International Journal of Precision Engineering and Manufacturing (2020) 21:883–894 Online ISSN 2005-4602

https://doi.org/10.1007/s12541-019-00303-w Print ISSN 2234-7593

REGULAR PAPER

Model Predictive Control of Autonomous Delivery Robot

with Non‑minimum Phase Characteristic
Dongil Choi1 

Received: 10 October 2019 / Revised: 5 December 2019 / Accepted: 15 December 2019 / Published online: 2 January 2020
© Korean Society for Precision Engineering 2020

Abstract
This paper introduces the concept of motion planning of delivery robot in an autonomous driving mode using an inverted
pendulum model that can effectively control disturbance. The inverted pendulum model exhibits the non-minimum phase
characteristic caused by the right half-plane zero. An effective method of reducing this characteristic is examined. A motion
platform with 3-degree-of-freedom motion and a touch sensor are installed on a wheeled omnidirectional mobile platform.
A steel ball is placed on the touch sensor and controlled to be located at the center. As the autonomous delivery robot moves,
the steel ball is subjected to various disturbances and goes off the center. The influence of disturbance can be predicted by
measuring the distance the steel ball moves away from the center. In this paper, linear quadratic regulator, preview control,
and model predictive control are applied to the inverted pendulum model for motion planning, and thus the reduction of the
non-minimum phase characteristic can be comparatively analyzed via simulation. The decrease in the disturbance is experi-
mentally compared according to motion planning. Consequently, this paper proposes an effective motion planning method
for an autonomous delivery robot with non-minimum phase characteristic.

Keywords  Model predictive control · Motion planning · Autonomous robot

1 Introduction factories. There are various categories of autonomous robots

such as a wheeled mobile platform, drone, and legged robot.
Delivery robots are a type of autonomous driving platforms These robots are applied to diverse areas and are likely to
for moving along a specified path with the assistance of be continuously developed in the future. In the case of an
various environmental sensors. Autonomous delivery robots autonomous delivery robot, its function is to deliver goods
exploit advanced technologies such as deep learning and as fast and safely as possible. Wheeled mobile robots move
artificial intelligence to utilize map data and diverse sensor easily on flat surfaces both indoor and outdoor. However,
information (GPS, IMU, CAMERA, LIDAR) to avoid obsta- when the center of mass increases, they are likely to flip over
cles and determine an optimal path. During the development and can hardly operate at high velocity. For this reason, each
of a delivery robot, the latest robot control and recognition delivery robot is designed with a pyramid shape such that
technologies can be utilized and tested on a real platform. A there is a wide area in contact with the ground and a narrow
mobile platform, environmental perception sensor, obstacle top to reduce the center of mass.
recognition, and an optimal path-generation algorithm are However, such a design is not enough for high veloc-
the core technologies of delivery robots. After over 30 years ity and acceleration. The motion of a robot necessarily
of research, these core delivery robot technologies have involves acceleration and deceleration. Moreover, when a
recently been commercialized and can now be identified robot moves in a curve, a centripetal force is experienced.
in food delivery robots, parcel delivery robots, automated If the accelerating and decelerating forces are not appropri-
logistics robots and automated guided vehicles of smart ately controlled, then it is not stable. In particular, a food
delivery robot is often loaded with liquid food and needs to
* Dongil Choi be effectively controlled. Otherwise, safe delivery is impos-
[email protected] sible. Moreover, if the ground surface is not sufficiently flat,
the vibrating wheels cause the top to vibrate, which makes
1
Myongji University, 116 Myongji‑ro, Cheoin‑gu, Yongin‑Si, it difficult to perform this task even at a low velocity. In this
Gyeonggi‑do 17058, South Korea

13
Vol.:(0123456789)

884 International Journal of Precision Engineering and Manufacturing (2020) 21:883–894

paper, we aim to address the issue of the low velocity of cur- force was compensated by the motion planning methods.
rently operating delivery robots and to develop an autono- The performance of the motion planning by linear quadratic
mous delivery robot with a motion planning technology for regulator (LQR) and MPC respectively was tested to deter-
fast and stable delivery. Figure 1 shows conceptual images mine which control method was most appropriate for the
of the robot. One is a differential drive mobile platform, and autonomous delivery robots.
the other is an omnidirectional mobile platform.
The acceleration applied to a load is controlled in real-
time by calculating the tilt angle of a motion platform. Con- 2 Inverted Pendulum Model
sidering this, we designed a controller based on an inverted
pendulum model. If a torque is not applied to the inverted 2.1 B‑POR (Ball‑Plate Omni Robot)
pendulum, the zero-moment point (ZMP) is always located
at the center of rotation. Thus, when the system is con- As shown in Fig. 2, ball-plate omni robot (B-POR) consisted
trolled by maintaining ZMP at 0, it can achieve stability of two modules and was developed as a miniature test plat-
under situations with various accelerations. However, the form [15]. A mobile platform with omnidirectional wheels
inverted pendulum model with no direct torque is basically constituted the lower part (base) and a motion platform
an underactuated system and thus shows non-minimum
phase characteristic. Accordingly, undesired undershoot
occurs when the system is controlled. To prevent this situ-
ation, various studies have been conducted [1–4]. Control
using a planned motion trajectory can be effective in reduc-
ing the non-minimum phase characteristic of the inverted
pendulum model. This method typically adopts the preview
control [5–8] and model predictive control (MPC) [9–14]
as the control techniques. To verify the method, the perfor-
mance of each controller was tested using an experimental
platform that was smaller than the real autonomous delivery
robot. The experimental platform was an omnidirectional
robot equipped with a 3 degree-of-freedom (DOF) motion
platform. The motion platform had a resistive touch sen-
sor. The position of a steel ball on the touch sensor can be
measured in real-time. In the experiment, during autono-
mous movement of the platform, the position of the ball
was measured to determine how effectively the accelerating Fig. 2  Ball-plate omni robot

Fig. 1  Conceptual images of
two types of high-speed autono-
mous delivery robots

13
International Journal of Precision Engineering and Manufacturing (2020) 21:883–894 885

corresponded to the upper part, which could move in the ⎡p⎤ ⎡0 1 0 0 ⎤⎡ p ⎤ ⎡ 0 ⎤

roll, pitch and height directions. A resistive touch sensor d ⎢ ṗ ⎥ ⎢ 0 0 a1 0 ⎥⎢ ṗ ⎥ ⎢ b1 ⎥
(1)
dt ⎢⎢ 𝜃 ⎥⎥ ⎢⎢ 0 1 ⎥⎥⎢⎢ 𝜃 ⎥⎥ ⎢⎢ 0 ⎥⎥ c
= + F
was installed on the motion platform to monitor the position 0 0
of the steel ball in real-time. The position was measured at ⎣ 𝜃̇ ⎦ ⎣ 0 0 a2 0 ⎦⎣ 𝜃̇ ⎦ ⎣ b2 ⎦
250 Hz. PID control was applied to the motion platform so
that the steel ball could be kept at the center of the touch where
sensor. Owing to the characteristics of PID control, the
a1 = mp 2 lp 2 g∕M0 , a2 = (mp + mc )mp lp g∕M0
errors converged to 0 in the steady state. However, the ball
2
was temporarily off the center under a disturbance. Depend- b1 = mp lp ∕M0 , b2 = mp lp ∕M0
ing on the intensity of the disturbance, the steel ball can be M0 = mc mp lp . 2

stabilized again or released. In this regard, the B-POR is

the same as the existing ball-plate systems for which vari- The continuous-time state-space equation of the inverted
ous control methods have been investigated [16–20]. In this pendulum model can be expressed as following forms:
paper, B-POR was used not to examine the control of the
ẋ = Ao x + Bo u
ball-plate systems, but to evaluate the control stability of the (2)
stall ball while the entire system was subjected to an accel- y = Co x
erating force. Basically, the PID control is not sufficient to
where
stably control a ball under an excessive disturbance. As such,
an additional motion plan is needed. We propose a motion [ ]T
x= p ṗ 𝜃 𝜃̇ , u = Fc
planning method using an inverted pendulum model. When [ ]
the mobile platform is accelerated, the inverted pendulum y = p,
̇ Co = 0 1 0 0 .
model can be used to calculate a control angle to compensate
The inverted pendulum model is effective in expressing
for the accelerating force of the motion platform. In addition
the simplified relationship between CoM (Center of Mass)
to the PID control, an angle is applied in the feed-forward
and ZMP. Both CoM and ZMP of a mobile platform can
mode to control the motion platform.
be stably controlled using the inverted pendulum model.
Table 1 presents the physical properties of B-POR used in
2.2 Cart Inverted Pendulum Model this investigation. The physical properties were estimated
using 3D-CAD data.
As illustrated in Fig.  3, B-POR can be modeled as an
inverted pendulum system. The B-POR with omnidirec-
tional mobility was modeled. In this case, the forward planes
3 Motion Planning
and the lateral planes need to be separately modeled. Rota-
tion was not considered in this investigation. B-POR had a
3.1 LQR (Linear Quadratic Regulator)
rectangular shape like A4 paper. However, it was assumed
that the modeling characteristics were identical in both the
The inverted pendulum model was controlled via LQR. This
forward and lateral directions. The linearized equation of
is the most common control method that can adjust Q, which
motion of the inverted pendulum model can be expressed
is the weight of the state, and R, which is the weight of con-
as follows:
trol input, and thus can set the response performance of the
system [21–23]. For an autonomous delivery robot, the posi-
tion and velocity control capacities are crucial in avoiding
an obstacle and reaching the destination. The gains obtained
via LQR control are used as state feedback to control the
position and velocity of the inverted pendulum model. The
control input of the LQR is as follows:

Table 1  Physical properties of ball-plate omni robot

Symbol Unit Value Description

mc kg 2 Mass of cart
mp kg 0.5 Mass of pendulum
lp m 0.12 Length of CoM of pendulum
Fig. 3  Cart inverted pendulum model

13

886 International Journal of Precision Engineering and Manufacturing (2020) 21:883–894

u = Klqr (xref − x) (3) approximately 2 s. For this reason, there is an error of the
initial position and the actual behavior has an undesired
where reverse action. Such a reverse action corresponds to the non-
[ ]T minimum phase characteristic that is typical of systems with
xref = 0 ṗ des 0 0 . a zero point in the right half-plane [24–27]. Accordingly, the
preview control and MPC were proposed to minimize the
The cost function for the LQR gain is expressed as non-minimum phase characteristic.
follows:
∞

∫0
J= (xT Qx + uT Ru) (4) 3.2 Preview Control

where The preview control is usually applied to control walking

in a humanoid robot. After it was originally proposed by
0 = ATo S + SAo − SBo R−1 BTo S + Q Katayama [5], the preview control began to be extensively
Klqr = R−1 BTo S. applied to the humanoid robot HRP-series by Kajita [6, 7,
28] and is still widely utilized for walking pattern genera-
In this investigation, the weights of Q, R were set as tion. The preview control superposes the future control of
follows. reference input on the LQI control in the FIFO mode. Thus,
the tracking performance can be enhanced by responding
⎡ 50 0 0 0⎤ to target reference inputs in advance. The continuous-time
⎢0 10 0 0⎥ state-space equation of the inverted pendulum model, Eq. 2
Q=⎢
0 ⎥⎥
, R = 0.1 (5)
⎢0 0 1 can be converted to the discrete-time state-space equations
⎣0 0 0 1⎦ as follows.
These weights were determined so that the control error for xk+1 = Ad xk + Bd 𝛥uk
position and velocity in the steady-state could converge to (6)
yk = Cd xk
0 and the time of velocity control could be close to 2 s. Fig-
ure 4 shows the simulation result. If the incremental state vector and the incremen-
The profile of velocity ṗ followed a step input ṗ ref tal control vector are expressed as 𝛥xk = xk − xk−1
after the initial undershoot, and the setting time was and 𝛥uk = uk − uk−1 respectively, and the augmented

Fig. 4  LQR simulation

13
International Journal of Precision Engineering and Manufacturing (2020) 21:883–894 887

T ( )−1
state-space equation is expressed as x̃ k = [ ek xk ] and Gdk = R + B̃ T P̃ B̃ B̃ T X̃ k−1
the discrete-time state-space equation can be arranged as ( ) (12)
X̃ k = Ã Tc X̃ k−1 l = 2, … , NL
follows.
[ ] [ ] where à c = à − B̃ K,
̃ Gd = GI , X̃ 1 = −Ã T P̃ C̃ T  . The preview
[ ] 1 c
̃A = I Cd Ad
, B=̃ Cd Bd
, C̃ = 1 01×4 (7) control simulation was conducted for the motion planning of
0 Ad Bd the inverted pendulum model. In this simulation, the number
where the tracking error, ek is yk − yk  , and the optimal con-
ref of preview samples, NL was 150, and the control interval was
trol problem can be constructed to minimize the following 100 Hz. The weight parameters were set as follows.
cost function. ⎡ 0.01 0 0 0⎤
∑[
∞
] ̃ x = ⎢⎢ 0
̃ e = 10, Q 0.1 0 0⎥
R = 1e−4
0 ⎥⎥
Q ,
J= eTi Qe ei + 𝛥xiT Qx 𝛥xi + 𝛥uTi R𝛥ui (8) ⎢0 0 1
i=k ⎣0 0 0 1⎦
Here, the optimal control equation is given as follows: As shown in the simulation result of Fig. 5, when the pre-
view control was applied to the inverted pendulum model, the
∑
k
∑
NL
uk = −GI ei − Gx xk −
ref
Gdi yk+i (9) control input was created at 0.5 s, which is 1.5 s ahead of
i=1 i=1 the time when the step input started and tracked the reference
input. Given that the preview control tracks the reference input
where NL is the number of preview samples and the gain is
in advance, the undershoot can be reduced. However, con-
obtained as the solution of the following algebraic Riccati
sidering that the future input should be known, the preview
equation.
control is available only when the reference input is predeter-
( )−1 mined for a period of at least 1.5 s from a given time. In case a
P = Ã T P̃ Ã − Ã T P̃ B̃ R + B̃ T P̃ B̃ B̃ T P̃ Ã + Q
̃ (10)
humanoid robot’s walking pattern is created using the preview
{ }
Here, the weight is Q ̃ = diag Qe , Qx  , from which the fol- control, the parameter is determined beforehand. As such, this
lowing optimal gain is obtained. a control method is advantageous. The position tracking per-
formance was drastically improved compared to the case when
[ ] ( )−1
K̃ = GI Gx = R + B̃ T P̃ B̃ B̃ T P̃ Ã (11) the LQR control was used. Nevertheless, a slight undershoot
and overshoot were observed with respect to velocity tracking.
In addition, the optimal preview gain is calculated by the Although various weight parameters and a different number of
following recursive formulas: preview samples were applied to the simulation, similar results

Fig. 5  Preview control simula-

tion

13

888 International Journal of Precision Engineering and Manufacturing (2020) 21:883–894

were obtained. As will be seen in the next section, MPC was ⎡ CB 0 0 ⋯ 0 ⎤

used in the simulation to enhance the tracking performance. ⎢ CAB CB 0 0 ⎥
⎢ ⎥
⋯
𝛷 = ⎢ CA2 B CAB CB ⋯ 0 ⎥
3.3 MPC (Model Predictive Control) ⎢ ⋮ ⎥
⎢ CANp −1 B CANp −2 B CANp −Nc B ⎥⎦
⎣ ⋯ ⋯
MPC is a method of obtaining a control input that optimally
Np
tracks the reference input on the prediction horizon [9, 10]. ⎡ CA ⎤ �������������
This method can impose constraints on state and control inputs ⎢ CA2 ⎥ ̄ � �T
F=⎢ ⎥, Rs = 1 1 ⋯ 1 , R̄ = rw INc ×Nc .
and renew the optimal gain that satisfies the constraints each
⎢ ⋮N ⎥
time, thereby achieving a better gain for a limited control hori- ⎣ CA p ⎦
zon. Given that MPC enables the reference input information (17)
to be utilized for control in the future, a better tracking perfor- where Np is the number of prediction horizons, Nc is the
mance for the reference input can be realized [11–13]. In this number of control horizons, and rw is the control parameter
regard, the information on the predetermined reference input is for the performance of closed-loop. Given the state variable
necessary. In case MPC is used in real time, a delay command vector, Xk = [ 𝛥xk T yk ]T and the structures of matrices C and
needs to be issued to the system, which degrades the real- A, it is noteworthy that the last column of the matrix F is the
time control performance compared to liner quadratic regu- same as R̄s . Thus, Kmpc = [ Kx Ky ] can be used again. In this
lator (LQR) control. However, autonomous delivery robots case, Kx corresponds to the feedback grain vector related to
use various sensors and a reference trajectory can be prepared 𝛥xk T  , and Ky corresponds to the feedback gain related to yk .
beforehand. Accordingly, MPC is more advantageous in the Equation 15 can be expressed as follows:
control of autonomous delivery robots because it can minimize [ ]
the non-minimum phase characteristic. In this investigation, [ ] 𝛥xk
𝛥uk = Ky rk − x K K
we applied the discrete model predictive control to the inverted y
yk (18)
pendulum model. The augmented state-space equation can be = Ky (rk − yk ) − Kx 𝛥xk .
expressed as follows using the incremental state vector 𝛥xk
and the output yk. Figure 6 shows the feedback control diagram of the discrete
[ ] [ ][ ] [ ] model predictive control for the SISO system. Figure 7 pre-
𝛥xk+1 Ad 0Td 𝛥xk Bd sents the MPC simulation result. The prediction horizon Np
= + 𝛥uk
yk+1 Cd Ad 1 yk Cd Bd was 150, the control horizon Nc was 20, rw was 1, and the
[ ] (13)
[ ] 𝛥xk simulation had a control interval of 100 Hz.
yk = 0d 1
yk The simulation result shows that the control input umpc
was created 1.5 s before the beginning of the step refer-
Xk+1 = AXk + B𝛥uk ence input pref  . Thus, the non-minimum phase charac-
(14) teristic could be minimized. The velocity tracking result
yk = CXk
indicates that the excessive reverse action caused by LQR
[ ] [ ]
Ad 0Td Bd [ ] was removed. Consequently, the undershoot and the over-
where A = , B= , C = 0d 1 ,
[ ] C A
d d 1 C B
d d shoot were significantly smaller than those of the pre-
𝛥xk [ ] view control simulation. As such, MPC was effective in
Xk = , 0d = 0 0 ⋯ 0 .
yk reducing the non-minimum phase characteristic and thus
Thus, the increase in control input 𝛥uk is as follows: exhibited superior performance compared to the preview
control. In the next section, the experimental platform is
𝛥uk = Ky rk − Kmpc Xk (15) controlled via motion planning via LQR and MPC, and
its performance is tested.
[ ]T
̄ −1 𝛷T R̄s
Ky = 1 0 0 ⋯ 0 (𝛷T 𝛷 + R)
[ ]T (16)
̄ −1 𝛷T F
Kmpc = 1 0 0 ⋯ 0 (𝛷T 𝛷 + R)

Fig. 6  Discrete model predictive control diagram

13
International Journal of Precision Engineering and Manufacturing (2020) 21:883–894 889

Fig. 7  DMPC simulation

4 Experiment

4.1 Experimental Platform

The B-POR is a miniature test platform for an autono-

mous delivery robot. An experiment was performed using
B-POR to compare LQR and MPC with respect to control
performance. A steel ball was placed on the motion plat-
form and the position of the ball was measured using a
resistive touch sensor. Data from the touch sensor were
acquired via a USB port. The steel ball was controlled
towards the center of the touch sensor using PID con-
trol on an Arduino Mega. Figure 8 illustrates the control
structure of B-POR. In the experiment, while the omni-
directional mobile platform was stationary, the ball-plate
system was controlled for 20 s via the PID control. Fig-
Fig. 8  Control system structure of B-POR
ure 9 shows the experimental result. The ball was stabi-
lized within approximately 7 mm in the experiment. The
PID control used for B-POR stabilized the steel ball when 4.2 LQR: Tracking Rectangular Trajectory
the omnidirectional mobile platform was in the stationary
condition. However, when the platform moves with vari- An experiment was performed whereby a rectangular tra-
ous accelerations, control stability cannot be ensured. To jectory was tracked via LQR motion planning. The veloc-
stabilize the steel ball against movement, feed-forward ity command of 0.3 m/s was applied as the step input
control was performed by the motion plans using LQR for 5 s sequentially in the forward, right, backward and
and MPC respectively.

13

890 International Journal of Precision Engineering and Manufacturing (2020) 21:883–894

stabilize the ball than for the stationary condition. How-

ever, the feed-forward tilt angles of a motion platform can
be calculated via the LQR motion planning and utilized to
ensure more stable control. Figure 12 shows two-dimen-
sional motion trajectories for more intuitive visualization.
B-POR did not perfectly follow the reference rectangular
trajectory, and a particularly large error was observed at
the right-angled curves. This occurred because the accu-
racy of estimating a position decreases owing to the non-
minimum phase characteristic. Consequently, the course
deviated outward.

4.3 MPC: Tracking Rectangular Trajectory

Another experiment was conducted whereby a rectangular

trajectory was tracked by MPC motion planning. A veloc-
Fig. 9  PID control of ball-plate system ity of 0.5 m/s was applied as the step input, which was
larger than that of the LQR experiment. Given that MPC
had higher control stability, a higher reference velocity than
left directions. Figure 10 shows the position and velocity that of LQR was applied. Figure 13 shows the position and
graphs in the time domain. The experimental results were velocity trajectories.
in close agreement with the simulation results. The feed- Similar to the simulation result, both position and veloc-
forward tilt angles of the motion platform are illustrated ity were tracked more smoothly. As shown in Fig. 14, the
in Fig. 11. The maximum tilt angle was approximately feed-forward tilt angles were less than 3°. These values were
6° and the steel ball was controlled within approximately lower than those of LQR. The steel ball was stabilized within
45  mm. If a step-type velocity command is given, an 20 mm. This indicates that the MPC motion planning can
accelerating force occurs at the moment that the velocity provide more stable control at a higher speed than the LQR
changes. Such an accelerating force becomes a distur- motion planning. Table 2 numerically presents the RMS val-
bance to the steel ball. In this case, it is more difficult to ues for steel ball positions. When MPC motion planning was

Fig. 10  Position and velocity in

time domain using LQR

13
International Journal of Precision Engineering and Manufacturing (2020) 21:883–894 891

Fig. 11  Tilt angle and position

of the ball in time domain using
LQR

Fig. 12  2-Dimensional plots of the B-POR and the ball using LQR

applied, the steel ball was controlled with an RMS value of was experimentally demonstrated that the MPC motion plan-
approximately 4 mm, which was close to 3 mm or the control ning achieved more reliable control than the LQR motion
level in the stationary condition. As illustrated in Fig. 15, planning. The experimental video is available online [29].
when MPC was used, the rectangular trajectory was smoothly
tracked with no deviation from the corner. Consequently, it

13

892 International Journal of Precision Engineering and Manufacturing (2020) 21:883–894

Fig. 13  Position and velocity in

time domain using MPC

Fig. 14  Tilt angle and position

of the ball in time domain using
MPC

5 Conclusion delivery robot was modeled as an inverted pendulum sys-

tem and the ZMP control was performed. In addition, the
This paper examined motion planning methods for stably motion platform was dynamically stabilized to minimize
controlling autonomous delivery robots. An autonomous the impact of disturbance on the loads. For the predictive
control methods, the preview control and MPC were tested

13
International Journal of Precision Engineering and Manufacturing (2020) 21:883–894 893

Fig. 15  2-Dimensional plots of the B-POR and the ball using MPC

with respect to reducing the non-minimum phase charac- the rotational movement is not considered. For the rotation
teristic of the inverted pendulum model. The performance in place, the problem is not significant because it does not
of the MPC motion planning was demonstrated by both produce lateral acceleration, but in the case of turning, the
simulation and experiment. In the experiment, while the centrifugal acceleration may occur and become a problem.
steel ball was stabilized using the PID control, its posi- This issue requires more complex models and interpreta-
tion was measured to indirectly identify acceleration of tions. In the author’s past paper [24], I conducted a study
the steel ball during movement. The experimental result using double inverted pendulum, which enabled better object
revealed that the MPC motion planning ensured much stabilization and the rotational movement is considered.
more stable control and the accelerations were effectively However, the paper had limitations that are limited to LQR
compensated during movement. control. So, I have a plan to apply the MPC motion planning
The main reason why the MPC performs better is because to double inverted pendulum model for more sophisticated
it uses the future command information in advance. This stabilization performance. Finally, the performance of the
inevitably leads to lower real-time performance. The MPC MPC motion planning needs to be demonstrated in a real
motion planning used in this paper did not adopt any con- autonomous delivery robot in the future.
straint. The real-time performance of the MPC motion plan-
ning can be further improved by applying constraints. If con- Acknowledgements  This work was supported by 2018 Research Fund
of Myongji University.
straints are applied, a delivery robot could be controlled to
quickly avoid a sudden obstacle in real-time.
In this paper, the single inverted pendulum model is used
to represent the system with minimal degree of freedom.
A more appropriate model would be double inverted pen-
References
dulum to model the movement of the steel ball. And also, 1. Choi, D., & Oh, J. (2012). ZMP stabilization of rapid mobile
manipulator. In Proceedings of the IEEE international conference
on robotics and automation (ICRA) (pp. 883–888). IEEE.
Table 2  Root-mean-square (RMS) value of ball position 2. García, P., Albertos, P., & Hägglund, T. (2016). Control of unsta-
ble non-minimum-phase delayed systems. Journal of Process
Motion PID LQR MPC Control, 16(10), 1099–1111.
Stationary Rectangular Rectangular 3. Nakamura, R., & Amino, A. (2017). Perfect tracking control using
a phase plane for a wheeled inverted pendulum under hardware
X-axis (mm) 3.19 8.81 4.11 constraints. In Proceedings of the IEEE international conference
Y-axis (mm) 3.33 6.61 4.15 on robotics and automation (ICRA) (pp. 4377–4382). IEEE.

13

894 International Journal of Precision Engineering and Manufacturing (2020) 21:883–894

4. Dallali, H., Brown, M., & Vanderborght, B. (2009). Using the 20. Jørgensen, V. (1974). A ball-balancing system for demonstration
torso to compensate for non-minimum phase behaviour in ZMP of basic concepts in the state-space control theory. International
bipedal walking. In T. Kröger & F. M. Wahl (Eds.), Advances in Journal of Electrical Engineering Education, 11(4), 367–376.
robotics research (pp. 191–202). Berlin: Springer. 21. Choi, D., & Oh, J. (2008). Human-friendly motion control of
5. Katayama, T., Ohki, T., Inoue, T., & Kato, T. (1985). Design of an a wheeled inverted pendulum by reduced-order disturbance
optimal controller for a discrete-time system subject to preview- observer. In 2008 IEEE international conference on robotics and
able demand. International Journal of Control, 41(3), 677–699. automation (ICRA) (pp. 2521–2526). IEEE.
6. Akachi, K., Kaneko, K., Kanehira, N., Ota, S., Miyamori, G., 22. Choi, D., Oh, J. H. (2011). Four and two wheel transformable
Hirata, M., et al. (2005). Development of humanoid robot HRP- dynamic mobile platform. In 2011 IEEE international conference
3P. In 5th IEEE-RAS international conference on humanoid robots on robotics and automation (ICRA) (pp. 1–4). IEEE.
(pp. 50–55). IEEE. 23. Kim, M., & Choi, D. (2019). Design and development of a vari-
7. Kajita, S., Kanehiro, F., Kaneko, K., Fujiwara, K., Harada, K., able configuration delivery robot platform. International Journal
Yokoi, K., et al. (2003). Biped walking pattern generation by using of Precision Engineering and Manufacturing, 20(10), 1757–1765.
preview control of zero-moment point. In 2003 IEEE international 24. Choi, D., & Oh, J. (2014). Motion planning for a rapid mobile
conference on robotics and automation (ICRA) (pp. 1620–1626). manipulator using model-based ZMP stabilization. Robotica. https​
IEEE. ://doi.org/10.1017/S0263​57471​40025​01.
8. Cho, B.-K., & Kim, J.-Y. (2018). Dynamic posture stabilization of 25. Choi, D., Kim, M., & Oh, J. H. (2012). Development of a rapid
a biped robot SUBO-1 on slope-changing grounds. International mobile robot with a multi-degree-of-freedom inverted pendulum
Journal of Precision Engineering and Manufacturing, 19(7), using the model-based zero-moment point stabilization method.
1003–1009. Advanced Robotics, 26(5–6), 515–535.
9. Holkar, K. S., & Waghmare, L. M. (2010). An overview of model 26. Kim, M., Choi, D., & Oh, J. H. J. (2010). Stabilization of a
predictive control. International Journal of Control and Automa- rapid four-wheeled mobile platform using the ZMP stabiliza-
tion, 3(4), 47–63. tion method. In 2010 IEEE/ASME international conference on
10. Pannocchia, G., & Rawlings, J. B. (2003). Disturbance models advanced intelligent mechatronics (AIM) (pp. 317–322). IEEE.
for offset-free model-predictive control. AIChE Journal, 49(2), 27. Canete, L., & Takahashi, T. (2014). Development of a single
426–437. controller for the compensation of several types of disturbances
11. Wieber, P.-B. B. (2006). Trajectory free linear model predictive during task execution of a wheeled inverted pendulum assistant
control for stable walking in the presence of strong perturbations. robot. In 2014 IEEE/RSJ international conference on intelligent
In 2006 6th IEEE-RAS international conference on humanoid robots and systems (IROS) (pp. 2414–2420). IEEE.
robots (pp. 137–142). IEEE. 28. Kajita, S., Morisawa, M., Harada, K., Kaneko, K., Kanehiro, F.,
12. Giselsson, P. (2009). Model predictive control in a pendulum sys- Fujiwara, K., et al. (2006). Biped walking pattern generator allow-
tem. In 2009 American control conference (pp. 2335–2340). ing auxiliary ZMP control. In 2006 IEEE/RSJ international con-
13. Lafaye, J., Collette, C., & Wieber, P. B. (2015). Model predictive ference on intelligent robots and systems (IROS) (pp. 2993–2999).
control for tilt recovery of an omnidirectional wheeled humanoid IEEE.
robot. In 2015 IEEE international conference on robotics and 29. [Experiment] Ball Plate Omni Robot. (2019). Retrieved December
automation (ICRA) (pp. 5134–5139). IEEE. 5, 2019, from https​://youtu​.be/WEnwp​PVgp6​Q.
14. Lim, H., et al. (2014). Experimental verification of nonlinear
model predictive tracking control for six-wheeled unmanned Publisher’s Note Springer Nature remains neutral with regard to
ground vehicles. International Journal of Precision Engineering jurisdictional claims in published maps and institutional affiliations.
and Manufacturing, 15(5), 831–840.
15. Choi, D., Kim, M., Kim, H., Choe, J., & Nah, M. C. (2018). Real-
time motion planning of autonomous personal transporter using
model predictive control for minimizing non-minimum phase Dongil Choi  received BS, MS,
motion. In 2018 15th international conference on ubiquitous Ph.D degrees in Mechanical
robots (UR) (pp. 362–368). IEEE. E n g i n e e r i n g f r o m Ko r e a
16. Zeeshan, A., Nauman, N., & Jawad Khan, M. (2012). Design, Advanced Institute of Science
control and implementation of a ball on plate balancing system. and Technology (KAIST), in
In Proceedings of 2012 9th international Bhurban conference on 2005, 2007, 2012, respectively.
applied sciences and technology (IBCAST) (pp. 22–26). IEEE. He is currently a professor at
17. Awtar, S., Bernard, C., Boklund, N., Master, A., Ueda, D., & Myongji University. His research
Craig, K. (2002). Mechatronic design of a ball-on-plate balancing interests include design and con-
system. Mechatronics, 12(2), 217–228. trol of autonomous robot, legged
18. Fan, X., Zhang, N., & Teng, S. (2004). Trajectory planning and robot and mobile manipulator.
tracking of ball and plate system using hierarchical fuzzy control
scheme. Fuzzy Sets and Systems, 144(2), 297–312.
19. Fabregas, E., Chacón, J., Dormido-Canto, S., Farias, G., & Dor-
mido, S. (2015). Virtual laboratory of the ball and plate system.
IFAC-PapersOnLine, 48(29), 152–157.

13