Mechatronics 106 (2025) 103288

Contents lists available at ScienceDirect

Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics

Fixed-structure sampled-data feedforward control for multivariable motion

systems✩
Masahiro Mae a ,∗, Max van Haren b , Koen Classens b , Wataru Ohnishi a , Tom Oomen b,c ,
Hiroshi Fujimoto a
a
The University of Tokyo, Japan
b Eindhoven University of Technology, The Netherlands
c Delft University of Technology, The Netherlands

ARTICLE INFO ABSTRACT

Keywords: Increasing performance requirements in high-precision mechatronic systems lead to a situation where both
Feedforward control multivariable and sampled-data implementation aspects need to be addressed. The aim of this paper is to
Reference tracking develop a design framework for a multi-input multi-output feedforward controller to improve continuous-
Multi-input multi-output system
time tracking performance through learning. The sampled-data feedforward controller is designed with
Sampled-data control
physically interpretable tuning parameters using a multirate zero-order-hold differentiator. The developed
Multirate inversion
Iterative learning control
approach enables interaction compensation for multi-input multi-output systems and the feedforward controller
parameters are updated through learning. The performance improvement is experimentally validated in a
multi-input multi-output motion system compared to the conventional feedforward controllers.

1. Introduction enables low-complexity parameterization with physical interpretabil-

ity and flexibility for varying references. In conventional approaches
Feedforward control is essential in increasing performance require- [14–16], the differentiator for the basis function design is implemented
ments for motion control of high-precision mechatronic systems in by the backward differentiator, and the sampled-data characteristic is
industries such as semiconductor lithography systems [1–3], wire bon- not considered. The gap between the backward differentiator and the
ders [4], atomic force microscopy [5], machine tools [6], industrial zero-order-hold characteristics of the sampled-data system results in the
robots [7], magnetic bearing [8], boost converters [9], 2D/3D print- limitation of the control performance in continuous time.
ers [10–12], and CT scanners [13]. Iterative Learning Control (ILC) Sampled-data feedforward control improves the continuous-time
is one of the algorithms to update the feedforward controller by the tracking performance of high-precision mechatronic systems where the
error data of the previous iteration, and the error is reduced through sampling frequency is not sufficiently high compared to the motion
learning. To overcome the limitation of the interpretability in ILC such profile [17]. In industrial control applications, the controlled system
as a Finite Impulse Response (FIR) filter-based structure in [1,11], is discretized by sampler and zero-order-hold and these characteristics
it is beneficial in industries that the controller consists of physically
should be considered in feedforward controller design to improve not
interpretable tuning parameters to achieve both intuitive tuning and
only on-sample but also intersample performance [18]. State-tracking
data-driven learning.
feedforward control [19,20] and ILC [21] with multirate inversion
Physical interpretability and intuitive tuning of the data-driven
can improve the continuous-time tracking performance in sampled-
feedforward controller are desirable in industrial applications. It is
data systems. These controllers enable on-sample state-tracking and it
achieved by structure analysis of the controlled system [14,15], and
leads to physically natural intersample behavior. For the application
the controller can be parameterized intuitively by linear combinations
to the complex mechatronic systems, there is no guarantee of a perfect
with tuning parameters and basis functions [10,16]. Basis functions
typically consist of a reference signal and its derivatives [16] and model for model inversion and there must be a modeling error between
nonlinear functions such as friction compensation [6]. This structure the identified model and the actual system. The multirate feedforward

✩ This paper was recommended for publication by Associate Editor Cheng-Wei Chen.
∗ Corresponding author.
E-mail address: [email protected] (M. Mae).

https://doi.org/10.1016/j.mechatronics.2024.103288
Received 5 April 2024; Received in revised form 9 December 2024; Accepted 20 December 2024
Available online 8 January 2025
0957-4158/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Mae et al. Mechatronics 106 (2025) 103288

control also can be extended to the multivariable systems [22,23].

However, the complex mechatronic systems are represented by the
multivariable model in many cases and it results in many tuning
parameters for the inverse-based controller design.
Although important contributions have been made to improve the
performance and interpretability of feedforward control, the sampled-
data characteristics with sampler and zero-order-hold are not taken into
account in the basis function design, and the structure of multi-input
multi-output (MIMO) feedforward control is not discussed in the con- Fig. 1. Tracking control diagram. The continuous-time system 𝑮 is controlled by the
feedforward controller 𝑭 (𝜽) and the discrete-time feedback controller 𝑲 with sampler 
text of intuitive tuning and leaning from experimental data. The aim of
and zero-order-hold . The objective is to minimize the continuous-time tracking error
this paper is to design the MIMO feedforward controller to improve the 𝒆(𝑡). The solid and dotted lines denote the continuous-time and discrete-time signals,
continuous-time tracking performance through learning. Compared to respectively.
conventional approaches, the developed basis function design considers
the sampled-data characteristics. In this paper, the feedforward con-
troller is parameterized with basis functions for MIMO motion systems Definition 1 (Sampler). The sampler  with sampling time 𝑇𝑠 is defined
and it enables physical interpretation of the feedforward controller as
parameters and analytic solution of data-driven parameter tuning. The
 ∶ 𝒓(𝑡) ↦ 𝒓[𝑘], 𝒓[𝑘] = 𝒓(𝑘𝑇𝑠 ). (1)
present paper substantially extends the preliminary result in [24],
including the comparison between the exact model inversion based on
the multirate feedforward control, the generalization for the application Definition 2 (Zero-order-hold). The zero-order-hold  with sampling
in MIMO motion systems, the data-driven tuning algorithm, and the time 𝑇𝑠 is defined as
experimental validation.  ∶ 𝒖[𝑘] ↦ 𝒖(𝑡), 𝒖(𝑘𝑇𝑠 + 𝜏) = 𝒖[𝑘], 𝜏 = [0, 𝑇𝑠 ). (2)
The main contribution of this paper is the fixed-structure feedfor-
ward controller design considering sampled-data characteristics and
The control objective in this paper is to minimize the continuous-
interactions in MIMO motion systems. The contributions include:
time tracking error 𝒆(𝑡). Traditionally, the conventional discrete-time
controller only focuses on the on-sample performance with the discrete-
Contribution 1. Discrete-time basis functions are designed for continuous-
time tracking error 𝒆[𝑘]. To improve the continuous-time tracking error
time reference considering sampled-data characteristics to improve 𝒆(𝑡), not only on-sample but also intersample performance should be
continuous-time tracking performance. considered. The improvement of continuous-time tracking performance
is defined as follows.
Contribution 2. ILC with basis functions is formulated with physically
Definition 3 (Continuous-time Tracking Performance). The optimization
interpretable tuning parameters considering the dynamics and interaction of problem to improve the continuous-time tracking performance in the
MIMO motion systems. sampled-data motion system is defined as
The outline is as follows. In Section 2, the problem is formulated. In minimize ‖𝒆[𝑘]‖𝑾 (3)
Section 3, the basis function design considering continuous-time track- 𝜽

ing performance is described, constituting Contribution 1. In Section 4, subject to RMS(𝑒𝑙 [𝑘]) ≈ RMS(𝑒𝑙 (𝑡)), (4)
∀𝑙
the continuous-time tracking performance of the developed approach
is demonstrated by the simulation in SISO motion systems compared where 𝜽 is the tuning parameter, ‖ ∙ ‖𝑾 is the weighted 2-norm with
to the conventional approaches. In Section 5, ILC with basis functions the weighting matrix 𝑾 , 𝑒𝑙 [𝑘] and 𝑒𝑙 (𝑡) are the tracking errors of the 𝑙t h
is formulated in MIMO motion systems, constituting Contribution 2. In axis in discrete time and continuous time, and RMS(∙) is the operator
to calculate the Root Mean Square.
Section 6, the performance improvement with interaction compensa-
tion and the sampled-data characteristics is experimentally validated. It is the practical constraint that the controller can be tuned using
the data of the discrete-time tracking error 𝒆[𝑘] although the objective
In Section 7, conclusions are presented.
of the control problem is the improvement of the continuous-time
tracking error 𝒆(𝑡). To satisfy approximately the same condition of
2. Problem formulation the discrete-time and continuous-time tracking errors, the controller
should be designed not to generate the control inputs that cause the
In this section, the problem to improve continuous-time track- intersample oscillation.
ing performance in MIMO motion systems is formulated. First, the
2.2. Decoupling control for interaction compensation
reference tracking problem is defined in continuous time. Second, in-
teraction compensation is investigated for reference tracking in MIMO
In the controller design of MIMO motion systems, the static de-
motion systems. Finally, the problems in this paper are described. coupling is applied by the input decoupling matrix 𝑻 𝑢 and the output
decoupling matrix 𝑻 𝑦 . The decoupled system 𝑻 𝑦 𝑮𝑻 𝑢 should be square
2.1. Continuous-time tracking performance in sampled-data control and diagonally dominant. In many cases, the single-input single-output
(SISO) controller is designed after the static decoupling. Even if the
system is statically decoupled, the off-diagonal terms still remain and it
The considered tracking control configuration in a 𝑛𝑢 -input 𝑛𝑦 -
results in interaction between inputs and outputs [1]. Therefore, static
output continuous-time linear time-invariant system 𝑮 is shown in
decoupling is not sufficient in practice and it limits the control per-
Fig. 1, with reference 𝒓(𝑡) ∈ R𝑛𝑦 , control input 𝒖(𝑡) ∈ R𝑛𝑢 , and output
formance of MIMO motion systems. In complex mechatronic systems,
𝒚(𝑡) ∈ R𝑛𝑦 . interaction analysis is difficult and the model of coupling dynamics
The system is controlled by the sampled-data controller that consists always has a modeling error. In this paper, the feedforward controller is
of feedforward controller 𝑭 (𝜽), feedback controller 𝑲, sampler , and designed considering both diagonal and off-diagonal dynamics to com-
zero-order-hold , where sampler and zero-order-hold are defined as pensate for interaction through learning from data with less modeling
follows. effort of MIMO motion systems.

2
M. Mae et al. Mechatronics 106 (2025) 103288

2.3. Problem description

The backward differentiator does not take into account the sampled-
data characteristics with sampler and zero-order-hold, the performance
In this paper, the controller design problem is with respect to the
deteriorates when the sampling frequency is not sufficiently high.
following requirements:
3.2. Single-rate zero-order-hold differentiator for on-sample performance
Requirement 1. The sampled-data characteristics with sampler and zero-
order-hold should be considered in the discrete-time basis function design to The state-space representation of the continuous-time 𝑛t h order
improve continuous-time tracking performance. integrator in the controllable canonical form is given by
⎡ 0 1 0 0 0 ⎤
Requirement 2. The basis functions should be parameterized with [ ] ⎢ 0 ⋱ ⋱ 0 ⋮ ⎥
𝑠 𝑨𝑛𝑐 𝒃𝑛𝑐 ⎢ ⎥
physically interpretable tuning parameters considering the dynamics and 𝐻𝑛𝑐 = =⎢ 0 0 ⋱ 1 0 ⎥, (7)
𝒄 𝑛𝑐 0 ⎢ ⎥
interaction of MIMO motion systems. 0 0 0 0 1
⎢ ⎥
Requirement 1 is dealt with in Section 3 and it results in Contribution 1. ⎣ 1 0 ⋯ 0 0 ⎦
Requirement 2 is dealt with in Section 5 and it results in Contribution 2. where 𝑨𝑛𝑐 ∈ R𝑛×𝑛 , 𝒃𝑛𝑐 ∈ R𝑛×1 , and 𝒄 𝑛𝑐 ∈ R1×𝑛 .
To consider the sampled-data characteristics for on-sample tracking
3. Basis function design for sampled-data motion system performance, the single-rate zero-order-hold differentiator is defined as
follows.
In this section, the basis function design using a sampled-data
differentiator is presented. The approach improves the continuous-time Definition 6 (Single-rate Zero-order-hold Differentiator). Considering the
tracking performance of the feedforward controller in sampled-data inverse of the continuous-time 𝑛t h order integrator discretized by sam-
control. First, the conventional approach using a backward differ- pler  and zero-order-hold  with one sample shift, the 𝑛t h order
entiator is analyzed in sampled-data control. Second, the single-rate single-rate zero-order-hold differentiator 𝜉𝑠𝑟𝑛 is given by
zero-order-hold differentiator is introduced for on-sample performance 𝑛
{( ) } −1
in sampled-data control. Third, the idea of state compatibility is de- 𝜉𝑠𝑟 = 𝐻𝑛𝑐  𝑧 . (8)
fined to improve intersample performance. Finally, the multirate zero-
order-hold differentiator is developed to design the sampled-data basis Although the discrete-time signal with the continuous-time signal
functions that satisfy state compatibility. It results in Contribution 1. and 𝑛t h order single-rate zero-order-hold differentiator is compatible
on-sample with the 𝑛t h order derivative signal of the continuous-time
3.1. Challenge in sampled-data basis function design signal, the generated discrete-time signal can be oscillated or diverge.
The reason is that the single-rate zero-order-hold differentiator has
The continuous-time feedforward controller can be parameterized unstable or oscillating poles because of the inverse of discretization
using the reference signals 𝒓 and its derivatives. It results in the zeros when the degree of the continuous-time integrator is 2 or more
combination with the tuning parameters 𝜽 and the continuous-time as Euler-Frobenius polynomials [25]. Therefore, even if the single-
basis functions that consist of a continuous-time differentiator d𝑡d . For rate zero-order-hold differentiator is compatible on-sample, there are
example, the continuous-time acceleration feedforward controller 𝐹 (𝜃) mismatches in other states and it deteriorates intersample performance.
can be designed for a single-mass motion system 𝐺(𝑠) = 𝑚𝑠1 2 as
2 2
𝐹 (𝜃) = 𝜃 d . In this example, the basis function is 𝛹 = d 𝑟(𝑡) and the 3.3. State compatibility for intersample performance
d𝑡2 d𝑡2
tuning parameter is 𝜃 = 𝑚. However, for applications of mechatronic
systems in industries, the motion controllers are typically implemented The sampled-data characteristics with intersample performance can
be taken into account by the state-tracking control framework [19–21].
in discrete time. Therefore, to design the discrete-time basis function
In the 𝑛 samples lifted system, the exact state-tracking can be achieved
𝛹 , the continuous-time differentiator d𝑡d should be replaced by the
in every 𝑛 sample. In such cases, the states in every 𝑛 sample are given
sampled-data differentiator 𝜉 defined as follows.
by the multirate sampler defined as follows.
Definition 4 (Sampled-data Differentiator). The 𝑛t h order sampled-data
Definition 7 (Multirate Sampler). The multirate sampler 𝑛 in every 𝑛
differentiator 𝜉 𝑛 with sampling time 𝑇𝑠 is the conversion from the
sample with sampling time 𝑇𝑠 is defined as
continuous-time signal 𝑟(𝑡) to the discrete-time signal 𝛹𝑛 [𝑘] that is
compatible with the 𝑛t h order derivative of 𝑟(𝑡) and defined as 𝑛 ∶ 𝑟(𝑡) ↦ 𝑟[𝑖𝑛 ], 𝑟[𝑖𝑛 ] = 𝑟(𝑖𝑛 𝑛𝑇𝑠 ). (9)
𝑛
𝛹𝑛 [𝑘] = 𝜉 𝑟(𝑡). (5)
To improve both on-sample and intersample performance in
sampled-data systems with zero-order-hold and integrators, the basis
In the conventional approach [10,16], the discrete-time basis func-
functions should satisfy the state compatibility defined as follows.
tions are designed by the continuous-time reference 𝑟(𝑡) and the back-
ward differentiator defined as follows.
Definition 8 (State Compatibility). The discrete-time signal 𝛹𝑛 [𝑘], that
is compatible with the 𝑛t h order derivative signal of the continuous-
Definition 5 (Backward Differentiator). The 𝑛t h order backward differ-
𝑛 is given by
time signal 𝑟(𝑡), satisfies state compatibility if the output through the
entiator 𝜉𝑏𝑑 system, that consists of the continuous-time (𝑛 − 𝑚)t h order integrator
( −1 )𝑛 𝑛
⎧ 1−𝑧 𝐻(𝑛−𝑚)𝑐 and zero-order-hold , is equal to the continuous-time 𝑚t h
⎪ 𝑇 𝑧2  (𝑛 ∶ even)
𝑛
𝜉𝑏𝑑 = ⎨( 𝑠−1 )𝑛 −1 𝑛+1
, (6) order derivative signal of 𝑟(𝑡) in every 𝑛 sample with multirate sampler
⎪ 1−𝑧 1+𝑧
𝑧 2  (𝑛 ∶ odd) 𝑛 and defined as
⎩ 𝑇𝑠 2
d𝑚
where 𝑧 is the discrete-time shift operator with sampling time 𝑇𝑠 𝑛 𝑚 𝑟(𝑡) = 𝑛 𝐻(𝑛−𝑚)𝑐 𝛹𝑛 [𝑘], (10)
𝑛 d𝑡
defined as 𝑧𝑛 𝑟[𝑘] = 𝑟[𝑘+𝑛]. 𝑧 2 denotes the phase compensation. When 𝑛 where 𝑚 = 0, 1, … , 𝑛 − 1.
1
is odd, the phase compensation consists of the half sample shift 𝑧 2 that The sampled-data differentiator that satisfies the state compati-
is a combination of one sample shift and the first order approximation bility enables the feedforward controller parameterization with basis
of averaging the current and previous value [16]. functions to improve continuous-time tracking performance.

3
M. Mae et al. Mechatronics 106 (2025) 103288

3.4. Multirate zero-order-hold differentiator with state compatibility

To improve the intersample performance in the discrete-time sys-

tem, not only the output but also the states of the reference trajectory
are considered. The multirate zero-order-hold differentiator is designed Fig. 2. Block diagram of basis function design using multirate zero-order-hold differ-
entiator. The dotted and dashed lines denote the discrete-time signal sampled by 𝑇𝑠
by the inverse of the continuous-time integrator discretized by sampler
and 𝑛𝑇𝑠 , respectively.
and zero-order-hold to satisfy the state compatibility. In this paper,
the reference is assumed to be sufficiently smooth and satisfies the
following assumption.

Assumption 1 (Smoothness of Reference). The continuous-time refer- Satisfying the state compatibility, the relationship between the ref-
ence 𝑟(𝑡) for 𝑛 states tracking is  𝑛−1 class and differentiable at least erence and the states is given by
𝑛 − 1 times. 𝑟𝑛 [𝑖𝑛 ] = 𝒙𝑛 [𝑖𝑛 ], (21)
To satisfy the 𝑛 states compatibility in every 𝑛 sample, the lifted
where
signal is considered using the lifting operator defined as follows. [ ]𝖳
d d𝑛−1
𝑟𝑛 [𝑖𝑛 ] = 𝑛 1 ⋯ 𝑟(𝑡)
d𝑡 d𝑡𝑛−1
Definition 9 (Lifting Operator). The lifting operator 𝑛 in every 𝑛 [ ]𝖳
= 𝑟0 [𝑖𝑛 ] 𝑟1 [𝑖𝑛 ] ⋯ 𝑟𝑛−1 [𝑖𝑛 ] ∈ R𝑛 . (22)
sample is defined as
𝑛 ∶ 𝑢[𝑘] ↦ 𝑢[𝑖𝑛 ], (11)
From the discussions above, the multirate zero-order-hold differen-
where tiator is defined as follows.
[ ]𝖳
𝑢[𝑖𝑛 ] = 𝑢[𝑛𝑖𝑛 ] 𝑢[𝑛𝑖𝑛 + 1] ⋯ 𝑢[𝑛𝑖𝑛 + (𝑛 − 1)] ∈ R𝑛 . (12)
Definition 11 (Multirate Zero-order-hold Differentiator). From (18) and
(21), considering the inverse of the state equation in the continuous-
The 𝑛 samples lifted system is defined as follows. time 𝑛t h order integrator discretized by sampler and zero-order-hold
using the multirate feedforward control [19], the 𝑛t h order multirate
𝑧 𝑛 that satisfies the state compatibility
Definition 10 (Lifted System). Consider a discrete-time system 𝐻𝑑 = zero-order-hold differentiator 𝜉𝑚𝑟
𝑪 𝑑 (𝑧𝑰 − 𝑨𝑑 )−1 𝑩 𝑑 + 𝑫 𝑑 . The relation between the input and the output is given by
[ ]𝖳
in the 𝑛 samples lifted system of 𝐻𝑑 is given by 𝑛
= −1 −1 𝑛 d d𝑛−1
(23)
𝜉𝑚𝑟 𝑛 𝑩 𝑛𝑑 (𝑧 𝑰 𝑛 − 𝑨𝑛𝑑 )𝑛 1 ⋯ 𝑛−1
d𝑡
.
d𝑡
𝑦[𝑖𝑛 ] = 𝑛 𝑦[𝑘] = (𝑛 𝐻𝑑 −1
𝑛 )(𝑛 𝑢[𝑘]) = 𝐻 𝑑 𝑢[𝑖𝑛 ], (13)

where The basis function design procedure using the multirate zero-order-
[ ]𝖳 hold differentiator is shown in Fig. 2. The sampled-data character-
𝑦[𝑖𝑛 ] = 𝑦[𝑛𝑖𝑛 ] 𝑦[𝑛𝑖𝑛 + 1] ⋯ 𝑦[𝑛𝑖𝑛 + (𝑛 − 1)] ∈ R𝑛 , (14)
istics are not considered in conventional differentiator implementa-
and the lifted system 𝐻 𝑑 is defined as tions. The multirate zero-order-hold differentiator has the advantage
[ ] that it can consider the sampled-data characteristics only replacing
𝑧𝑛 𝑨𝑑 𝑩 𝑑
𝐻 𝑑 = 𝑛 𝐻𝑑 −1 = the continuous-time differentiator and it results in continuous-time
𝑛 𝑪𝑑 𝑫𝑑 tracking performance improvement.
⎡ 𝑨𝑛𝑑 𝑨𝑛−1
𝑑 𝑩𝑑 𝑨𝑛−2
𝑑 𝑩𝑑 ⋯ 𝑨𝑑 𝑩 𝑑 𝑩𝑑 ⎤
⎢ 𝑪𝑑 𝑫𝑑 𝑶 ⋯ ⋯ 𝑶 ⎥ 4. Demonstration in SISO sampled-data motion system
⎢ ⎥
⎢ 𝑪 𝑑 𝑨𝑑 𝑪 𝑑 𝑩𝑑 𝑫𝑑 ⋱ ⋮ ⎥.
= (15)
⎢ ⋮ ⋮ ⋱ ⋱ ⋱ ⋮ ⎥ In this section, the continuous-time tracking performance of the
⎢ 𝑪 𝑨𝑛−2 𝑪 𝑑 𝑨𝑛−3 𝑪 𝑑 𝑨𝑛−4 ⋱ ⎥ linearly parameterized feedforward control using the multirate zero-
⎢ 𝑑 𝑛−1 𝑑 𝑑 𝑩𝑑 𝑑 𝑩𝑑 𝑫𝑑 𝑶 ⎥
⎣ 𝑪𝑑𝑨 𝑪 𝑑 𝑨𝑛−2 𝑪 𝑑 𝑨𝑛−3 ⋯ ⎦ order-hold differentiator is demonstrated. First, the comparison is con-
𝑑 𝑑 𝑩𝑑 𝑑 𝑩𝑑 𝑪 𝑑 𝑩𝑑 𝑫𝑑
ducted in the acceleration feedforward control using the backward
Considering the states in discrete-time, the 𝑛t h order integrator differentiator, the single-rate zero-order-hold differentiator, and the
discretized by sampler and zero-order-hold is given by multirate zero-order-hold differentiator. Second, the comparison is con-
[ ] ducted between the linearly parameterized feedforward control using
𝑧 𝑨𝑛𝑑 𝒃𝑛𝑑 the multirate zero-order-hold differentiator and the multirate feedfor-
𝐻𝑛𝑑 = 𝐻𝑛𝑐  =
𝒄 𝑛𝑑 0 ward control based on exact model inversion.
[ 𝑨 𝑇 ]
𝑒 𝑛𝑐 𝑠 𝑨−1𝑛𝑐 (𝑒𝑨𝑛𝑐 𝑇𝑠 − 𝑰)𝒃
𝑛𝑐
= . (16) 4.1. Comparison in acceleration feedforward control
𝒄 𝑛𝑐 0

To design the inverse of the 𝑛t h order integrator discretized by The continuous-time tracking performance improvement by consid-
sampler and zero-order-hold, the 𝑛 samples lifted system is given by ering sampled-data characteristics can be seen clearly in a single-mass
[ ] motion system example.
𝑧𝑛 𝑨𝑛𝑑 𝑩 𝑛𝑑
𝐻 𝑛𝑑 = 𝑛 𝐻𝑛𝑑 −1
𝑛 = , (17)
𝑪 𝑛𝑑 𝑫 𝑛𝑑 4.1.1. Conditions
The controlled system is given as 𝐺(𝑠) = 𝑚𝑠1 2 where 𝑚 = 1 is the
and in state-space representation defined as
mass of the rigid body. The sampling time is set to 𝑇𝑠 = 5 ms. The
𝒙𝑛 [𝑖𝑛 + 1] = 𝑨𝑛𝑑 𝒙𝑛 [𝑖𝑛 ] + 𝑩 𝑛𝑑 𝑢[𝑖𝑛 ] (18) continuous-time reference is the 1st order polynomial trajectory. The
top of Fig. 3 shows control inputs with the acceleration feedforward
𝑦[𝑖𝑛 ] = 𝑪 𝑛𝑑 𝒙𝑛 [𝑖𝑛 ] + 𝑫 𝑛𝑑 𝑢[𝑖𝑛 ] (19) control using the backward differentiator 𝑢[𝑘] = 𝑚𝜉𝑏𝑑 2 𝑟(𝑡), that using
2 𝑟(𝑡), and that
the single-rate zero-order-hold differentiator 𝑢[𝑘] = 𝑚𝜉𝑠𝑟
where 2 𝑟(𝑡). The
using the multirate zero-order-hold differentiator 𝑢[𝑘] = 𝑚𝜉𝑚𝑟
[ ]𝖳 simulation is conducted in an open loop without a feedback controller.
𝒙𝑛 [𝑖𝑛 ] = 𝑥0 [𝑖𝑛 ] 𝑥1 [𝑖𝑛 ] ⋯ 𝑥𝑛−1 [𝑖𝑛 ] ∈ R𝑛 . (20)

4
M. Mae et al. Mechatronics 106 (2025) 103288

Fig. 4. Graphical description of state compatibility with multirate zero-order hold

differentiator in 2nd order integrator and zero-order-hold (ZOH).

Fig. 3. Simulation results of the open-loop tracking in a single-mass motion system multirate zero-order-hold differentiator is identical to the exact model
with acceleration feedforward control using the backward differentiator ( ), that inversion using the multirate feedforward control only if it consists of
using the single-rate zero-order-hold differentiator ( ), and that using the multirate only one basis function.
zero-order-hold differentiator ( ). Top: control inputs. Center: 1st order reference
Although the sampling time in industrial applications is typically
( ) and outputs. Bottom: tracking errors. The acceleration feedforward using
multirate zero-order-hold differentiator considers state compatibility and it results in shorter than 5 ms, sampled-data dynamics are affected by the relative
smaller error. (∙) and (◦) show the sampling points every 𝑇𝑠 and 2𝑇𝑠 . condition between the length of sampling time and the steepness of
reference. The preliminary result [24] also shows that the continuous-
time tracking performance improvement is experimentally validated in
4.1.2. Comparison in continuous-time tracking performance a SISO multi-modal motion system but not in MIMO motion systems.
In Fig. 3, the center and the bottom show the comparison of
outputs and tracking errors in the open-loop simulation. It shows 4.2. Comparison with multirate feedforward control
that the acceleration feedforward control using the multirate zero-
order-hold differentiator outperforms because of the state compatibility The comparison of continuous-time tracking performance between
compensating for the controlled system discretized by sampler and the linearly parameterized feedforward control using the multirate
zero-order-hold. zero-order-hold differentiator and the exact model inversion using
The limitation of the continuous-time tracking performance with the the multirate feedforward control is demonstrated in motion system
acceleration feedforward control using the backward differentiator is examples of a 2nd order mass-damper-spring model and a 4t h order
described by the sampled-data analysis that is given by
{ }{ } mass–spring-mass model.
𝑇𝑠2 (1 + 𝑧−1 ) (1 − 𝑧−1 )2
𝑦[𝑘] = {𝐺} 𝑢[𝑘] = 𝑧𝑟(𝑡)
2𝑧(1 − 𝑧−1 )2 𝑇𝑠2 4.2.1. Controller design for mass-damper-spring motion system
1 + 𝑧−1 The model 𝐺2 of the mass-damper-spring motion system shown in
= 𝑟[𝑘]. (24)
2 Fig. 5(a) is given by
The result shows that the on-sample error appears as the 1st or- 1
𝐺2 (𝑠) = , (26)
der approximated half-sample delay of the reference because of the 𝑚𝑠2 + 𝑏𝑠 + 𝑘
zero-order-hold. It means that the perfect on-sample tracking can be where 𝑚 = 4 × 10−4 , 𝑏 = 8 × 10−2 , and 𝑘 = 4.
achieved by a half-sample forward shifted reference with the 1st or- In the linearly parameterized feedforward control, the basis func-
der reference condition but cannot be achieved with higher-order tions are given by
references. [ ]
𝜳 [𝑘] = 1 𝜉𝑚𝑟 1 2 𝑟(𝑡),
𝜉𝑚𝑟 (27)
The intersample oscillation in the continuous-time tracking perfor-
mance with the acceleration feedforward control using the single-rate
and the tuning parameters are given by
zero-order-hold differentiator is described by the sampled-data analysis
[ ]𝖳
that is given by 𝜽= 𝑘 𝑏 𝑚 . (28)
2
{( ) }−1 2(𝑧 − 1)2
𝜉𝑠𝑟 = 𝐻2𝑐  𝑧 = . (25)
𝑇𝑠2 𝑧(𝑧 + 1)
The multirate feedforward control provides perfect state tracking at
The result shows that the 2nd order single-rate zero-order-hold differ- every 2 sample for the 2nd order model without modeling error.
entiator has a pole at 𝑧 = −1 and that causes the oscillation. If the
degree of the continuous-time integrator is more than 2, the single-
4.2.2. Controller design for mass–spring-mass motion system
rate zero-order-hold differentiator has unstable poles and generates the
The model 𝐺4 of the mass–spring-mass motion system shown in
unbounded signal. Although the unstable poles of the single-rate zero-
Fig. 5(b) is given by
order-hold differentiator can be compensated by the stable inversion
1
approach [18], the oscillating pole at 𝑧 = −1 cannot be compensated. 𝐺4 (𝑠) = 𝑚 𝑚 , (29)
1 2 4
The state compatibility of the multirate zero-order-hold differentia- 𝑠 + (𝑚1 + 𝑚2 )𝑠2
𝑘
tor in 2nd order integrator and zero-order-hold is illustrated in Fig. 4.
The multirate zero-order-hold differentiator stands on not only the where 𝑚1 = 𝑚2 = 2 × 10−4 and 𝑘 = 20.
reference trajectory but also its derivatives and it results in better In the linearly parameterized feedforward control, the basis func-
continuous-time tracking performance with higher-order references. tions are given by
Note that the linearly parameterized feedforward control using the [ 2 ]
4 𝑟(𝑡),
𝜳 [𝑘] = 𝜉𝑚𝑟 𝜉𝑚𝑟 (30)

5
M. Mae et al. Mechatronics 106 (2025) 103288

Table 1
Root Mean Square error RMS(𝑒(𝑡)) with multirate feedforward (MRFF) control and
linearly parameterized feedforward (LPFF) control using multirate zero-order-hold
differentiator in simulation.
RMS(𝑒(𝑡)) 𝐺2 𝐺4
MRFF with model parameters 1.81 × 10−3 4.65 × 10−4
LPFF with model parameters 1.78 × 10−1 4.14 × 10−3
LPFF with tuning min‖𝑒[𝑘]‖2 2.20 × 10−3 8.35 × 10−4
Fig. 5. Model of motion systems.

4.2.4. Comparison in continuous-time tracking performance

The tracking errors of simulations in the mass-damper-spring mo-
tion system are shown in Fig. 6(b). It shows that the linearly parame-
terized feedforward control with the model parameter makes a large
error at constant velocity regions because of the mismatch between
the model of the feedforward controller and the discretized model of
the controlled system. After tuning as min‖𝑒[𝑘]‖2 , the linearly param-
eterized feedforward control provides a smaller error than that of the
multirate feedforward control at constant velocity regions. Note that
although the multirate feedforward control guarantees the perfect state
tracking of position and velocity for every 2 sample in the 2nd order
motion system, it causes the intersample oscillation because of the
mismatch of acceleration and jerk in the continuous-time reference of
the 4t h order polynomial trajectory.
The tracking errors of simulations in the mass–spring-mass motion
system are shown in Fig. 6(c). It shows that the linearly parameterized
feedforward control with the model parameter makes a large oscillating
error because of the mismatch of the resonance frequency between the
model of the feedforward controller and the discretized model of the
controlled system. The large oscillating error is improved in the linearly
parameterized feedforward control with tuning as min‖𝑒[𝑘]‖2 .
The Root Mean Square of the continuous-time tracking errors with
the multirate feedforward control and the linearly parameterized feed-
forward control using the multirate zero-order-hold differentiator in
simulations is shown in Table 1. It shows that the tracking error of the
linearly parameterized feedforward control with tuning as min‖𝑒[𝑘]‖2 is
around the same scale as that of the multirate feedforward control even
though the solution space of the linearly parameterized feedforward
control is limited in the linear space. It means that the basis functions
using the multirate zero-order-hold differentiator provide a reason-
able linear solution space for the sampled-data feedforward control.
Fig. 6. Simulation results using multirate feedforward control with model parameters
As a result, it is shown that the linearly parameterized feedforward
( ), linearly parameterized feedforward control with model parameters ( ), control can provide around the same performance as the multirate
and linearly parameterized feedforward control with tuning min‖𝑒[𝑘]‖2 ( ). feedforward control through a tuning process using the experimental
data.

and the tuning parameters are given by 5. ILC with MIMO structured basis functions
[ 𝑚1 𝑚2 ] 𝖳
𝜽 = 𝑚1 + 𝑚2 . (31)
𝑘 In this section, the feedforward controller parameterization and
the parameter updating framework using ILC with basis functions
The multirate feedforward control provides perfect state tracking at are presented. The structured feedforward controller parameterization
every 4 sample for the 4t h order model without modeling error. for MIMO motion systems is formulated with physically interpretable
tuning parameters. Parameter update through learning is described
4.2.3. Conditions with the monotonic convergence condition in MIMO motion systems.
The continuous-time reference 𝑟(𝑡) is the 4t h order polynomial tra- It results in Contribution 2.
jectory shown in Fig. 6(a). The sampling time of the discrete-time
controller is 𝑇𝑠 = 5 ms. The continuous-time output 𝑦(𝑡) is obtained 5.1. MIMO fixed-structure feedforward controller parameterization
by higher sampling frequency in every 0.5 ms only for evaluation of
the continuous-time tracking error 𝑒(𝑡). The continuous-time tracking The dynamics of mechatronic systems are typically dominated by
error 𝑒(𝑡) is compared to the multirate feedforward control with model the mechanics assuming that electronics are much faster than me-
parameters, the linearly parameterized feedforward control with model chanics. This results in a situation where rigid body modes dominate
parameters, and the linearly parameterized feedforward control with the lower frequency and there are several flexible modes at a higher
tuning as min‖𝑒[𝑘]‖2 . The simulation is conducted in an open loop frequency due to limited mechanical stiffness. The 𝑛𝑢 -input 𝑛𝑦 -output
without a feedback controller. continuous-time multi-modal motion system [26] is defined as

6
M. Mae et al. Mechatronics 106 (2025) 103288

𝑮𝑐 (𝑠) = 𝑮𝑟 (𝑠) + 𝑮𝑓 (𝑠)

𝑛𝑟 𝑛
∑ 𝒄 𝑘𝑟 𝒃𝑘𝑟 ∑
𝑓 𝒄 𝑘𝑓 𝒃𝑘𝑓
= + , (32)
𝑘𝑟 =1 (𝑠2 + 2𝜁𝑘𝑟 𝜔𝑘𝑟 𝑠 + 𝜔2𝑘 ) 𝑘𝑓 =1 (𝑠2 + 2𝜁𝑘𝑓 𝜔𝑘𝑓 𝑠 + 𝜔2𝑘 )
𝑟 𝑓
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
rigid body modes flexible modes

where 𝑛𝑟 ∈ N+ and 𝑛𝑓 ∈ N+ are the number of rigid body and flexible

modes, 𝜔 ∈ R+ and 𝜁 ∈ R+ are the resonance angular frequency and the
damping coefficient. The vectors 𝒃 ∈ R1×𝑛𝑢 and 𝒄 ∈ R𝑛𝑦 ×1 are associated
with the inputs, the outputs, and the mode shapes. In this paper, the
system is assumed to be square as 𝑛𝑢 = 𝑛𝑦 after the static decoupling
based on the coordinate transformation.
To compensate for not only the rigid body modes but also the
Fig. 7. Block diagram of iterative learning control with basis functions.
flexible modes in MIMO motion systems, the traditional rigid body
feedforward control is extended with the additional snap feedforward
control [14] and the ideal feedforward controller 𝑭 ∗ (𝑠) is defined as
5.2. Norm-optimal ILC with basis functions
𝑭 ∗ (𝑠) = 𝑮−1 ∗ 4
𝑟 (𝑠) + 𝑫 (𝑠)𝑠 , (33)
where 𝑫 ∗ (𝑠) is the coefficient of the snap feedforward control aiming Achieving higher performance and ease of tuning for the MIMO
to compensate for the compliance of the flexible modes. feedforward controller, ILC with basis functions is implemented. ILC
The objective of the feedforward controller 𝑭 ∗ (𝑠) is to minimize the with basis functions has an advantage in task flexibility compared to
closed-loop error given by traditional ILC. The controller structure is shown in Fig. 7. To update
𝒆(𝑠) = 𝑺(𝑠)𝒓(𝑠) − 𝑺(𝑠)𝑮𝑐 (𝑠)𝑭 ∗ (𝑠)𝒓(𝑠), (34) the parameters of the feedforward controller through learning, the
optimization criterion from the present study [10] is defined as follows.
where 𝑺(𝑠) denotes the sensitivity function matrix that is defined as
𝑺(𝑠) = (𝑰 + 𝑮𝑐 (𝑠)𝑲 𝑐 (𝑠))−1 . It results in 𝑭 ∗ (𝑠) = 𝑮−1 ∗
𝑐 (𝑠) and 𝑫 (𝑠) is Definition 12 (Norm-optimal MIMO ILC with Basis Functions). The op-
given by timization criterion for norm-optimal ILC with basis functions is given
1
𝑫 ∗ (𝑠) = (𝑮−1 (𝑠) − 𝑮−1
𝑟 (𝑠)). (35) by
𝑠4 𝑐
Assuming the reference trajectory in the mechatronic systems  (𝜽𝑗+1 ) = ‖𝒆𝑗+1 ‖𝑾 𝑒 + ‖𝒖𝑓 𝑓𝑗+1 ‖𝑾 𝑓 𝑓 + ‖𝒖𝑓 𝑓𝑗+1 − 𝒖𝑓 𝑓𝑗 ‖𝑾 𝛥𝑓 𝑓 , (39)
mainly contains the low-frequency components and the resonance
frequencies of the rigid body modes are enough smaller than that where the weighting matrices are 𝑾 𝑒 ≻ 0, 𝑾 𝑓 𝑓 , 𝑾 𝛥𝑓 𝑓 ⪰ 0, the param-
of flexible modes approximated to 𝜔𝑘𝑟 ≃ 0, the compliance that eters of the feedforward controller are 𝜽𝑗 ∈ R𝑛𝜽 , and the feedforward
corresponds to the low-frequency behavior of the flexible modes is control input in next iteration is 𝒖𝑓 𝑓𝑗+1 = 𝑭 (𝜽𝑗+1 )𝒓.
given by Here, the weighting matrices 𝑾 𝑒 , 𝑾 𝑓 𝑓 , 𝑾 𝛥𝑓 𝑓 correspond to op-
{ }
1 −1 timal performance, robustness for model uncertainty, and robustness
𝑫 = lim 𝑫 ∗ (𝑠) = lim (𝑮𝑐 (𝑠) − 𝑮−1 𝑟 (𝑠))
𝑠→0 𝑠→0 𝑠4 for trial varying disturbances including noise, respectively. For the
( 𝑛 )−1 𝑛𝑓 ( 𝑛 )−1
∑ 𝑟 ⎛ ∑ 𝒄 𝑘 𝒃𝑘 ⎞ ∑ 𝑟 practical tuning procedure of the weighting matrices in the experiment,
=− 𝒄 𝑘𝑟 𝒃 𝑘𝑟 ⎜ 𝑓 𝑓
⎟ 𝒄 𝑘𝑟 𝒃𝑘𝑟 . (36) see [10].
⎜𝑘 =1 𝜔2 ⎟ 𝑘 =1
𝑘𝑟 =1 ⎝ 𝑓 𝑘𝑓 ⎠ 𝑟 The error in trial 𝑗 + 1 is given by
Hence, the fixed-structure feedforward controller for MIMO motion 𝒆𝑗+1 = 𝑺 𝒓 − 𝑺 𝑮𝒖𝑓 𝑓𝑗+1 (40)
systems is parameterized as
= 𝒆𝑗 − 𝑺 𝑮(𝒖𝑓 𝑓𝑗+1 − 𝒖𝑓 𝑓𝑗 ), (41)
d d2 d4
𝑭 (𝜽) = 𝜣 𝑝 + 𝜣 𝑣 + 𝜣 𝑎 + 𝜣𝑠 , (37)
d𝑡 d𝑡2 d𝑡4 where 𝑺 = (𝑰 + 𝑮𝑲)−1 .
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏟⏟
rigid body compliance The feedforward controller parameter update is given by
compensation compensation
𝜽∗𝑗+1 = ar g min  (𝜽𝑗+1 ). (42)
where 𝜣 𝑝 , 𝜣 𝑣 , 𝜣 𝑎 , 𝜣 𝑠 ∈ R𝑛𝑢 ×𝑛𝑦 are the parameter matrices of the feed- 𝜽𝑗+1
forward controller corresponding to the position, velocity, acceleration
and snap basis functions. Finally, the continuous-time differentiator d𝑡d When the feedforward control input is linearly parameterized in
is replaced by the sampled-data differentiator 𝜉, and the fixed-structure parameters 𝜽𝑗+1 and basis functions 𝜳 , and is defined as
sampled-data feedforward controller for MIMO motion systems is given
by 𝒖𝑓 𝑓𝑗+1 = 𝑭 (𝜽𝑗+1 )𝒓 = 𝜳 𝜽𝑗+1 , (43)

𝑭 (𝜽) = 𝜣 𝑝 + 𝜣 𝑣 𝜉 + 𝜣 𝑎 𝜉 2 + 𝜣 𝑠 𝜉 4 . (38) the optimization criterion (39) is quadratic in 𝜽𝑗+1 from (41), and an
analytic solution to (42) exists [10]. By solving the necessary condition
It enables low-complexity parameterization with physical interpretabil- 𝜕 (𝜽 )
for optimality 𝜕𝜽 𝑗+1 = 0 with basis functions 𝜳 = 𝜕𝜽𝜕 𝑭 (𝜽𝒋 )𝒓 ∈ R𝑛𝑢 ×𝑛𝜽
ity, flexibility for varying references, and consideration of sampled-data 𝑗+1 𝑗
and weighting matrices 𝑾 𝑒 , 𝑾 𝑓 𝑓 , 𝑾 𝛥𝑓 𝑓 , the analytic solution of (42)
dynamics at the same time. Although using jerk feedforward control as
for the parameter update law that minimizes  (𝜽𝑗+1 ) is given by
a basis function can improve control performance, in this paper from
the viewpoint of interpretability, jerk feedforward control has a less 𝜽𝑗+1 = 𝑸𝜽𝑗 + 𝑳𝒆𝑗 , (44)
physical meaning, and only snap feedforward control for compliance
compensation is included in basis functions in addition to rigid body where the learning filters 𝑸 and 𝑳 are given by
feedforward control.
𝑸 = (𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝑓 𝑓 + 𝑾 𝛥𝑓 )𝜳 )−1 𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝛥𝑓 𝑓 )𝜳 ,
In this paper, the developed approach combines the multirate zero-
order-hold differentiator in (23) and the fixed-structure sampled-data (45)
feedforward controller for MIMO motion systems in (38). The devel- 𝑳 = (𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝑓 𝑓 + 𝑾 𝛥𝑓 𝑓 )𝜳 )−1 𝜳 𝖳 (𝑺 𝑮)𝖳 𝑾 𝑒 . (46)
oped approach considers both continuous-time tracking performance
in Requirement 1 and interaction compensation in Requirement 2.

7
M. Mae et al. Mechatronics 106 (2025) 103288

From (40), (43), and (44), the parameter update law is written as
𝜽𝑗+1 = (𝑸 − 𝑳𝑺 𝑮𝜳 )𝜽𝑗 + 𝑳𝑺 𝒓. (47)

The parameter update law (47) leads to the monotonic convergence

condition of the parameters 𝜽𝑗 if the provided weighting matrices 𝑾 𝑒 ,
𝑾 𝑓 𝑓 , 𝑾 𝛥𝑓 𝑓 are selected properly to satisfy
𝜎(𝑸 − 𝑳𝑺 𝑮𝜳 ) < 1 ⇔
𝜎((𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝑓 𝑓 + 𝑾 𝛥𝑓 𝑓 )𝜳 )−1 𝜳 𝖳 𝑾 𝛥𝑓 𝑓 𝜳 ) < 1, (48)

where 𝜎(⋅) is the maximum singular value of the matrix.

The monotonic convergence of the parameters 𝜽𝑗 results in the
convergence of the feedforward control input 𝒖𝑓 𝑓𝑗 from (43) when the
basis functions 𝜳 from the reference 𝒓 are fixed through iterations.
From (40), (43), and (44), the feedforward control input update law
is written as
𝒖𝑓 𝑓𝑗+1 = (𝑸′ − 𝜳 𝑳𝑺 𝑮)𝒖𝑓 𝑓𝑗 + 𝜳 𝑳𝑺 𝒓, (49)

where the learning filter 𝑸′ is given by

𝑸′ = 𝜳 (𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝑓 𝑓 + 𝑾 𝛥𝑓 )𝜳 )−1 𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝛥𝑓 𝑓 ).

(50)

The feedforward control input update law (49) leads to the monotonic
convergence condition of the feedforward control input 𝒖𝑓 𝑓𝑗 if the
provided weighting matrices 𝑾 𝑒 , 𝑾 𝑓 𝑓 , 𝑾 𝛥𝑓 𝑓 are selected properly to
satisfy
𝜎(𝑸′ − 𝜳 𝑳𝑺 𝑮) < 1 ⇔
𝜎(𝜳 (𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝑓 𝑓 + 𝑾 𝛥𝑓 𝑓 )𝜳 )−1 𝜳 𝖳 𝑾 𝛥𝑓 𝑓 ) < 1. (51)

From (48) and (51), the monotonic convergence conditions of both 𝜽𝑗

and 𝒖𝑓 𝑓𝑗 are guaranteed if
𝜳 𝖳 ((𝑺 𝑮)𝖳 𝑾 𝑒 (𝑺 𝑮) + 𝑾 𝑓 𝑓 + 𝑾 𝛥𝑓 𝑓 )𝜳 ≻ 0. (52)

Note that these monotonic convergence conditions of the parameters

𝜽𝑗 and the feedforward control input 𝒖𝑓 𝑓𝑗 are derived from that of
the norm-optimal ILC [27,28] that also can be applied to MIMO sys-
tems [29–31]. Specifically, the monotonic convergence condition of the
norm-optimal ILC with basis functions is derived in Section 3.1 and
Section 4.1 of [32].
In [10], the basis functions are designed using the backward dif- Fig. 8. Experimental flexible beam setup of MIMO motion system.

ferentiator that does not take into account the sampled-data character-
istics with sampler and zero-order-hold. The theoretical performance
limitation is linked to how much the basis functions contain the dy- 6.1. Motion system
namics of the controlled system, and the lack of the sampled-data
characteristics deteriorates the tracking performance. In this paper, the The experimental flexible beam setup of a MIMO motion system is
developed approach considers the sampled-data characteristics by using shown in Fig. 8. The setup exhibits dominant flexible behavior and cou-
the multirate zero-order-hold differentiator in (23). [10] also does not pling dynamics that are expected to arise in high-precision mechatronic
contain the guidelines for extending to MIMO motion systems. This systems in industries. Although typical high-precision mechatronic sys-
paper introduces the guidelines with the fixed-structure sampled-data tems operate in six degrees of freedom, the four degrees of freedom
feedforward controller for MIMO motion systems in (38). The perfor- are elastically suspended by the leaf spring to facilitate the control
mance improvement of the developed approach with the combination design and analysis as shown in Fig. 8(a). The real-time controller
of (23) and (38) is validated in Section 6. based on Raspberry Pi with EtherCAT connection is used with the com-
putation frequency 1024 Hz. After the static decoupling of the system
with dual-inputs (𝑢1 , 𝑢2 ) and dual-outputs (𝑦1 , 𝑦2 ) based on coordinate
transformation as shown in Fig. 8(b), the controlled system 𝑮 is given
6. Validation in MIMO sampled-data motion system
in translation and rotation with dual-inputs (𝐹𝑦 , 𝑇𝑧 ) and dual-outputs
(𝑦, 𝑅𝑧 ) as shown in Fig. 8(c). The frequency response data obtained by
In this section, the developed approach combining Section 3 and multisine excitation, the continuous-time model 𝑮𝑐 with the higher-
Section 5 is applied to a MIMO motion system. The results demon- order dynamics for the simulation, and the discrete-time model 𝑮𝑑
strate the performance improvement with interaction compensation with the only diagonal rigid body dynamics for parameter update and
and sampled-data characteristics in both the simulation and the exper- feedback controller design are shown in Fig. 9.
iment.

8
M. Mae et al. Mechatronics 106 (2025) 103288

Fig. 9. Bode magnitude plot of the experimental setup: frequency response data (
), continuous-time model 𝑮𝑐 ( ) with the higher-order dynamics for the simulation,
and discrete-time model 𝑮𝑑 ( ) with the only diagonal rigid body dynamics for
parameter update and feedback controller design. Nyquist frequency of the controller
is shown in a black dotted line ( ).

6.2. Conditions

The continuous-time reference of the translation 𝑦 is the 4t h order

polynomial trajectory as shown in Fig. 10, and that of the rotation
𝑅𝑧 is set to 0 r ad for all time. The sampling frequency of the discrete-
time controller is 𝐹𝑠 = 128 Hz, as Nyquist frequency is enough higher
than the first resonance mode, and the sampling time is 𝑇𝑠 = 1∕𝐹𝑠 .
Fig. 10. Reference of 𝑦: continuous-time 4t h order polynomial trajectory and its
Although the sampling frequency in industrial applications is typically
derivatives. Reference of 𝑅𝑧 is set to 0 r ad for all time.
higher than 128 Hz, sampled-data dynamics are affected by the relative
condition between the length of sampling time and the steepness
of reference. The continuous-time outputs 𝑦 and 𝑅𝑧 are also mea- conventional approach is parameterized as
[ ]
sured in higher sampling frequency 1024 Hz only for evaluation of the 𝜳𝑦 𝟎 [ ]𝖳
𝒖𝑓 𝑓 = 𝑭 (𝜽)𝒓 = 𝜳 𝜽 = 𝜽11 𝜽22 . (57)
continuous-time tracking errors 𝒆(𝑡). To investigate the intersample 𝟎 𝜳 𝑅𝑧
performance by using the real-time controller with the computation
frequency 1024 Hz, the control frequency is set to 128 Hz, and the In the developed approach, not only the diagonal terms but also
continuous-time output is measured in 8 times higher sampling fre- the off-diagonal terms of the feedforward controller are taken into
quency 1024 Hz that is equal to the computation frequency of the account. The off-diagonal terms also can be obtained for interaction
real-time controller. The feedback controller 𝑲 is designed diagonally compensation through learning even if only a diagonal model is used
with a PD controller and a notch filter as a 5 Hz closed-loop band- for parameter update. The feedforward control input in the developed
width and a 6 dB modulus margin for compensation to disturbance and approach is parameterized as
unmodeled dynamics.
𝒖𝑓 𝑓 = 𝑭 (𝜽)𝒓 = 𝜳 𝜽
[ ]
6.3. Feedforward controller design 𝜳 𝑦 𝜳 𝑅𝑧 𝟎 𝟎 [ ]𝖳
= 𝜽11 𝜽12 𝜽21 𝜽22 . (58)
𝟎 𝟎 𝜳𝑦 𝜳 𝑅𝑧
From Section 5, the fixed-structure feedforward controller for a
MIMO motion system is parameterized as In both the simulation and the experiment, the weighting matrices
[ ][ ] [ ][ ] are set to 𝑾 𝑒 = 𝑰 and 𝑾 𝑓 𝑓 = 𝑾 𝛥𝑓 𝑓 = 𝑶, and all approaches in this
𝜃 𝜃𝑝12 1 𝜃 𝜃𝑣12 𝜉
𝑭 (𝜽) = 𝑝11 + 𝑣11 validation satisfy the monotonic convergence condition in (48). The
𝜃𝑝21 𝜃𝑝22 1 𝜃𝑣21 𝜃𝑣22 𝜉
[ ] [ 2] [ ][ ] basis functions corresponding to the reference signal and its derivatives
𝜃 𝜃𝑎12 𝜉 𝜃 𝜃𝑠12 𝜉 4 are orthogonal and have an analytical solution in the norm-optimal ILC.
+ 𝑎11 2 + 𝑠11 , (53)
𝜃𝑎21 𝜃𝑎22 𝜉 𝜃𝑠21 𝜃𝑠22 𝜉 4
where 𝜉 is a sampled-data differentiator. The basis functions of each 6.4. Validation with interaction compensation
output are defined as
[ ] To validate the performance improvement with interaction compen-
𝜳 𝑦 = 𝑟𝑦 𝜉 𝑟𝑦 𝜉 2 𝑟𝑦 𝜉 4 𝑟𝑦 , (54)
[ ] sation, the simulations and the experiments without and with interac-
2
𝜳 𝑅𝑧 = 𝑟𝑅𝑧 𝜉 𝑟𝑅𝑧 𝜉 𝑟𝑅𝑧 𝜉 𝑟𝑅𝑧 ,4 (55) tion compensation in (57) and (58) are conducted through 20 iterations.
In both approaches, the multirate zero-order-hold differentiator 𝜉𝑚𝑟 in
and the tuning parameter vectors are defined as
[ ] [ ] (23) is used as a sampled-data differentiator 𝜉 to consider sampled-data
𝜽11 = 𝜃𝑝11 𝜃𝑣11 𝜃𝑎11 𝜃𝑠11 , 𝜽12 = 𝜃𝑝12 𝜃𝑣12 𝜃𝑎12 𝜃𝑠12 , characteristics. The continuous-time tracking errors, Root Mean Square
[ ] [ ] of tracking errors, and tuning parameters learned through iterations
𝜽21 = 𝜃𝑝21 𝜃𝑣21 𝜃𝑎21 𝜃𝑠21 , 𝜽22 = 𝜃𝑝22 𝜃𝑣22 𝜃𝑎22 𝜃𝑠22 . (56)
are shown in Fig. 11, Fig. 12, and Fig. 13 for the simulation, and
Fig. 14, Fig. 15, and Fig. 16 for the experiment. The results show that
In the conventional approach, only the diagonal terms of the feed- the errors are roughly converged after 10t h iteration. The translation
forward controller are considered. The feedforward control input in the error 𝑒𝑦 is also improved a little in the simulation and the experiment,

9
M. Mae et al. Mechatronics 106 (2025) 103288

Fig. 11. Tracking error 𝒆(𝑡) in simulation using the multirate zero-order-hold differen- Fig. 14. Tracking error 𝒆(𝑡) in experiment using the multirate zero-order-hold differ-
tiator: without ( ) and with ( ) interaction compensation. Rotation error 𝑒𝑅𝑧 entiator: without ( ) and with ( ) interaction compensation. Rotation error
is improved about factor 100. 𝑒𝑅𝑧 is improved about factor 10.

Fig. 12. Root Mean Square (RMS) of tracking error in simulation using the multirate Fig. 15. Root Mean Square (RMS) of tracking error in experiment using the multirate
zero-order-hold differentiator: 𝒆[𝑘] ( ) 𝒆(𝑡) ( ) without and 𝒆[𝑘] ( ) 𝒆(𝑡) ( zero-order-hold differentiator: 𝒆[𝑘] ( ) 𝒆(𝑡) ( ) without and 𝒆[𝑘] ( ) 𝒆(𝑡)
) with interaction compensation. ( ) with interaction compensation.

Fig. 13. Tuning parameters learned through iterations in simulation using the multirate Fig. 16. Tuning parameters learned through iterations in experiment using the multi-
zero-order-hold differentiator. 𝜃∙11 ( ) is without interaction compensation. 𝜃∙11 ( rate zero-order-hold differentiator. 𝜃∙11 ( ) is without interaction compensation. 𝜃∙11
) and 𝜃∙21 ( ) are with interaction compensation. Other tuning parameters are ( ) and 𝜃∙21 ( ) are with interaction compensation. Other tuning parameters
𝜽12 = 𝜽22 = 𝑶 because the reference of the rotation 𝑅𝑧 is set to 0 r ad for all time. are 𝜽12 = 𝜽22 = 𝑶 because the reference of the rotation 𝑅𝑧 is set to 0 r ad for all time.

10
M. Mae et al. Mechatronics 106 (2025) 103288

Fig. 17. Tracking error 𝒆(𝑡) in simulation with interaction compensation: using Fig. 20. Tracking error 𝒆(𝑡) in experiment with interaction compensation: using
the backward differentiator ( ) and the multirate zero-order-hold differentiator the backward differentiator ( ) and the multirate zero-order-hold differentiator
( ). ( ).

Fig. 18. Root Mean Square (RMS) of tracking error in simulation with interaction com- Fig. 21. Root Mean Square (RMS) of tracking error in experiment with interaction
pensation: 𝒆[𝑘] ( ) 𝒆(𝑡) ( ) using the backward differentiator and 𝒆[𝑘] ( ) 𝒆(𝑡) compensation: 𝒆[𝑘] ( ) 𝒆(𝑡) ( ) using the backward differentiator and 𝒆[𝑘] (
( ) using the multirate zero-order-hold differentiator. ) 𝒆(𝑡) ( ) using the multirate zero-order-hold differentiator.

Fig. 19. Tuning parameters learned through iterations in simulation with interaction Fig. 22. Tuning parameters learned through iterations in experiment with interaction
compensation. 𝜃∙11 ( ) and 𝜃∙21 ( ) are using the multirate zero-order-hold compensation. 𝜃∙11 ( ) and 𝜃∙21 ( ) are using the multirate zero-order-hold
differentiator. 𝜃∙11 ( ) and 𝜃∙21 ( ) are using the multirate zero-order-hold differentiator. 𝜃∙11 ( ) and 𝜃∙21 ( ) are using the multirate zero-order-hold
differentiator. Other tuning parameters are 𝜽12 = 𝜽22 = 𝑶 because the reference of the differentiator. Other tuning parameters are 𝜽12 = 𝜽22 = 𝑶 because the reference of the
rotation 𝑅𝑧 is set to 0 r ad for all time. rotation 𝑅𝑧 is set to 0 r ad for all time.

11
M. Mae et al. Mechatronics 106 (2025) 103288

but the interaction effect is not serious in the translation 𝑦 because the Classens: Writing – review & editing, Validation, Software, Formal
reference of the rotation 𝑅𝑧 is set to 0 r ad for all time. The frequency analysis, Data curation. Wataru Ohnishi: Writing – review & editing,
of the residual error 𝑒𝑦 with the feedback controller corresponds to Supervision, Software, Formal analysis. Tom Oomen: Writing – review
the sensitivity peak around 5 Hz of the closed-loop bandwidth. The & editing, Supervision, Resources, Funding acquisition, Conceptual-
rotation error 𝑒𝑅𝑧 is improved significantly with interaction compen- ization. Hiroshi Fujimoto: Writing – review & editing, Supervision,
sation of about factor 100 in the simulation and of about factor 10 in Resources, Project administration, Funding acquisition, Conceptualiza-
the experiment. Note that the scales of the errors in simulation and tion.
experiment are different of about factor 10 in translation 𝑦 and of
about factor 4 in rotation 𝑅𝑧 because of the dynamics not included in Declaration of competing interest
the simulation model, measurement noise, quantization of the sensors
and actuators, and communication delay. The controller using the The authors declare that they have no known competing finan-
multirate zero-order-hold differentiator does not generate the control cial interests or personal relationships that could have appeared to
inputs that cause the intersample oscillation, and Fig. 12 and Fig. 15 influence the work reported in this paper.
show that Root Mean Square of the discrete-time and continuous-time
tracking errors are approximately the same that is the definition of the Acknowledgments
continuous-time tracking performance improvement. The validation re-
sults demonstrate that effective interaction compensation can improve This research was partly supported by JSPS KAKENHI Grant Num-
the tracking performance in multivariable motion systems. ber 21J13196 and 23K19116.

6.5. Validation with sampled-data characteristics Data availability

To validate the performance improvement with sampled-data char- Data will be made available on request.
acteristics, the simulations and the experiments with interaction com-
pensation using the backward differentiator in (6) and the multirate
zero-order-hold differentiator in (23) are conducted through 20 it- References
erations. The continuous-time tracking errors, Root Mean Square of
[1] Heertjes M, Hennekens D, Steinbuch M. MIMO feed-forward design in wafer
tracking errors, and tuning parameters learned through iterations are scanners using a gradient approximation-based algorithm. Control Eng Pract
shown in Fig. 17, Fig. 18, and Fig. 19 for the simulation, and Fig. 20, 2010;18(5):495–506.
Fig. 21, and Fig. 22 for the experiment. The translation error 𝑒𝑦 is [2] Steinbuch M, Oomen T, Vermeulen H. Motion control, mechatronics design, and
also improved significantly of about factor 5 in the simulation but is Moore’s law. IEEJ J Ind Appl 2021;2(4):21006010.
[3] Li M, Chen T, Cheng R, Yang K, Zhu Y, Mao C. Dual-loop iterative learning
improved a little in the experiment because of the unmodeled dynamics
control with application to an ultraprecision wafer stage. IEEE Trans Ind Electron
in the simulation such as communication delay. The rotation error 𝑒𝑅𝑧 2022;69(11):11590–9.
is improved a little in both the simulation and the experiment. Note that [4] Poot M, Portegies J, Mooren N, van Haren M, van Meer M, Oomen T. Gaussian
the scales of the errors in simulation and experiment are different about processes for advanced motion control. IEEJ J Ind Appl 2022;11(3):396–407.
factor 10 in translation 𝑦 and about factor 20 in rotation 𝑅𝑧 because [5] Nikooienejad N, Maroufi M, Moheimani SOR. Iterative learning control
for video-rate atomic force microscopy. IEEE/ASME Trans Mechatronics
of the dynamics not included in the simulation model, measurement 2021;26(4):2127–38.
noise, and quantization of the sensors and actuators. The validation [6] Hayashi T, Fujimoto H, Isaoka Y, Terada Y. Projection-based iterative learn-
results demonstrate that considering sampled-data characteristics has ing control for ball-screw-driven stage with consideration of rolling friction
the potential to push the envelope of the tracking performance in compensation. IEEJ J Ind Appl 2020;9(2):132–9.
[7] Lee Y-H, Hsu S-C, Chi T-Y, Du Y-Y, Hu J-S, Tsao T-C. Industrial robot accurate
sampled-data motion systems.
trajectory generation by nested loop iterative learning control. Mechatronics: Sci
Intell Mach 2021;74:102487.
7. Conclusion [8] Chen C-W, Rai S, Tsao T-C. Iterative learning of dynamic inverse fil-
ters for feedforward tracking control. IEEE/ASME Trans Mechatronics
Fixed-structure feedforward control considering sampled-data char- 2020;25(1):349–59.
[9] Miyoshi S, Ohnishi W, Koseki T, Sato M. Output voltage precise tracking control
acteristics and interactions in MIMO motion systems is developed.
for boost converters based on noncausal and nonlinear feedforward control. IEEJ
The feedforward controller that is parameterized by MIMO sampled- J Ind Appl 2023;12(6):1114–26.
data basis functions and physically interpretable tuning parameters [10] Bolder J, Oomen T, Koekebakker S, Steinbuch M. Using iterative learning control
are updated through learning. Application to the sampled-data MIMO with basis functions to compensate medium deformation in a wide-format inkjet
printer. Mechatronics: Sci Intell Mach 2014;24(8):944–53.
motion system demonstrates a significant improvement in tracking
[11] Blanken L, Koekebakker S, Oomen T. Data-driven feedforward tuning using non-
performance with interaction compensation compared to the conven- causal rational basis functions: With application to an industrial flatbed printer.
tional diagonal approach in both the simulation and the experiment. In Mechatronics: Sci Intell Mach 2020;71:102424.
engineering practice, the discrete-time basis functions that correspond [12] Wang Z, Pannier CP, Barton K, Hoelzle DJ. Application of robust mono-
to the continuous-time reference are designed using the multirate zero- tonically convergent spatial iterative learning control to microscale additive
manufacturing. Mechatronics: Sci Intell Mach 2018;56:157–65.
order-hold differentiator. The feedforward control signal is generated
[13] Wache A, Aschemann H, Krause BJ, Kurth J. Iterative learning control of
by the MIMO fixed-structure feedforward controller parameterization a pneumatically actuated lung tumour mimic model for an improvement of
using the basis functions. The tuning parameters of the feedforward PET/CT-imaging. Mechatronics: Sci Intell Mach 2022;88:102915.
controller are updated through iterative learning control on batch-to- [14] Boerlage M. MIMO jerk derivative feedforward for motion systems. In: 2006
American control conference. IEEE; 2006, p. 3892–7.
batch. Ongoing research focuses on ILC with rational sampled-data
[15] van der Meulen SH, Tousain RL, Bosgra OH. Fixed structure feedforward
basis functions and basis function design with higher-order dynamics. controller design exploiting iterative trials: Application to a wafer stage and a
desktop printer. J Dyn Syst, Meas, Control 2008;130(5):051006.
CRediT authorship contribution statement [16] Lambrechts P, Boerlage M, Steinbuch M. Trajectory planning and feed-
forward design for electromechanical motion systems. Control Eng Pract
2005;13(2):145–57.
Masahiro Mae: Writing – original draft, Visualization, Valida-
[17] Chen T, Francis BA. Optimal sampled-data control systems. London: Springer;
tion, Software, Methodology, Investigation, Funding acquisition, For- 1995.
mal analysis, Data curation, Conceptualization. Max van Haren: Writ- [18] van Zundert J, Oomen T. On inversion-based approaches for feedforward and
ing – review & editing, Software, Methodology, Formal analysis. Koen ILC. Mechatronics: Sci Intell Mach 2018;50:282–91.

12
M. Mae et al. Mechatronics 106 (2025) 103288

[19] Fujimoto H, Hori Y, Kawamura A. Perfect tracking control based on multirate Wataru Ohnishi received the B.E., M.S., and Ph.D. degrees
feedforward control with generalized sampling periods. IEEE Trans Ind Electron from The University of Tokyo, Japan, in 2013, 2015, and
2001;48(3):636–44. 2018, respectively.
[20] van Zundert J, Ohnishi W, Fujimoto H, Oomen T. Improving intersample He is currently an Associate Professor with the Depart-
behavior in discrete-time system inversion: With application to LTI and LPTV ment of Electrical Engineering and Information Systems,
systems. IEEE/ASME Trans Mechatronics 2020;25(1):55–65. Graduate School of Engineering, The University of Tokyo.
[21] Ohnishi W, Strijbosch N, Oomen T. State-tracking iterative learning control in He held a visiting position at the Eindhoven University of
frequency domain design for improved intersample behavior. Internat J Robust Technology.
Nonlinear Control 2023;33(7):4009–27. His research interests include high-precision motion
[22] Mae M, Ohnishi W, Fujimoto H. MIMO multirate feedforward controller de- control and optimization.
sign with selection of input multiplicities and intersample behavior analysis. He is a Member of the Institute of Electrical and
Mechatronics: Sci Intell Mach 2020;71:102442. Electronics Engineers and a Senior Member of the Institute
[23] Mae M, Ohnishi W, Fujimoto H. Multirate feedforward control with mode of Electrical Engineers of Japan.
decomposition for intersample performance in multivariable motion systems.
Control Eng Pract 2023;141:105694.
[24] Mae M, van Haren M, Ohnishi W, Oomen T, Fujimoto H. Feedforward with Tom Oomen received M.Sc. degree (cum laude) and Ph.D.
Acceleration and Snap using Sampled-Data Differentiator for a Multi-Modal degree from the Eindhoven University of Technology, Eind-
Motion System. In: The 2nd modeling, estimation and control conference, vol. hoven, The Netherlands.
55. Elsevier Ltd; 2022, p. 253–8. He is currently a full professor with the Department
[25] Åström KJ, Hagander P, Sternby J. Zeros of sampled systems. Automatica: J IFAC of Mechanical Engineering at the Eindhoven University of
Int Fed Autom Control 1984;20(1):31–8. Technology. He is also a part-time full professor with the
[26] Gawronski W. Advanced structural dynamics and active control of structures. Delft University of Technology. He held visiting positions
Springer Science & Business Media; 2004. at KTH, Stockholm, Sweden, and at The University of
[27] Gunnarsson S, Norrlöf M. On the design of ILC algorithms using optimization. Newcastle, Australia.
Automatica: J IFAC Int Fed Autom Control 2001;37(12):2011–6. His research interests are in the field of data-driven
[28] Norrlöf M, Gunnarsson S. Time and frequency domain convergence properties in modeling, learning, and control, with applications in pre-
iterative learning control. Int J Control 2002;75(14):1114–26. cision mechatronics. He is a recipient of the 7th Grand
[29] Barton KL, Alleyne AG. A norm optimal approach to time-varying ILC with
Nagamori Award, the Corus Young Talent Graduation
application to a multi-axis robotic testbed. IEEE Trans Control Syst Technol:
Award, the IFAC 2019 TC 4.2 Mechatronics Young Re-
Publ IEEE Control Syst Soc 2011;19(1):166–80.
search Award, the 2015 IEEE Transactions on Control
[30] van Zundert J, Bolder J, Koekebakker S, Oomen T. Resource-efficient ILC for
Systems Technology Outstanding Paper Award, the 2017
LTI/LTV systems through LQ tracking and stable inversion: Enabling large
IFAC Mechatronics Best Paper Award, the 2019 IEEJ Journal
feedforward tasks on a position-dependent printer. Mechatronics: Sci Intell Mach
of Industry Applications Best Paper Award, and recipient of
2016;38:76–90.
a Veni and Vidi personal grant.
[31] Wang Y, Hsiao T. Multivariable iterative learning control design for precision
He is a Senior Member of Institute of Electrical and
control of flexible feed drives. Sensors (Basel, Switzerland) 2024;24(11):3536.
Electronics Engineers. He has been a Senior Editor of IEEE
[32] van de Wijdeven J, Bosgra OH. Using basis functions in iterative learning control:
Control Systems Letters (L-CSS) and Co-Editor-in-Chief of
analysis and design theory. Int J Control 2010;83(4):661–75.
IFAC Mechatronics, and he has served on the editorial board
of IEEE Transactions on Control Systems Technology. He has
Masahiro Mae received the B.E. degree in electrical and also been vice-chair for IFAC TC 4.2 and a member of the
electronic engineering, M.S. degree in advanced energy, Eindhoven Young Academy of Engineering.
and Ph.D. degree in electrical engineering and information
systems from The University of Tokyo, Tokyo, Japan, in
2018, 2020, and 2023, respectively. Hiroshi Fujimoto received the Ph.D. degree in electrical
He is currently an Assistant Professor with the Depart- engineering from the Department of Electrical Engineering,
ment of Electrical Engineering and Information Systems, The University of Tokyo, Japan, in 2001.
Graduate School of Engineering, The University of Tokyo. In 2001, he joined the Department of Electrical Engi-
His research interests include multivariable control and neering, Nagaoka University of Technology, Niigata, Japan,
data-driven optimization for mechatronics and energy sys- as a Research Associate. From 2002 to 2003, he was a
tems in industrial applications. Visiting Scholar with the School of Mechanical Engineering,
He is a Member of Institute of Electrical and Electronics Purdue University, West Lafayette, IN, USA. In 2004, he
Engineers, Institute of Electrical Engineers of Japan, Society joined the Department of Electrical and Computer Engineer-
of Instrument and Control Engineers, Society of Automotive ing, Yokohama National University, Yokohama, Japan, as a
Engineers of Japan, and Japan Society of Energy and Lecturer, and became an Associate Professor in 2005. He
Resources. was an Associate Professor with The University of Tokyo
from 2010 to 2020 and has been a Professor since 2021.
Max van Haren received the M.Sc. degree (cum laude) His interests include control engineering, motion con-
in Systems and Control from the Eindhoven University of trol, nanoscale servo systems, electric vehicle control, motor
Technology, The Netherlands, in 2021. drive, visual servoing, and wireless power transfer. He was a
He is currently pursuing a Ph.D. degree in the De- recipient of the Best Paper Award from the IEEE TRANSAC-
partment of Mechanical Engineering at the Eindhoven TIONS ON INDUSTRIAL ELECTRONICS in 2001 and 2013,
University of Technology. Isao Takahashi Power Electronics Award in 2010, Best
His research interests include data-driven control and Author Prize of SICE in 2010, the Nagamori Grand Award in
identification for complex mechatronic systems. 2016, First Prize Paper Award for the IEEE TRANSACTIONS
He is a Student Member of the Institute of Electrical ON POWER ELECTRONICS in 2016, and IEEJ Industry
and Electronics Engineers. Applications Society Distinguished Transaction Paper Award
in 2018 and 2023.
Koen Classens received the M.Sc. degree (cum laude) in He is a Fellow of Institute of Electrical and Electronics
Systems and Control and the M.Sc. degree (cum laude) in Engineers and a Senior Member of Institute of Electrical
Mechanical Engineering from the Eindhoven University of Engineers of Japan. He is also a Member of the Society of In-
Technology, The Netherlands, in 2019. strument and Control Engineers, Robotics Society of Japan,
He is currently pursuing a Ph.D. degree in the De- and Society of Automotive Engineers of Japan. He has been
partment of Mechanical Engineering at the Eindhoven a Senior Editor for the IEEE/ASME TRANSACTIONS ON
University of Technology. He is a recipient of the Unilever MECHATRONICS since 2022 and an Associate Editor for
Research Prize. the IEEE Industrial Electronics Magazine since 2006. He has
His research interests include control and fault diagnosis been a Chairperson of the JSAE Technology Board since
for complex high-precision mechatronic systems. 2022 and was a past Chairperson of the IEEE/IES Technical
He is a Student Member of the Institute of Electrical Committee on Motion Control from 2012 to 2013 and the
and Electronics Engineers. JSAE vehicle electrification committee from 2014 to 2020.

13