Sample-efficient Model Predictive Control Design of Soft Robotics by

Bayesian Optimization
Anuj Pal, Tianyi He∗ , Wenpeng Wei

Abstract— This paper presents a sample-efficient data-driven With the appropriate models, different control techniques
method to design model predictive control (MPC) for cable- can be applied. Model predictive control can effectively ad-
actuated soft robotics using Bayesian optimization. Instead of dress the constraints of the states and inputs. This technique
modeling the complex dynamics of the soft robots, the proposed
approach uses Bayesian optimization to search the best-guessed has been used in a pneumatic humanoid robot [8], [9]. To
account for the model uncertainty, robust H∞ control is used
arXiv:2210.08780v1 [cs.RO] 17 Oct 2022

low-dimensional prediction model and its associated controller

to minimize the objective function of closed-loop responses. The to control distributed actuators in a segmented soft robotic
prediction model is updated by Bayesian optimization from arm [10]. A detailed review of the model-based control of
the closed-loop input-output data in each iteration. A linear soft robotics can be found in [11]. It is widely recognized that
MPC is then designed based on the updated prediction model,
and evaluated based on the closed-loop responses. Different the performances of model-based methods heavily rely on the
from directly searching controller parameters, the closed-loop accuracy of the model. Therefore, the model-based control
system stability, and inputs/outputs constraints can be easily approach has another obvious challenge in the trade-off
handled in the MPC design. After a few iterations, a convergent between accuracy/performance and adaptability/complexity.
solution of a (sub-)optimal controller can be obtained, which Recent advances in the data-driven (or learning-based)
minimizes the user-defined closed-loop performance index. The
proposed method is simulated and validated by a high-fidelity control method provide an alternative approach to the model-
simulation of a cable-actuated soft robot. The simulation results based control approach. Traditional machine learning (ML)
demonstrate that the proposed approach can achieve desired or deep learning (DL) methods can be applied to approximate
tracking controller for the soft robot without a prior-known the complex model without a prior understanding of the
model. system dynamics. The input-output data will be used to
establish a black-box model. However, the black-box model
I. INTRODUCTION is hard to be implemented in the control design. The closed-
loop system stability is still an unsolved issue. Therefore, its
Soft robots are made of continuously deformable materials applications in real-time control are limited. A review paper
or structures to mimic the biological continuum motions. The on the machine learning of soft robots provides more details
distributed softness or continuum brings unique compliance [12].
and flexibility compared to the conventional rigid robots. A promising middle-point method is to integrate data-
Therefore, the soft robots are showing advantages in the driven and model-based control to tackle the control of
applications of medical devices [1], human-robot interac- complex soft robots. This emerging method is considered
tions [2]. However, the softness leads to complex dynamics, a powerful candidate to address the complexity and non-
which challenges the development of appropriate control linearity. Some works of such methods are reported in the
algorithms to exploit its advantages. literature. An iterative learning model predictive control can
The soft robot is ideally an infinite-dimensional system. improve the model accuracy by gradually updating the model
The continuum structure theory [3] can accurately describe parameters using the data from repetitive processes [13].
the dynamics using PDE. However, these models are infinite- Koopman operator theory offers a data-driven method to
dimensional, so they are hard to be used for controller design. generate a linear model in the lifted high-dimensional space
In the model-based control approach, the key challenge is to approximate the nonlinear model. Linear control methods,
to develop a low-dimensional model but accurate enough to including linear MPC and LQG, can then be easily imple-
achieve good control performance. Many modeling methods mented [14], [15]. However, both the ML/DL and Koopman
of finite-dimensional approximations are reported, including operator approaches need a large number of experiments,
Piecewise Constant Strain (PCC) approach [4], Variable collecting data to learn a good model. This drawback leads
Strain (VS) approach [5], Finite Element Model (FEM) [6] to expensive costs in the learning process and hinders the
and its following model-reduction methods [7] . applications of soft robots.
Bayesian optimization is a promising approach that has
been proven to reduce the computational burden of identi-
Anuj Pal and Wenpeng Wei are with the Department of Mechanical
Engineering, Michigan State University, East Lansing, Michigan, 48824, fying the optimal parameters for any complex system. The
USA. Emails: [email protected], [email protected]. approach involves the use of a data-driven model and an
Tianyi He is with the Department of Mechanical and Aerospace actual system to iteratively improve the system knowledge
Engineering, Utah State University, Logan, Utah, 84342, USA. Email:
[email protected]. as per the desired performance function. The approach has
* Corresponding author. This work has been submitted to ACC 2023. been validated in easing the computational burden for various
problems ranging from parameter calibration and control
design. References [16], [17], [18], [19], [20], [21], [22]
implemented the Bayesian optimization framework in auto-
motive domain for performing the engine calibration. Apart
from automotive applications, the Bayesian optimization
approach has also been successfully implemented in other
(𝒙𝟏 , 𝒙𝟐 )
applications such as analog/rf circuit design [23], ground-
water reactive transport model [24], actuator modeling [25],
and designing natural-gas liquefaction plant [26]. All these
works have shown the capability of data-driven approaches
Fig. 1. Cable-actuated soft robot and positioning of its end-point on 2D
for modeling a complex system with a relatively simple plane.
model for computation.
In this paper, we present a sample-efficient data-driven
method to design the MPC of a cable-actuated soft robot will point at the x − y plane, and the tracking controller is
by Bayesian optimization. Due to the nonlinearity and expected to track the reference trajectory on the x − y plane.
complexity of the soft robot, the nonlinear mapping from The pre-defined task is to track a given reference trajectory
the model parameters and associated control parameters to rt in the time horizon (t = 1, 2, . . . , N), written as [1, N]. The
ultimate system performance is hard to model. We treat the control law (2)
nonlinear mapping as a Gaussian process and use Bayesian ut = g(yt , rt ) (2)
optimization to find the best prediction model that matches
with online input-output data. The best-guessed prediction is expected to enforce yt tracks reference rt . The objective
model is then used to design a linear MPC. The closed- is to find a control law that minimizes the weighted tracking
loop data is collected to evaluate the control performance errors and control inputs within the entire time horizon,
and update the prediction model by Bayesian optimization. as indicated in the problem formulation in (3). Q, R are
An optimal solution of the best-guessed model and MPC will weighting matrices for tracking errors and inputs, and Q ≥
be obtained after iterating a few experiments. 0, R > 0.
To the best knowledge of the authors, this is the first time N
that MPC for the soft robot has been designed using the min J = min
N
{ut }t=1 N
{ut }t=1
∑ (yt − rt )T Q(yt − rt ) + utT Rut (3a)
t=1
Bayesian optimization technique. The main contributions of
this work are three-fold: 1) Proposing a Bayesian optimiza- subject to: unknown : xt+1 = f (xt , ut ) (3b)
tion framework for the soft robot in identifying the optimal unknown : yt = h(xt , ut ) (3c)
reduced-order system dynamics approximation 2) Integrating umin ≤ ut ≤ umin (3d)
linear MPC with Bayesian optimization to achieve accurate ymin ≤ yt ≤ ymin (3e)
tracking control of soft robots 3) Demonstrating the track-
ing control performance and computational burdens of the The optimal solution of the control law is denoted as
proposed method. ut∗ = g∗ (yt , rt ). The unknown dynamics f (·) and h(·) make
The rest of this paper is organized as follows. Section II the mapping from control law g∗ impossible to be evaluated
formulates the problem and shows the overview of the by the unknown models. In the traditional model-based ap-
control scheme. Section III introduces Bayesian optimization proach or Koopman operator approach, a tremendous amount
and the algorithm for control design. Section IV then presents of data needs to be collected by conducting sufficiently many
the simulation results in a high-fidelity environment. At last, experiments at operating points covering the whole range of
conclusions are made, and future work is discussed. interest. Therefore, solving the optimal control law is very
expensive for the traditional methods that need to establish
II. P ROBLEM FORMULATION the mapping from control inputs to the ultimate system
performance. However, a given control law can be evaluated
Consider a soft robot actuated by cables, which is a multi- by conducting experiments/simulations, collecting the input-
input-multi-output (MIMO) system. Its nonlinear dynamic is output data, and analyzed in the cost function of (3). In other
described by the discrete-time nonlinear system (1) words, the control law can be sampled and improved through
xt+1 = f (xt , ut ) the online input-output data. The Bayesian optimization is a
(1) suitable tool to achieve this control objective without precise
yt = h(xt , ut )
dynamic models.
where xt , ut , yt denote the state, control inputs and outputs at Instead of using Bayesian optimization to directly auto-
time index t. The nonlinear function f (·) and h(·) are the tune the controller [27], Bayesian optimization is used in this
nonlinear dynamic model of the soft robot that is assumed paper to improve the approximated prediction model based
to be unknown. on the best ’guess’ in each iteration. The well-studied MPC
The soft robot is actuated by three cables embedded in the will be used to design the controller, and the performance
body, and the end point is installed by a laser. The end-point index will be evaluated on the closed-loop system with MPC.
 Output feedback constraints can be designed in a unified way of linear MPC.
Soft robot The closed-loop stability and robustness can be guaranteed
by the well-established linear MPC design; 3) This method
Linear MPC 𝑢𝑘 (𝑡)
Reference 𝒓𝒕 𝑲𝑴𝑷𝑪 (𝜽) Output
improves the computational efficiency than existing methods
(𝒙, 𝒚)
𝑦𝑘 (𝑡) because a relatively low-dimensional prediction model is
used in the MPC.
Learning-model
𝓟(𝜽) III. BAYESIAN OPTIMIZATIONS
Bayesian Bayesian optimization (BO) denotes a class of algorithms
optimization
for black-box global optimization problems in which data
Closed-loop
𝒌→𝒌+𝟏 Input-output data collection is expensive [9], and thus, only a few evaluations
Fig. 2. Overview of tracking control scheme of soft robotics by Bayesian
optimization and MPC. t - time index, k - iteration index. are possible. To deal with scarce data, BO (i) assumes
a probabilistic prior about the objective function and (ii)
chooses wisely the next combination of parameters to try
The overview of the proposed control scheme is shown in on the system according to a pre-established acquisition
Fig. 2. The prediction model P in the state-space represen- function.
tation to approximate the input-output relationship at a local Bayesian optimization models the objective function as a
operating point is denoted as GP model and utilizes the model estimations to select the
next function evaluation until it finds the global minimum
zt+1 = A p zt + B p ut or hits the maximum allowable iteration number. The BO
(4)
yt = C p zt + D p ut algorithm aims at approaching the optimal solution θ ∗ , by
which the linear MPC law KMPC (θ ) optimizes the control
where zt is the state of the prediction model, whose dimen-
performance on the entire task time horizon. Θ is the search
sion can be selected. A p , B p ,C p , D p are the prediction model
space of the parameter θ .
matrices, the elements of which are model parameters. For
ease of expression, the vector θ denotes model parameters θ ∗ = arg min J(θ ) (6)
θ ∈Θ
to be tuned that are stacked from all the matrices elements.
The dimension of the prediction model P(θ ) cannot be The evaluation of objective function J is conducted by
selected too high to avoid the curse of dimension. Normally, the closed-loop experiments/simulation, which samples the
the dimension of θ is chosen as 10 − 20. parameter θ in the search domain. Two key steps involved in
At each iteration, the MPC uses the updated prediction performing the BO are the model development and then for-
model to design the control inputs in the prediction horizon mulating the acquisition function to intelligently search for
Tp . Note that the prediction horizon is usually smaller than the optimal solution. A Gaussian process model is developed,
the entire time horizon [1, N] of the tracking control task. which predicts both mean and uncertainty quantification.
At each time step, the control input is obtained from the Both these parameters are then combined in a way to perform
MPC design by solving the following optimization, and the efficient exploration and exploitation of the design/search
controller is denoted as KMPC (θ ). The control inputs at the space to identify the optimal solution.
entire time horizon can be obtained by repeating solving (5) A. Gaussian process
until the task is complete.
Gaussian process regression (GPR) is used to formulate
Tp
the latent function f capturing the interaction between system
min
Tp
∑ (yt − rt )T Q(yt − rt ) + utT Rut (5a) performance (cost function J) based on input parameters to
{ut }t=1 t=1
be optimized (θ ).
subject to: updated prediction model P(θ ) The expression for the latent function is given as
umin ≤ ut ≤ umin f ∼ GP(m, k) (7)
(5b)
ymin ≤ yt ≤ ymin
where, f is the latent function, capturing the input-output
The unknown system dynamic equations (3) is replaced behavior, and m and k are mean and covariance. If the system
by the ’guessed’ prediction model (4), and it will be used to is corrupted with noise, then the output can be written as:
design a linear model predictive control. Using the closed-
y = J(θ ) = f + ε, where ε ∼ N(0, σn2 ) (8)
loop input-output data, the performance in the entire time
horizon can be evaluated. The Bayesian optimization is then Here, ε is the measurement noise considered as Gaussian
used to tune the prediction model P(θ ) such that the input- with zero mean and variance σn2 . Combining both the terms
output data can be well matched. in the above expression, the output cost can be written as:
The advantages of this method are three-fold: 1) The
J(θ ) ∼ GP(m, k + σn2 δ pq ) (9)
precise model is not needed, and the prediction model will be
improved in each iteration based on the evaluations of closed- where δ pq = 1 iff p = q is the Kronecker’s delta. The output is
loop system performance; 2) The controller subject to the a GP with mean m and covariance k + σn2 δ pq . Several mean
and covariance functions are available to choose to develop function utilizes both mean and variance estimates to balance
the model. For this work, m = 0 is selected, which is a the exploration-exploitation of the design space, which is
very common mean function used for model development. crucial for finding global optima.
A Gaussian covariance function is selected, which introduces There are various ways to formulate the acquisition func-
several hyperparameters to the model. Hyperparameters are tion [ref]. For this work, we are using Expected Improvement
the unknown parameters that are optimized to get the best (EI) [29]. It is derived by formulating an improvement func-
possible fit of the function to the available data. The op- tion I(x) and then taking its expectation. The improvement
timal hyperparameters are obtained by maximizing the log function for the current problem is defined as:
marginal likelihood function.
Let us assume that we initially have N p points to fit the I(θ ) = max{0, Jm (θ ) − J ∗ } (13)
Gaussian Process model. Therefore, the output cost function where, J ∗ = maxx∈N J is the best value of the cost function
with measurement noise can be written as: obtained so far, and Jm (θ ) is the cost predicted by the
Gaussian model. Taking the expectation of the improvement
J(θ ) ∼ GP(m, K + σn2 I) (10)
function, we obtain the expression for EI, given as:
where, θ represent an input matrix, m is the vector with  J ∗ − J (θ ) 
m
mean values and K is the covariance matrix given as Ak (θ ) = EI(θ ) = (Jm (θ ) − J ∗ )Φ
Jσ (θ )
(k(xi , x j )) ∀ i, j ∈ [1, N p ].  J ∗ − J (θ )  (14)
After fitting the model, it is important to perform predic- m
+ Jσ (θ )φ
tions at any unknown input point. Let J(θ ) be the known Jσ (θ )
outputs, and J(θ 0 ) is the unknown output that we want to Here, Φ and φ are the standard normal cumulative distri-
predict at input θ 0 using the Gaussian Process model. Since bution function and probability distribution function, respec-
the process is Gaussian, we can write the joint distribution tively. The goal here is to maximize the EI function, which
of both known and unknown points as: depends on the mean value and variance estimate from the
model, shown in the first and second part of the equation

J(θ )
 
m
 
K + σn2 I k(θ , θ 0 ) 
 (14), respectively.
∼N , T (11) The overall Bayesian optimization algorithm to design
J(θ 0 ) m(θ 0 ) k (θ , θ 0 ) k(θ 0 , θ 0 ))
MPC for soft robot is summarized as follow.
Here, J(θ ) is the training set, and J(θ 0 ) is the unknown
output. k(θ , θ 0 ) = k(θi , θ 0 ) ∀ i ∈ [1, N p ] represent the covari- Algorithm 1: Bayesian optimization of MPC of soft
ance calculation between known data set θ and unknown robots
test point θ 0 . Using the above expression, the conditional Data: Q ≥ 0, R > 0, umin , umax , ymin , ymax , N > Tp > 0,
distribution of J(θ 0 ) given J can easily be obtained. For search space Θ, null dataset Θset , Jset .
the sake of brevity, we have eliminated the mathematical Result: θ , i.e. A p , B p ,C p , D p
discussion for the model development. Readers are encour- for k < kmax do
aged to go through the reference [28] for more details. After • Update the mean and variance of the objective
performing mathematical calculations, the model estimates function J(θ )
at any unknown point (θ 0 ) can be written as: • Optimize the acquisition functin to determine next
evaluation point θk+1
Jm (θ 0 ) = m(θ 0 ) + k(θ , θ )T K−1 (J(θ ) − m) θ ∗ ← arg minθ ∈Θ J(θ )
(12) • Design linear MPC and conduct experiments using
Jσ 2 (θ 0 ) = k(θ 0 , θ 0 ) − k(θ , θ 0 )T K−1 k(θ , θ 0 )
KMPC (θk+1 ) and measure J(θ ˆ )
where, Jm (θ 0 ) and Jσ 2 (θ 0 ) are the posterior mean and • Augment the data set
variance estimate from the Gaussian process at any unknown Θset,k+1 ← Θset,k ∪ θk+1
point θ 0 . The variance estimate becomes crucial for intel- Jset,k+1 ← Jset,k ∪ J (θk+1 )
ligently identifying the optimal region. The next section end
discusses the acquisition function, which combines both θopt = θopt,kmax = arg minθ ∈Θ J(θ );
mean and variance estimates from the model to direct the
Bayesian optimization algorithm toward the optimal region.
B. Acquisition function: Expected improvement IV. S IMULATION R ESULTS AND DISCUSSIONS
The expected improvement seeks to obtain the next evalua- To validate the effectiveness of the proposed method in
tion point, where the objective function is expected to be im- the control of soft robots, several simulations are conducted
proved most over the best cost function value J ∗ collected so in a high-fidelity simulation environment - RoBoSim. The
far. One major advantage of the Gaussian process model is its simulation platform can simulate both the static and dy-
mean and variance estimation. The mean value suggests the namics of the continuum and soft robots [30], [5]. The
region of the optimal solution, whereas the variance estimate simulation of the soft robot is based on the Cosserat-beam
shows the uncertainty in unexplored regions. The acquisition theory. The beam shape is characterized by the nonlinear
parameterization by strain fields, and the order-reduction is The responses of the best design at x1 and x2 are shown
based on a functional basis of strain modes. While remaining in Fig.4. From the initial condition point, the states can con-
geometrically exact, the simulation platform provides us with verge rapidly to the origin point. Because the measurement
a minimal set of ordinary differential equations that can is subject to measurement noises, the position responses are
be easily implemented for analysis and control design. The fluctuating around 0.
accuracy of the model is comparably accurate to the well-
0.01
validated finite element method in the nonlinear structural
statics and dynamics. 0.005
A. Simulation setup

x1
0
The Bayesian optimization is conducted using the
Bayesian Optimization Toolbox bayesopt by MATLAB. The -0.005
soft robot in SoRoSim and the Bayesian optimization are
-0.01
co-simulated with the sampling time 50 ms in simulation 0 5 10 15 20
time of 20 seconds. The soft material of the robot body is
0.04
selected as PDMS. The outputs x1 and x2 displacement are
constrained by [−0.1, 0.1]. The control input is constrained

x2
0.02
by [−10, 10]N. The parameters of robot and Bayesian opti-
mization are summerized in the Table I.
0
TABLE I
0 5 10 15 20
S UMMARY OF PARAMETERS IN THE SIMULATION
Time/s
BO parameter value soft robot values
Fig. 4. Position tracking of the origin under perturbation.
white noise variance σ 2 0.012 density 1000 kg/m3
maximum iteration kmax 100 Young’s modulus 1e6 N/m2
number of seed point 10 Poisson’s ratio 0.5 2) Scenario 2: tracking a circular shape: The reference
GPA active set size 300 length 0.6 m trajectory to be tracked is a circle with a radius of 0.05
prediction horizon Tp 10 radius 0.1 m m. The equation is expressed as x12 + x22 = 0.052 . The con-
weighting matrix Q 103 I3 actuation cable 3
weighting matrix R I3 damping 0.11e5 Pa.s vergence of the performance relative to iterations is plotted
in Fig. 5. Although it takes longer iterations to achieve the
optimal, the sub-optimal can be obtained in a few iterations.
B. Simulation results The responses of the positions x1 and x2 are plotted in
1) Scenario 1: position initialization to origin: In this Fig.6. The red curve is the response from the actual soft
simulation scenario, the soft robot is driven to position to robot, and the blue curve is the desired circular shape. In
the origin point to initialize the system. The tracking target the time horizon of the tracking task, the soft robot starts
is the constant reference trajectory at origin point (0, 0). from the origin point and converges to the desired circle.
Gradually, the position curve matches the desired trajectory.
12

10 12
Current point
8 Current best point
Overall best point 10 Current point
6 Current best point
8 Overall best point
4
6
2

0 4

-2 2

-4
0 20 40 60 80 100 0

-2
Fig. 3. Performance index in iterations of Bayesian optimizaion. 0 20 40 60 80 100

The performance index J at the iterations is plotted in Fig. 5. Performance index in iterations of Bayesian optimization.
the Fig.3. The best performance point is achieved at the
green point of the 42th iteration. At the early stage of
convergence, a sub-optimal solution can be achieved by V. C ONCLUSIONS AND FUTURE WORK
around 10 iterations. This indicates that a good controller can This paper presents a systematic framework for approx-
be tuned by the Bayesian optimization in a few iterations. imating the complex dynamics of a soft robot using the
 0.05
[9] P. Hyatt and M. D. Killpack, “Real-time nonlinear model predictive
0.04 control of robots using a graphics processing unit,” IEEE Robotics and
Real trajectory
Automation Letters, vol. 5, no. 2, pp. 1468–1475, 2020.
0.03 Desired trajectory [10] A. Doroudchi, S. Shivakumar, R. E. Fisher, H. Marvi, D. Aukes, X. He,
S. Berman, and M. M. Peet, “Decentralized control of distributed
0.02 actuation in a segmented soft robot arm,” in 2018 IEEE Conference
on Decision and Control (CDC). IEEE, 2018, pp. 7002–7009.
0.01 [11] C. Della Santina, C. Duriez, and D. Rus, “Model based control of soft
robots: A survey of the state of the art and open challenges,” arXiv
x2

0
preprint arXiv:2110.01358, 2021.
-0.01 [12] D. Kim, S.-H. Kim, T. Kim, B. B. Kang, M. Lee, W. Park, S. Ku,
D. Kim, J. Kwon, H. Lee et al., “Review of machine learning methods
-0.02 in soft robotics,” Plos one, vol. 16, no. 2, p. e0246102, 2021.
[13] Z. Q. Tang, H. L. Heung, K. Y. Tong, and Z. Li, “A novel iterative
-0.03 learning model predictive control method for soft bending actuators,”
in 2019 International Conference on Robotics and Automation (ICRA),
-0.04 2019, pp. 4004–4010.
[14] D. Bruder, C. D. Remy, and R. Vasudevan, “Nonlinear system iden-
-0.05
-0.05 0 0.05 tification of soft robot dynamics using koopman operator theory,” in
x1 2019 International Conference on Robotics and Automation (ICRA).
IEEE, 2019, pp. 6244–6250.
[15] D. Bruder, X. Fu, R. B. Gillespie, C. D. Remy, and R. Vasudevan,
Fig. 6. Position trajectory tracking of the desired circular shape trajectory. “Data-driven control of soft robots using koopman operator theory,”
IEEE Transactions on Robotics, vol. 37, no. 3, pp. 948–961, 2020.
[16] A. Pal, Y. Wang, L. Zhu, and G. G. Zhu, “Engine calibration
optimization based on its surrogate models,” in Dynamic Systems and
Bayesian optimization (BO) framework. The study integrates Control Conference, vol. 59155. American Society of Mechanical
the framework to identify the low-dimensional linear dy- Engineers, 2019, p. V002T12A002.
namic model, which is then used with a linear MPC con- [17] L. Zhu, Y. Wang, A. Pal, and G. Zhu, “Engine calibration using
global optimization methods with customization,” SAE Technical
troller to achieve the desired tracking performance. For BO, Paper 2020-01-0270, Tech. Rep., 2020.
the Gaussian process is used for model development, and [18] A. Pal, L. Zhu, Y. Wang, and G. G. Zhu, “Constrained surrogate-
expected improvement (EI) is used as an acquisition function based engine calibration using lower confidence bound,” IEEE/ASME
Transactions on Mechatronics, vol. 26, no. 6, pp. 3116–3127, 2021.
to perform exploration and exploitation of the search/design [19] A. Pal, Y. Wang, L. Zhu, and G. G. Zhu, “Multi-objective surrogate-
space to obtain optimal solutions. The results demonstrate assisted stochastic optimization for engine calibration,” Journal of
superior performance using our proposed approach in iden- Dynamic Systems, Measurement, and Control, vol. 143, no. 10, p.
101004, 2021.
tifying the approximate system dynamics, which is then used [20] A. Pal, L. Zhu, Y. Wang, and G. G. Zhu, “Multi-objective stochastic
with linear MPC to achieve desired system performance. bayesian optimization for iterative engine calibration,” in 2020 Amer-
The future work includes investigating the robustness of the ican Control Conference (ACC), 2020, pp. 4893–4898.
[21] T. Gutjahr, T. Kruse, and T. Huber, “Advanced modeling and optimiza-
proposed method and experimental validations by real-time tion for virtual calibration of internal combustion engines,” in NDIA
implementations. Ground Vehicle Systems Engineering and Technology Symposium,
2017.
[22] J. Tang, A. Pal, W. Dai, C. Archer, J. Yi, and G. Zhu, “Stochastic
R EFERENCES bayesian optimization for predicting borderline knock,” International
Journal of Engine Research, p. 14680874211065237, 2021.
[1] M. Runciman, A. Darzi, and G. P. Mylonas, “Soft robotics in min- [23] W. Lyu, F. Yang, C. Yan, D. Zhou, and X. Zeng, “Multi-objective
imally invasive surgery,” Soft robotics, vol. 6, no. 4, pp. 423–443, bayesian optimization for analog/rf circuit synthesis,” in Proceedings
2019. of the 55th Annual Design Automation Conference, 2018, pp. 1–6.
[2] P. Polygerinos, N. Correll, S. A. Morin, B. Mosadegh, C. D. Onal, [24] J. Zhou, X. Su, and G. Cui, “An adaptive kriging surrogate method for
K. Petersen, M. Cianchetti, M. T. Tolley, and R. F. Shepherd, “Soft efficient joint estimation of hydraulic and biochemical parameters in
robotics: Review of fluid-driven intrinsically soft devices; manufac- reactive transport modeling,” Journal of contaminant hydrology, vol.
turing, sensing, control, and applications in human-robot interaction,” 216, pp. 50–57, 2018.
Advanced Engineering Materials, vol. 19, no. 12, p. 1700016, 2017. [25] B. Liu, V. Grout, and A. Nikolaeva, “Efficient global optimization of
[3] D. Trivedi, A. Lotfi, and C. D. Rahn, “Geometrically exact models for actuator based on a surrogate model assisted hybrid algorithm,” IEEE
soft robotic manipulators,” IEEE Transactions on Robotics, vol. 24, Transactions on Industrial Electronics, vol. 65, no. 7, pp. 5712–5721,
no. 4, pp. 773–780, 2008. 2017.
[4] C. Della Santina, A. Bicchi, and D. Rus, “On an improved state [26] W. Ali, M. S. Khan, M. A. Qyyum, and M. Lee, “Surrogate-assisted
parametrization for soft robots with piecewise constant curvature and modeling and optimization of a natural-gas liquefaction plant,” Com-
its use in model based control,” IEEE Robotics and Automation Letters, puters & Chemical Engineering, vol. 118, pp. 132–142, 2018.
vol. 5, no. 2, pp. 1001–1008, 2020. [27] M. Neumann-Brosig, A. Marco, D. Schwarzmann, and S. Trimpe,
“Data-efficient autotuning with bayesian optimization: An industrial
[5] F. Boyer, V. Lebastard, F. Candelier, and F. Renda, “Dynamics of
control study,” IEEE Transactions on Control Systems Technology,
continuum and soft robots: A strain parameterization based approach,”
vol. 28, no. 3, pp. 730–740, 2019.
IEEE Transactions on Robotics, vol. 37, no. 3, pp. 847–863, 2020.
[28] C. K. Williams and C. E. Rasmussen, Gaussian processes for machine
[6] S. Grazioso, G. Di Gironimo, and B. Siciliano, “A geometrically exact
learning. MIT press Cambridge, MA, 2006, vol. 2, no. 3.
model for soft continuum robots: The finite element deformation space
[29] D. R. Jones, M. Schonlau, and W. J. Welch, “Efficient global optimiza-
formulation,” Soft robotics, vol. 6, no. 6, pp. 790–811, 2019.
tion of expensive black-box functions,” Journal of Global optimization,
[7] S. Sadati, S. E. Naghibi, L. Da Cruz, and C. Bergeles, “Reduced-order vol. 13, no. 4, pp. 455–492, 1998.
modeling and model order reduction for soft robots,” 2021. [30] F. Renda, C. Armanini, V. Lebastard, F. Candelier, and F. Boyer,
[8] C. M. Best, M. T. Gillespie, P. Hyatt, L. Rupert, V. Sherrod, and M. D. “A geometric variable-strain approach for static modeling of soft
Killpack, “A new soft robot control method: Using model predictive manipulators with tendon and fluidic actuation,” IEEE Robotics and
control for a pneumatically actuated humanoid,” IEEE Robotics & Automation Letters, vol. 5, no. 3, pp. 4006–4013, 2020.
Automation Magazine, vol. 23, no. 3, pp. 75–84, 2016.