Supplementary Materials for

Control of soft robots with inertial dynamics

David A. Haggerty et al.

Corresponding author: David A. Haggerty, [email protected]

Sci. Robot. 8, eadd6864 (2023)

The PDF file includes:

Discussion
Figs. S1 to S5
Legends for movies S1 to S5

Other Supplementary Material for this manuscript includes the following:

Movies S1 to S5
Supplementary Discussion

Testing Apparatus The experiments conducted in the framework of the study are performed
on an experimental system shown in Fig. S1.

A 1.8x1.8x1.5 m 80/20 frame was assembled to support the testing apparatus. To this frame
the motion capture cameras, soft arm, and control hardware were affixed. Information about the
position and shape of the manipulator is gathered via motion capture (PhaseSpace Inc. Impulse
X2E). This investigation utilized the motion capture system with 8 detectors (cameras) and 4

sets of trackers evenly spaced along the backbone of the soft arm; four LEDs are attached along
the axis of each muscle. The same motion capture system was used for both data collection used
in offline model construction and for closed-loop position feedback in control e xperiments. In

closed-loop experiments that are performed without predefined trajectory, additional four LED
trackers are mounted on an external object (a pole), their coordinates are averaged in real-time

in order to determine the central point, which then served as an arbitrary reference generator.

Festo VEAB-L-26-D2-Q4-V1-1R1 proportional pressure regulators with 0.01 to 2 bar out-

put range and approximately 15 liters/min of flow a t 1 b ar p ressure, a re u sed t o c ontrol the

pressure in the arm’s muscles. The body is held to a constant pressure of approximately 1.5 bar
The software used for running the system was LabVIEW 2019 with myRIO toolkit and real-

time module, whereas LabVIEW Python node is used to acquire the real-time data from motion
capture system. These information is then fed trough the fast network protocol to a myRIO

1900 control hardware. The same control hardware is also used to drive the pressure valves
whereas an additional circuitry based on operational amplifiers is used to adjust 0-5V voltage
levels generated by MyRIO hardware to be compatible with used proportional valves whose

input range is 0-10V. Exhaust air ports of the valves are connected to vacuum so as to improve
the dynamical response of the system.
Model Performance Metrics We perform a convergence study on the reconstruction power
of our Koopman models as a function of the number of snapshots for a range of observables.

This process allowed us to develop a dictionary of observables suitable for our system. Given
a particular choice of observables and number of training samples, we build the corresponding
linear input-output system with A, B, and C matrices. This linear model is applied to N = 8000

samples of sinusoidal verification data over a range of deflection amplitudes and speeds. These
particular samples are not included in the training data in order to give us a fair evaluation of

the predictive power of our models. The linear system produced via (14) and (5) evaluate the

evolution of these initial conditions over a single time step. The single-step reconstruction error

is given by
kx+,
i
predict
− x+,
i
actual
k2
ei = .
L
where x+,
i
actual
is the evolution of xi measured by the motion capture system, x+,
i
predict
is the

evolution predicted by the DMD model, and L is the length of the soft arm. We use the root

mean square (RMS) of the individual ei errors to score our model:

Koopman Spectral Quantities We are often interested in the spectral properties of the Koop-
man operator because they give us physical information about the multiple coupled time-dependent
processes inherent to our system. DMD can be used to approximate the discrete part of this

spectrum [30]. We seek the triplet (λi , φi (z(x)), vi ) of Koopman eigenvalues, eigenfunctions,
and modes, respectively. The eigenvalues and Koopman modes are simply the eigenvalues and
eigenvectors of the DMD matrix A. The Koopman modes in Fig. 2 are added to the time av-

erage mode associated with λ = 1 to give an impression of the effect of the mode on the soft
arm. Computation of the eigenfunctions requires wi which are the eigenvectors of the conjugate
transpose of A. After these are normalized so that hvi , wj i = δij , the eigenfunctions are given

by the complex inner product φi (z(x)) = hz(x), wi i. The eigenfunctions are shown here as
functions of the lifted state z(x). Their magnitude |φi (z(x))| is called the “mode power” and

gives the relative importance of the ith Koopman mode to the dynamics when the state is x. In
order to compare the influences of the Koopman modes to the dynamics, their colorings in Fig.
2 are shown scaled to the maximum value of the mode power attained over the entire training
data set.

Reconstruction of sinusoidally forced motion After choosing time delay observables, we at-

tempted to reconstruct the movement of the soft arm under a sinusoidal inputs with six different

frequencies (0.1, 0.2, 0.4, 0.8, 1, and 1.1Hz). We begin this process by providing a step input to
the muscles that corresponds to a static position on the sinusoid with an arbitrary phase, and then

complete two revolutions at a given frequency before incrementing up in speed. The position

of the physical system was recorded via motion capture system, and these inputs were pro-
vided to our above-developed model. This process was repeated for low (similar to [42, 34, 35],

approximately 15◦ ), medium (similar to [13, 14], approximately 25◦ ), and high deflection (the
single-actuator maximum of our system (robot 1), approximately 110◦ ), with results reported in

Fig. S2.

Disturbance Rejection Finally, to evaluate the disturbance rejection capabilities of our sys-

tem and to further distinguish the contribution of the static Koopman pregain, G, and the dy-
namic Koopman LQR gain, K, we commanded both stationary and circular references for the
soft arm tip and subjected the system to disturbances. The control effort was recorded and com-

pared to the control effort expected from the pregain term alone. Given a static reference, the
control effort from the pregain alone is constant in time. The results of the tests are shown in
Fig. S5, capturing the contribution of K, proportional to the disturbance.
 Control circuitry

Test superstructure

Motion capture
cameras

Robot backbone

Pneumatic muscles

LED trackers

Figure S1: Schematic representation of the experimental setup and its components. The
soft arm is mounted from above to the test superstructure. The soft arm backbone provides stiff-
ness and the pneumatic muscles generate movement. Four layers of LED trackers are tracked
by motion capture cameras positions around the arm.
 
  



       

  


       




       


Figure S2: Koopman reconstruction of circular motion. The dynamic Koopman model is
given a collection of sinusoidal inputs with a range of amplitudes and speeds and is tasked with
reconstructing the motion of the soft arm. The true trajectories are shown in dashed blue, and
the high (A), medium (B), and low (C) deflection reconstructions are given in red, yellow, and
purple, respectively. The reconstruction is restarted every time the frequency changes. The
reconstruction agrees with the true frequency, but is missing some of the amplitude in the fast
regime. The static Koopman operator and feedback control account for the improvement in
performance between this plot and Fig. 3
 
 
      


  
 
  
               
 
  
  




  
 
  
               
  


  
 
  
               
 

Figure S3: Y and Z components of the real-time closed-loop reference tracking experi-
ments. The soft arm tracked circular reference trajectories in the X-Y plane with frequencies
ranging from 0.1 to 1.1 Hz (0.2Hz step) at: (A,D) high, (B,E) medium, and (C,F) low de-
flection magnitudes. Plots show the Y (left column) and Z (right column) positions over time
compared to their respective references. The commanded references are dashed lines and the
control results are solid lines.
 Soft Robot System x(t) 


Training u Motion x
Inputs Capture

Training Path
Control Path n t
Common Path Control in Lifted Space
z(x)
Lifted Controller
Lifting zref
G K
xref
Pregain LQR EDMDc
Reference

Figure S4: Block diagram of the system, training method and K-LQR control approach.
Training inputs representing voltage signals are fed into pressure valves and 3D positions of the
soft arm are measured by using a motion capture system. The lifting procedure of the position
data provides the inputs needed to determine the Koopman model of the system and to calculate
optimal lifted controller parameters. The position of the soft arm is finally controlled in 3D
space by using obtained K-LQR.
  




  







   

 
  
­
€‚
  
­




        
 

Figure S5: Demonstration of our controller’s ability to reject impulse disturbances in real-
time. A) The position over time of the end effector is shown relative to a predefined reference.
The soft arm returns to the reference position after three large disturbances are applied. B)
The magnitude of the commanded control effort is shown. Note that the static Koopman input
component comes from the pregain term and is constant because the static reference doesn’t
change. The dynamic Koopman input adapts in real time to the disturbances.
Supplementary Movies

Movie S1: Movie S1 shows the performance of Robot 1 commanded to follow an arbitrary
trajectory throughout the workspace, ranging from low to high deflection. This Movie can be
found at https://drive.google.com/file/d/1njIB3zbG3J6U8v65Bim2_pXq-AcICDYn/

view?usp=share_link.

Movie S2: Movie S2 shows the performance of Robot 1 commanded to follow a series of

sinusoidal trajectories at increasing speeds (0.1 − 1.1Hz frequency) and increasing deflections,

as shown in Figures 3 and S3. This Movie can be found at https://drive.google.com/

file/d/1Nw1leGrLz1pqR6DUGd2UgYViBSGH6Srk/view?usp=share_link.

Movie S3: Movie S3 shows our system completing two real-world tasks: first, catching a

swinging ball that enters the workspace from two different directions; second, throwing a ball

into a bin positioned at two different locations in the workspace. This Movie can be found at

https://drive.google.com/file/d/1gj-WHwkOQRqoyR2_Rnnj38b4NF95zFo1/

view?usp=share_link.

Movie S4: Movie S4 shows our training, modeling, and control sequence for Robot 1.

This sequence uses 5 minutes of step inputs, approximately 3 seconds of model and controller
computation, followed by arbitrary reference tracking. This Movie can be found at https://
drive.google.com/file/d/1LwPXZfLh3xPuvP1YxFhmmarpRg44cq5P/view?usp=

share_link.

Movie S5: Movie S5 shows our training, modeling, and control sequence for Robot 2. The

same sequence is provided as in Movie S4, but with a robot capable of nearly 180◦ of deflection.
Arbitrary reference tracking is successful across the entire range of deflections. This Movie can
be found at https://drive.google.com/file/d/1dZw4ZKY7_9YPN2Y7mY4Mpwawpe9DZA_

U/view?usp=share_link.