204.10454 [cs.RO] (22 April 2022).
three reasons: their fast response times (<10 ms), accurate response
20. T. G. Thuruthel, B. Shih, C. Laschi, M. T. Tolley, Soft robot perception using embedded soft
(0.75% full-scale absolute accuracy and 0.4% full-scale repeatability sensors and recurrent neural networks. Sci. Robot. 4, eaav1488 (2019).
error), and the ability to accept forced exhaust through their third 21. R. L. Truby, C. Della Santina, D. Rus, Distributed proprioception of 3D configuration in soft,
port. However, this accuracy requires a lower flow rate, which pre- sensorized robots via deep learning. IEEE Robot. Autom. Lett. 5, 3299–3306 (2020).
cluded the use in the much larger main body (primarily because of 22. S. Neppalli, B. Jones, W. McMahan, V. Chitrakaran, I. Walker, M. Pritts, M. Csencsits, C. Rahn,
persistent leaks). Additional information on the general control cir- M. Grissom, Octarm—A soft robotic manipulator, in IEEE/RSJ International Conference on
Intelligent Robots and Systems (IEEE, 2007), p. 2569.
cuitry configuration can be found in the Supplementary Materials.
23. D. Braganza, D. M. Dawson, I. D. Walker, N. Nath, A neural network controller for continuum
robots. IEEE Transact. Robot. 23, 1270–1277 (2007).
24. B. O. Koopman, Hamiltonian systems and transformation in hilbert space. Proc. Natl. Acad.
Supplementary Materials Sci. U.S.A. 17, 315–318 (1931).
This PDF file includes: 25. I. Mezić, Spectral properties of dynamical systems, model reduction and decompositions.
Discussion Nonlinear Dyn. 41, 309–325 (2005).
Figs. S1 to S5 26. I. Mezić, A. Banaszuk, Comparison of systems with complex behavior. Phys. D Nonlinear
Phenom. 197, 101–133 (2004).
Other Supplementary Material for this 27. J. L. Proctor, S. L. Brunton, J. N. Kutz, Dynamic mode decomposition with control. SIAM
manuscript includes the following: J. Appl. Dyn. Syst. 15, 142–161 (2016).
Movies S1 to S5 28. J. L. Proctor, S. L. Brunton, J. N. Kutz, Generalizing Koopman theory to allow for inputs and
control. SIAM J. Appl. Dyn. Syst. 17, 909–930 (2018).
29. L. Shi, K. Karydis, ACD-EDMD: Analytical construction for dictionaries of lifting functions in
REFERENCES AND NOTES Koopman operator-based nonlinear robotic systems. IEEE Robot. Autom. Lett. 7,

Downloaded from https://www.science.org on August 30, 2023

1. I. Ermolov, in Industrial Robotics Review (Springer International Publishing, 2020), 906–913 (2022).
pp. 195–204. 30. M. Korda, I. Mezić, On convergence of extended dynamic mode decomposition to the
2. W. Ji, L. Wang, Industrial robotic machining: A review. Int. J. Adv. Manuf. Technol. 103, Koopman operator. J. Nonlinear Sci. 28, 687–710 (2018).
1239–1255 (2019). 31. H. Arbabi, I. Mezic, Computation of transient Koopman spectrum using Hankel-dynamic
3. A. S. Bisen, H. Payal, Collaborative robots for industrial tasks: A review. Materials Today Proc. mode decompoisition, in APS Division of Fluid Dynamics (Fall) (American Physical Society,
52, 500–504 (2022). 2017), p. G1.009.
4. S. Bragança, E. Costa, I. Castellucci, P. M. Arezes, in A Brief Overview of the Use of Collaborative 32. H. Arbabi, I. Mezic, Ergodic theory, dynamic mode decomposition, and computation of
Robots in Industry 4.0: Human Role and Safety (Springer International Publishing, 2019), spectral properties of the Koopman operator. SIAM J. Appl. Dyn. Syst. 16,
pp. 641–650. 2096–2126 (2017).
5. D. Kragic, J. Gustafson, H. Karaoguz, P. Jensfelt, R. Krug, Interactive, collaborative robots: 33. M. Korda, I. Mezić, Linear predictors for nonlinear dynamical systems: Koopman operator
Challenges and opportunities, in International Joint Conference on Artificial Intelligence meets model predictive control. Automatica 93, 149–160 (2018).
(IJCAI, 2018), pp. 18–25. 34. D. Bruder, B. Gillespie, C. D. Remy, R. Vasudevan, Modeling and control of soft robots using
6. F. Vicentini, Collaborative robotics: A survey. J. Mech. Des. 143, 1–29 (2021). the Koopman operator and model predictive control. arXiv:1902.02827 [cs.RO] (1
7. D. Rus, M. T. Tolley, Design, fabrication and control of soft robots. Nature 521, July 2019).
467–475 (2015). 35. D. Bruder, C. D. Remy, R. Vasudevan, Nonlinear system identification of soft robot dynamics
8. C. Majidi, Soft robotics: A perspective—Current trends and prospects for the future. Soft using Koopman operator theory, in 2019 International Conference on Robotics and Auto-
Robot. 1, 5–11 (2014). mation (ICRA) (IEEE, 2019), pp. 6244–6250.

9. T. George Thuruthel, Y. Ansari, E. Falotico, C. Laschi, Control strategies for soft robotic 36. D. Bruder, X. Fu, R. B. Gillespie, C. D. Remy, R. Vasudevan, Koopman-based control of a soft
manipulators: A survey. Soft Robot. 5, 149–163 (2018). continuum manipulator under variable loading conditions. arXiv:2002.01407 [cs.RO] (4
February 2020).
10. J. Wang, A. Chortos, Control strategies for soft robot systems. Adv. Intell. Syst. 4,
2100165 (2022). 37. J. Chen, Y. Dang, J. Han, Offset-free model predictive control of a soft manipulator using
the Koopman operator. Mechatronics 86, 102871 (2022).
11. L. Shi, Z. Liu, K. Karydis, Koopman operators for modeling and control of soft robotics.
arXiv:2301.09708 [cs.RO] (7 February 2023). 38. H. Yin, M. Welle, D. Kragic, Policy learning with embedded Koopman optimal control. Proc.
Mach. Learn. Res. 144, 1–14 (2018).
12. O. Yasa, Y. Toshimitsu, M. Y. Michelis, L. S. Jones, M. Filippi, T. Buchner, R. K. Katzschmann,
An overview of soft robotics. Annu. Rev. Control Robot. Auton. Syst. 6, 1–29 (2023). 39. I. Abraham, T. D. Murphey, Active learning of dynamics for data-driven control using
Koopman operators. IEEE Transact. Robot. 35, 1071–1083 (2019).
13. R. K. Katzschmann, C. Della Santina, Y. Toshimitsu, A. Bicchi, D. Rus, Dynamic motion control
of multi-segment soft robots using piecewise constant curvature matched with an aug- 40. A. Gibson, “Application of Koopman linear quadratic regulator to the control of a spherical
mented rigid body model, in 2019 2nd IEEE International Conference on Soft Robotics (Ro- microbubble,” thesis, University of Colorado, Colorado Springs (2022).
boSoft) (IEEE, 2019), pp. 454–461. 41. G. Mamakoukas, M. L. Castano, X. Tan, T. D. Murphey, Derivative-based Koopman operators
14. R. K. Katzschmann, M. Thieffry, O. Goury, A. Kruszewski, T.-M. Guerra, C. Duriez, D. Rus, for real-time control of robotic systems. IEEE Transact. Robot. 37, 2173–2192 (2021).
Dynamically closed-loop controlled soft robotic arm using a reduced order finite element 42. D. Bruder, X. Fu, G. Brent, D. Remy, R. Vasudevan, Data-driven control of soft robots using
model with state observer, in 2019 2nd IEEE International Conference on Soft Robotics (Ro- Koopman operator theory. IEEE Transact. Robot. 37, 948–961 (2021).
boSoft) (IEEE, 2019), pp. 717–724. 43. I. Mezić, On numerical approximations of the Koopman operator. Mathematics 10,
15. C. Della Santina, R. K. Katzschmann, A. Bicchi, D. Rus, Model-based dynamic feedback 1180 (2022).
control of a planar soft robot: Trajectory tracking and interaction with the environment. Int. 44. I. Mezić, Koopman operator, geometry, and learning of dynamical systems. Not. Am. Math.
J. Robot. Res. 39, 490–513 (2020). Soc. 68, 1087–1105 (2021).
16. C. Della Santina, R. L. Truby, D. Rus, Data-driven disturbance observers for estimating ex- 45. Y. Huang, M. Hofer, R. D’Andrea, Offset-free model predictive control: A ball catching ap-
ternal forces on soft robots. IEEE Robot. Autom. Lett. 5, 5717–5724 (2020). plication with a spherical soft robotic arm, in 2021 International Conference on Intelligent
17. T. G. Thuruthel, E. Falotico, F. Renda, C. Laschi, Model-based reinforcement learning for Robots and Systems (IROS) (IEEE/RSJ, 2021), pp. 563–570.
closed-loop dynamic control of soft robotic manipulators. IEEE Transact. Robot. 35, 46. B. D. Anderson, J. B. Moore, Optimal Control: Linear Quadratic Methods (Courier Corpora-
124–134 (2019). tion, 2007).
18. A. Centurelli, L. Arleo, A. Rizzo, S. Tolu, C. Laschi, E. Falotico, Closed-loop dynamic control of 47. M. M. Coad, L. H. Blumenschein, S. Cutler, J. A. Reyna Zepeda, N. D. Naclerio, H. El-Hussieny,
a soft manipulator using deep reinforcement learning. IEEE Robot. Autom. Lett. 7, U. Mehmood, J.-H. Ryu, E. W. Hawkes, A. M. Okamura, Vine robots: Design, teleoperation,
4741–4748 (2022). and deployment for navigation and exploration. IEEE Robot. Autom. Mag. 27,
120–132 (2020).

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 12 of 13

SCIENCE ROBOTICS | RESEARCH ARTICLE

48. N. D. Naclerio, E. W. Hawkes, Simple, low-hysteresis, foldable, fabric pneumatic artificial control algorithm, prepared the figures, and wrote the manuscript. E.K. designed the
muscle. IEEE Robot. Autom. Lett. 5, 3406–3413 (2020). electronics, performed the experiments, and wrote the manuscript. A.B.C. ran the experiments,
49. L. H. Blumenschein, N. S. Usevitch, B. H. Do, E. W. Hawkes, A. M. Okamura, Helical actuation performed the analysis, and prepared the figures. P.C.C. performed experiments. I.M. advised
on a soft inflated robot body, in 2018 IEEE International Conference on Soft Robotics (Ro- the design of the modeling and control algorithm, the paper, and the experiments. E.W.H.
boSoft) (IEEE, 2018), pp. 245–252. advised the design of the robot, wrote the manuscript, and advised the paper and experiments.
50. A. Kazemipour, O. Fischer, Y. Toshimitsu, K. W. Wong, R. K. Katzschmann, A robust adaptive Competing interests: D.A.H. is the founder and CEO of Vine Medical Inc., which has no financial
approach to dynamic control of soft continuum manipulators. arXiv:2109.11388 [cs.RO] (26 interest in the subject work, nor did it support the investigation in any way. The other authors
February 2021). declare that they have no competing interests. Data and materials availability: All data
needed to support the conclusions of this paper can be found in the Supplementary Materials
Acknowledgments or at https://doi.org/10.5281/zenodo.8184777.
Funding: This work was supported in part by the NSF grant no. 1935327 and the ARO-MURI
W911NF1710306: From Data-Driven Operator Theoretic Schemes to Prediction, Inference, and Submitted 28 June 2022
Control of Systems. Part of the work performed by E.K. was funded via the Fulbright Foundation. Accepted 2 August 2023
Author contributions: D.A.H. designed the robots and test apparatus, wrote the manuscript, Published 30 August 2023
prepared the movies, and performed the experiments. M.J.B. designed the modeling and 10.1126/scirobotics.add6864

Downloaded from https://www.science.org on August 30, 2023

Haggerty et al., Sci. Robot. 8, eadd6864 (2023) 30 August 2023 13 of 13

 Control of soft robots with inertial dynamics
David A. Haggerty, Michael J. Banks, Ervin Kamenar, Alan B. Cao, Patrick C. Curtis, Igor Mezi, and Elliot W. Hawkes

Sci. Robot., 8 (81), eadd6864.

DOI: 10.1126/scirobotics.add6864

View the article online

https://www.science.org/doi/10.1126/scirobotics.add6864

Downloaded from https://www.science.org on August 30, 2023

Permissions
https://www.science.org/help/reprints-and-permissions

Use of this article is subject to the Terms of service

Science Robotics (ISSN ) is published by the American Association for the Advancement of Science. 1200 New York Avenue NW,
Washington, DC 20005. The title Science Robotics is a registered trademark of AAAS.
Copyright © 2023 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim
to original U.S. Government Works