DOI: 10.48011/asba.v2i1.

1523

Explicit Model Predictive Control for

Inverted Pendulum Systems ?
Mateus Mussi Brugnolli, Bruno Augusto Angélico ∗
∗ Department of Telecommunications and Control Engineering
Escola Politécnica da USP, São Paulo, Brazil.
(e-mails: mateus_mmb@usp.br, angelico@lac.usp.br)

Abstract: Model Predictive Control is a control technique that has been greatly investigated in
recent years. It has the versatility of different types of models for the prediction of the system and
aptitude to handle the system constraints. In the last decade, the multi-parametric optimization
has been applied to the control theory that allowed for the MPC optimization to be performed
offline, which was denominated as explicit Model Predictive Control. This work investigates the
application of this control technique in Inverted Pendulum systems, which are commonly used
as didactic control systems. The complete control design is described considering its validation
for two Inverted Pendulum systems through simulations.

Keywords: Model Predictive Control, Explicit solution, Multi-parametric optimization,

Inverted Pendulum, Control Systems.

1. INTRODUCTION Since the MPC with linear models results in a quadratic

optimization problem, researchers started to apply the
Inverted pendulum systems are vastly used as didactic multi-parametric programming theory (Bemporad et al.
control tools (Sugaya et al. (2017)). This classic control (2002)). Briefly, in this approach a feasible parameter
problem relies on mechanical actuation to induce a torque set is assembled, that is a representation of all possible
sufficient to maintain a stem on its inverted position. states that the real system may reach. With this set,
This system has several mechanical configurations and each optimization restriction creates a finite number of
adaptability on degrees of freedom. Therefore, a vast set polytopes as a function of the parameters. Each region,
of control theories can be applied to this multi-variable that is related to a subset of parameters, can be interpreted
system, from the classic control, modern control, to the as the set of active or inactive constraints. Therefore,
most advanced and recent controllers (Jmel et al. (2020)). the optimization can be solved offline and it designs
the optimal MPC feedback control for each region. This
This work highlights the application of the Model Pre- approach has been known in the literature as the explicit
dictive Control (MPC) theory on inverted pendulum sys- MPC (eMPC) since the optimization solves the optimal
tems. This technique had its origins in the petrochemical control actions explicitly for each state subset (Borrelli
industries in the ’70s and appealed to many researchers. et al. (2017)).
Nowadays, this theory has several configurations and flex-
ibilities, for example, the usage of different sets of models The implementation of the eMPC uses a look-up table. It-
as linear, hybrid, nonlinear, deterministic, or stochastic eratively, the states of the system are collected and the ac-
models (Qin and Badgwell (2003); Lenz et al. (2015)). tive region of the current states determines the respective
MPC state feedback. Notice that if there is a considerable
The key concept of the MPC is to optimize the system number of restrictions or if the system dimension is large
performance through the prediction of the system model in enough, the multi-parametric optimization may result in
a closed-loop. Generally, the control signal is represented a large number of regions. Hence, this work simulates the
as discrete-time state feedback. Accordingly, the system eMPC considering multiple system dimensions. The main
states must be measured each time sampling to update the contributions for this paper are the investigation of the
MPC optimization parameters (states). Additionally, the eMPC for two inverted control systems and its validation
MPC theory includes states and control signal constraints through simulations.
for a feasibility guarantee. Thus, the optimization solution
is heavily influenced by the system complexity. A system This work is organized as follows: In Section 2, the model-
with a high dimension may demand high optimization ing of two distinct inverted control systems are addressed,
solution time that exceeds the time sampling of the control the Reaction Wheel Pendulum and the Rotary Inverted
system (Camacho and Alba (2013)). Pendulum (known as Furuta Pendulum). The next Sec-
tion 3 follows with the specifications for the eMPC with
? This work was supported by Fundação de Amparo à Pesquisa its tuning parameters and system constraints. Section 4
do Estado de São Paulo (FAPESP), grant 2017/22130-4 and Co-
develops the simulation setup using the nonlinear models
ordenação de Aperfeiçoamento de Pessoal de Nı́vel Superior - Brazil and the discussion of the results. Lastly, Section 5 points
(CAPES) - Finance Code 001. out the conclusions and topics for future works.
 2. INVERTED PENDULUM MODELING 1
Erw = (θ̇1 + α̇)Jrw (θ̇1 + α̇)
2
This work considers two configurations that has been re- 1
searched and built in the Laboratório de Controle Aplicado + (α̇`rw cos(θ1 ))mrw (α̇`rw cos(θ1 ))
2
(LCA) (Angélico (2016)). The systems are the reaction 1
wheel inverted pendulum (Block et al. (2007)) and the + (α̇`rw sin(θ1 ))mrw (α̇`rw sin(θ1 )) (3)
2
pendulum proposed by Katsuhisa Furuta, rotary inverted 1
pendulum (Mori et al. (1976)). Ep = α̇Jp α̇ (4)
2
2.1 Reaction Wheel Pendulum Using the complete Lagrangian function L of Equation (5),
Equation (6) is the Euler-Lagrange equation to obtain the
A representation of the reaction wheel pendulum is shown nonlinear models for the reaction wheel and the pendulum.
in Figure 1. A stem that represents the pendulum rotates The generalized forces Qi for each body are presented in
freely on the axis #1. In the opposite side of the pendulum, Equations (7,8). For the reaction wheel, the control signal
a DC motor on the stem. A second body, named as reaction is the P W M signal for the 6 Volts DC motor and there
wheel, is fixed and rotates on the axis #0 of the motor axis. is viscous damping in axis #0. For the pendulum, there is
The only actuator of the system is the DC motor, which viscous damping in axis #1.
directly applies a torque into the reaction wheel. In this
case, the angular position of axes #0 and #1, α and θ1 , L = Erw + Ep − Prw − Pp (5)
and their respective velocities are considered measured. 
d ∂L

∂L
− = Qi ; for δ = α, θ1 (6)
dt ∂ δ̇ ∂δ
K
Qrw = t (6 P W M − Ke α̇) − b0 α̇ (7)
R
Qp = −b1 θ̇1 (8)
Table 1 shows the parameters of the reaction wheel pen-
dulum built in the LCA.
Table 1. Parameters of the reaction wheel.
Parameter Value
mrw Reaction Wheel Mass [Kg] 0.144
g Gravity Acceleration [m/s2 ] 9.81
d Distance of CoM of rw to axis #1 [m] 0.0987
mp Pendulum Mass [Kg] 0.149
Jrw Inertia of the Reaction Wheel [Kgm2 ] 9.456×10−4
Figure 1. Illustration of the reaction wheel pendulum. `rw Distance of rw to axis #1 [m] 0.14298
Jp Inertia of the Pendulum [Kg] 6.2533×10−4
The control system objective is the application of a torque Kt Motor Torque Constant [N m/A] 0.307
into the reaction wheel that results in a reaction torque on R Armature Resistance [Ω] 5.325
Ke Back EMF Constant [V s/rad] 0.273
the pendulum that maintains it in the inverted position
b0 Viscous damping of axis #0 [N m/(rad/s] 1.0×10−5
(θ1 = 0). The control input is the Pulse Width Modulation b1 Viscous damping of axis #1 [N m/(rad/s] 1.0×10−6
(P W M ) signal applied to the DC motor. By the definition,
the signal ranges between 1 (applying the maximum av-
erage voltage) and 0 (null voltage). Additionally, negative The nonlinear equations are linearized applying the jaco-
values imply that the motor rotates in its inverted nominal bian with the state vector xrw = δθ1 δ θ̇1 δ α̇ T , where δ
 
rotation. It is considered the Euler-Lagrange modeling represents the linearization close to a operation point. The
method, with Ei as the kinematic energy and Pi as the h iT
potential energy of body i. This system was divided into chosen point was θ1R θ̇1R α̇R = [0 0 0]T . Note that
two bodies: the reaction wheel rw and the pendulum p. the dimension of the system is three, since the reaction
The potential energy of the reaction wheel Prw is the wheel angle α is not controllable in this configuration.
displacement of the Center of Mass (CoM) of the reaction Aiming the simulations using the MPC, that is a discrete-
wheel body in the vertical plane and for the pendulum Pp time control, a ZOH discretization is applied to the state-
is the displacement of its CoM in the same plane. space, using fs = 50 Hz. The discrete linear model for the
control is shown in Equation (9).
Prw = mrw gd cos(θ1 ) (1)
xrw (k + 1) = Arw xrw (k) + Brw P W M (k) (9)
Pp = mp gd cos(θ1 ) (2)
−0.0170
   
1.0189 0.0201 0.0008
The kinematic energy of the pendulum Ep is related to Arw =  1.8777 1.0189 0.0724 ; Brw = −1.5908
the torque applied on the stem. For the reaction wheel,
Erw has three distinct components. First, the torque −1.5951 −0.0170 0.6563 7.5526
applied to the reaction wheel with both angular velocities. The objective of the control system is to regulate all states
Additionally, there are vertical and horizontal translation to zero since it is the pendulum in the inverted position.
displacements of the CoM of the reaction wheel. Also, note that the angular position of the reaction wheel
does not impact on the system stability, and the actuation solved with Equation (15). The generalized forces Qi for
on the system is limited from -1 to 1 (negative values each body are defined in Equations (16,17). For the arm,
of P W M represent opposite rotation). Because of the the control signal is a P W M for the 12 Volts DC motor
mechanical configuration, the reaction torque can not and there is viscous damping in axis #0. Additionally,
stabilize the pendulum at any other position other than there is a viscous damping in axis #2. Table 2 shows the
the inverted position and the standby position. parameters of the Furuta pendulum built in the LCA.

2.2 Furuta Pendulum L = Earm + Ef − Parm − Pf (14)

 
The Figure 2 is a representation of the Furuta Pendulum. d ∂L ∂L
− = Qi ; for δ = θ0 , θ2 (15)
There is a DC motor accoupled to a fixed structure with a dt ∂ δ̇ ∂δ
beam fixed on axis #0. This beam is referred as the body K  
arm. There is also a stem that rotates around axis #2 on Qarm = t 12 P W M − Ke θ̇0 − barm θ̇0 (16)
R
the other edge of the arm. This stem is referred as the
body Furuta Pendulum f . The objective of this system Qf = −bf θ̇2 (17)
is to apply a torque to rotate the arm and stabilize the Table 2. Parameters of the Furuta Pendulum.
pendulum in the inverted position (θ2 = 0).
Parameter Value
mf Pendulum Mass [Kg] 0.076
g Gravity Acceleration [m/s2 ] 9.81
`f Distance of CoM of pendulum to axis #2 [m] 0.125
Jarm Inertia of the arm [Kgm2 ] 0.0046
Jf Inertia of the Pendulum [Kg] 0.00039
rf Distance of pendulum to axis #0 [m] 0.25
Kt Motor Torque Constant [N m/A] 0.118
R Armature Resistance [Ω] 3.6
Ke Back EMF Constant [V s/rad] 0.118
barm Viscous damping of axis #0 [N m/(rad/s] 3.0×10−5
bf Viscous damping of axis #2 [N m/(rad/s] 5.0×10−6

Afterwards, the nonlinear equations are linearized apply-

ing the jacobian with the state vector
xf = δθ0 δθ2 δ θ̇0 δ θ̇2 T ,
 

where δ represents the linearization close to a operation

point. It was linearized for
h iT
θ0R θ2R θ̇0R θ̇2R = [0 0 0 0]T .
Note that this system has dimension four since the angular
positions of the arm, of the pendulum, and their respective
Figure 2. Illustration of the Furuta pendulum. velocities are controllable in this configuration. Applying a
ZOH discretization in the state-space, the discrete model
There are two bodies for this system: arm and pendulum for the control is shown in Equation (18). This control
f . The potential energy of the arm Parm is equal to zero system is able to track a reference signal for the arm
because of the CoM of the arm maintains in the horizontal position while stabilizing the pendulum on its inverted
plane. The potential energy of the pendulum Pf is the position.
displacement of the CoM in the vertical plane.
xf (k + 1) = Af xf (k) + Bf P W M (k) (18)
Parm = 0 (10)
1.000 −0.005 0.020 0.000 0.014
   
Pf = mf g`f cos(θ2 ) (11) 0.000 1.019 0.001 0.020  −0.021
Af = 
0.000 −0.486 0.987 −0.005 ; Bf =  1.362 
  
The kinematic energy of the arm Earm is obtained from
the torque of its rotation. The energy Ef for the pendulum 0.000 1.914 0.020 1.019 −2.051
can be described using its rotation, the velocity of the CoM
in the directions of x, y, and z (Mori et al. (1976)). 3. EXPLICIT MODEL PREDICTIVE CONTROL

1 The MPC is a discrete-time controller that uses the model

Earm = (θ̇0 )Jarm (θ̇0 ) (12) to predict the optimal control sequence. For both systems,
2
1 mf  the control input is u(k) = P W M (k). Consider the cost
Ef = (θ̇2 )Jf (θ̇2 ) + `2 2 2 2
f θ̇2 + rf θ̇0 + . . . function of Equation (19).
2 2  NX−1
+`2 θ̇
f 0
2 sin2 (θ ) + 2r ` θ̇ θ̇ cos(θ )
2 f f 0 2 2 (13) J(k) = (x(k+j))T Q(x(k+j))+(u(k+j))T R(u(k+j))
With the Lagrangian function for the Furuta Pendulum in j=0
Equation (14), the nonlinear equations of this system are (19)
The element x(k + j) is the prediction of the states min J(u(k + j), x(k)) (26)
after j samples, considering the linear model presented u(k+j)
in Equation (9) for Reaction Wheel system and Equation subject to
(18) for Furuta Pendulum. Note that the measurement g(u(k + j), x(k)) ≤ 0 (27)
of x(k) is a requirement for the model prediction. The for all x(k) ∈ X , u(k + j) ∈ U, and j = 0, ... , N − 1. The
control sequence u(k + j) is the optimization vector with restrictions of Equation (27) form a polytope with a finite
the structure u(k + j) = F (j)x(k + j) + gj (j). The element number of partitions (regions). For each region, there is an
gj (j) is related to the input saturation. If F (j)x(k + j) optimal u∗ (k + j) vector that minimizes J(u(k + j), x(k)).
would saturate the control signal, gj (j) for the prediction j
Since Lemma 2 considers the sets X and U, this can be op-
provides the saturated value for u(k+j) to respect all input timized offline and its solution has a structure of a look-up
restrictions. For this approach, consider that N is finite table. The implementation of the eMPC does not require
and represents the prediction horizon. Therefore, there any online optimization. At each time sampling, the state
are N control signals for optimization for each input. The x(k) checks which region of Equation (27) is active through
elements Q and R are semi-definite and definite matrices, the process of branch search. After detecting the active
respectively, for tuning the system performance. Now, region, its respective u∗ (k + j) vector should be applied to
observe the last term of J(k), for j = N − 1: the system. The current control action (j = 0) is chosen
as the control signal and the controller holds for the next
J(k)j=N −1 = (x(k + N − 1))T Q(x(k + N − 1))
time sampling to update the states and repeat the process.
+(u(k + N − 1))T R(u(k + N − 1)) (20) This control setup has been programmed in the Multi-
For J(k) to be bounded, the term J(k)j=N −1 should Parametric Toolbox (MPT) for Matlab (Herceg et al.
be as small as possible, representing that at the end of (2013)). This toolbox has a complete set of functions
the prediction the system should be close to its desired for parametric optimization, computational geometry, and
states. For this bound certification, it is usual to include a especially a framework for MPC design. The MPT become
terminal penalty at the end of prediction, as in Equation a standard environment for eMPC since it has compatibil-
(21). The matrix P is called a terminal penalty and it ity with YALMIP (Löfberg (2004)) for MPC design and
is commonly computed using the Ricatti equation P = algorithms to solve multi-parametric optimizations.
AT P A + Q, where A is the state transition matrix and Q The MPT setup for MPC design has a list of requirements.
is the state weighting matrix. First, a Linear Time-Invariant (LTI) System object selects
a discrete linear model. Next, the set X is defined by
J(k)M P C = J(k) + (x(k + N ))T P (x(k + N )) (21)
inserting the upper and lower bounds of each state or
The MPC definition problem is stated in Lemma 1. declaring the polyhedra of constraints. The set U has
Lemma 1. The MPC controller computes the following similar requirements for the control input restrictions.
optimization (Camacho and Alba (2013)):
The matrices Q and R are included in the framework
min J(k)M P C (22) as quadratic functions for the states and control signals,
u(k+j) respectively. One remark is that it is possible to use one
subject to norm or infinity norm functions for MPC design. Addi-
tionally, there is a function that computes the terminal
x(k + 1) = Ax(k) + Bu(k) (23) penalty P using the usual LQR Penalty algorithm (Riccati
Equation). Finally, the MPT designs the MPC similar to
x(k + j) ∈ X (24)
Lemma 1. Furthermore, there is a function to explicit the
u(k + j) ∈ U (25) controller, applying Lemma 2.
for all j = 0, ... , N − 1 .
4. SIMULATION SETUP AND DISCUSSIONS
The Equation (23) is the constraint for the model pre-
diction. X is the parameters set that contains all feasible The software Simulink was chosen for the environment
values for the states and U is the set of all input signals that setup. Each system was simulated using the nonlinear
satisfies its restrictions, for example, the input saturations. equations of Section 2 and the eMPC has the objective
to control the system using the discrete linear model.
For each time sampling, the controller in Lemma 1 obtains The reaction wheel pendulum stabilizes the system on
the present state x(k) and solves the optimization problem. its inverted position θ1 = 0, subject to initial conditions
Hence, consider that the sets X , U, and the system model and disturbances. The Furuta pendulum stabilizes the
(Equation (23)) are known. Since all these restrictions are pendulum on its inverted position θ2 = 0 and tracks a
linear, they can be described as g(u(k + j), x(k)) ≤ 0. reference signal for arm angular position θ0 → θ0R , also
Also, note that u(k +j) is a function of x(k) because of the subject to initial conditions and disturbances.
feedback structure u(k +j) = K(j)x(k +j)+gj (j) and the
discrete linear model. Therefore, the cost function at the Since output measurements are usually noisy, a white noise
instant k can be described as JM P C = J(u(k + j), x(k)), signal was added for each output signal. The white noise
with u as the optimization vector and x the parameter signal had variance of 5 × 10−6 for angular measurements
vector. Finally, the Lemma 2 defines the eMPC problem. and 3 × 10−5 for angular velocities. Each noise signal was
Lemma 2. The eMPC controller computes the following generated with a distinct seed to minimize correlations
multi-parametric optimization (Borrelli et al. (2017)): between the signals.
Firstly, the parameter set X for the reaction wheel was 0.2

Pendulum Position θ1 (rad)

defined. After few simulations without viscous damping θ1 Disturbance
and noise signals, the maximum value for θ1 was 25◦ .
Considering that the DC motor has a symmetric operation, 0.1
the minimum value was θ1min = 25◦ (0.43 rad). Also,
θ̇1max = −θ̇1min = 5 rad/s; α̇max = −α̇min = 25 rad/s. 0
The tuning parameters for the reaction wheel pendulum
were:  
25 0 0 -0.1
Q =  0 10 0 ; R = 1.
0 0 1
-0.2
Next, the parameter set X for the Furuta pendulum was 0 2 4 6 8 10
defined similarly to the previous case. The arm angle po- Time (s)
sition was considered as a full rotation θ0max = −θ0min =
Figure 3. Reaction Wheel: Pendulum Position and Distur-
180◦ (π rad). The maximum angle for the pendulum was bance.
θ2max = −θ2min = 15◦ (0.26 rad). The maximum angle
15

Angular Velocities (rad/s)

velocities were
θ̇0max = −θ̇0min = θ̇2max = −θ̇2min = 5 rad/s.
10
The tuning parameters for the Furuta pendulum were:
3000 0 0 0
 
5
 0 50000 0 0 
Q=  0
 ; R = 1.
0 300 0  0
0 0 0 500
The control set U was the same for both systems since -5
its signal was modularized. Therefore, the maximum value θ̇1 α̇
was 1 and the minimum value was -1. Both systems had a -10
sampling frequency of fs = 50 Hz. The initial conditions 0 2 4 6 8 10
were all zeros except for each pendulum position, with Time (s)
θ1 (0) = θ2 (0) = 10◦ .
Figure 4. Reaction Wheel: Angular velocities.
The last parameter of the eMPC is the prediction horizon
N . For the reaction wheel, which has dimension 3, few 1
configurations were investigated. With N = 1, it had 3
Control Signal (PWM)

regions; N = 2 had 5 regions, N = 3 had 7 regions, up to

N = 30 with 61 regions. However, since the eMPC only 0.5
applies the first control signal (j = 0), it has been observed
that all cases for this system had the same performance.
This distinct behavior has a direct relation with the sets X 0
and U. For example, if a restrictive control signal would be
applied, the set U would be smaller and a larger prediction -0.5
horizon could improve the system performance. Since that
it was not the case for this system, the prediction horizon uMPC
N = 1 was chosen in the following simulations. -1
0 2 4 6 8 10
For the reaction wheel pendulum, Figure 3 shows the
pendulum position subject to an external disturbance. Time (s)
The explicit MPC was able to stabilize the system with a
initial condition and noisy measurements. The controller Figure 5. Reaction Wheel: Control Signal.
sustained with impulse disturbances and a small step chosen for simulations. Figure 6 shows the arm position
signal in t = 6 s. tracking the reference signal. In Figure 7, the pendulum
The Figures 4 and 5 shows the angular velocities and position is stabilized considering the external disturbances.
the control signal computed by the eMPC. Note that the Lastly, Figure 8 shows the angular velocites of the system
control signal had brief saturated impulses. and Figure 9 shows the control signal through the simula-
For the Furuta Pendulum, which has dimension 4, some tion.
configurations were investigated. N = 1 had 11 regions,
N = 3 had 296 regions, up to N = 5 with 2108 regions. 5. CONCLUSIONS
Note that the increase in system dimension strongly im-
pacts the number of regions. For this system, all configu- This work investigated the application of the eMPC tech-
rations also had the same performance, thus N = 1 was nique in two inverted pendulum control systems. The for-
 1.5 1
θ0 Set-Point θ0

Control Signal (PWM)

Arm Position θ0 (rad)

1
0.5
0.5

0 0
-0.5
-0.5
-1
uMPC
-1.5 -1
0 5 10 15 0 5 10 15
Time (s) Time (s)
Figure 6. Furuta: Arm Position and Reference Tracking. Figure 9. Furuta: Control Signal.
0.2 Bemporad, A., Morari, M., Dua, V., and Pistikopoulos,
E.N. (2002). The explicit linear quadratic regulator for
Pendulum Position θ2 (rad)

θ2 Disturbance
constrained systems. Automatica, 38(1), 3 – 20. doi:
0.1 https://doi.org/10.1016/S0005-1098(01)00174-1.
Block, D.J., Astrom, K.J., and Spong, M.W. (2007). The
Reaction Wheel Pendulum.
0 Borrelli, F., Bemporad, A., and Morari, M. (2017). Predic-
tive Control for Linear and Hybrid Systems. Cambridge
University Press. doi:10.1017/9781139061759.
-0.1 Camacho, E.F. and Alba, C.B. (2013). Model predictive
control. Springer Science & Business Media.
Herceg, M., Kvasnica, M., Jones, C., and Morari, M.
-0.2 (2013). Multi-Parametric Toolbox 3.0. In Proc. of the
0 5 10 15 European Control Conference, 502–510. Zürich, Switzer-
Time (s) land. http://control.ee.ethz.ch/~mpt.
Jmel, I., Dimassi, H., Hadj-Said, S., and M’Sahli, F.
Figure 7. Furuta: Pendulum Position and Disturbance. (2020). An adaptive sliding mode observer for inverted
pendulum under mass variation and disturbances with
2 experimental validation. ISA Transactions. doi:https://
Angular Velocities (rad/s)

doi.org/10.1016/j.isatra.2020.02.029.
1 Lenz, D., Kessler, T., and Knoll, A. (2015). Stochastic
model predictive controller with chance constraints for
0 comfortable and safe driving behavior of autonomous
vehicles. In 2015 IEEE Intelligent Vehicles Symposium
(IV), 292–297.
-1
Löfberg, J. (2004). Yalmip : A toolbox for modeling and
optimization in matlab. In In Proceedings of the CACSD
-2 Conference. Taipei, Taiwan.
θ̇0 θ̇2 Mori, S., Nishihara, H., and Furuta, K. (1976). Control
-3 of unstable mechanical system control of pendulum†.
0 5 10 15 International Journal of Control, 23(5), 673–692. doi:
Time (s) 10.1080/00207177608922192.
Qin, S.J. and Badgwell, T.A. (2003). A survey of industrial
Figure 8. Furuta: Angular velocities. model predictive control technology. Control Engineer-
ing Practice, 11(7), 733 – 764. doi:https://doi.org/10.
mulation of the parameter set X and the U is a crucial step 1016/S0967-0661(02)00186-7.
for eMPC design since unreachable states do not need to be Sugaya, J., Ohba, Y., and Kanmachi, T. (2017). Simula-
considered for the control system, reducing the number of tion of standing upright control of an inverted pendulum
constraints regions. Additionally, the small increase in the using inertia rotor and the swing type inverted pendu-
system dimension greatly affected the number of regions. lum for engineering education. In 2017 9th International
However, the increase of the prediction horizon N did not Conference on Information Technology and Electrical
improve the control performance for these control systems. Engineering (ICITEE), 1–6.

REFERENCES
Angélico, B.A. (2016). Website of lca (laboratório de
controle aplicado). https://sites.usp.br/lca/.