You are on page 1of 20

Received: 14 October 2020 Revised: 30 December 2020 Accepted: 2 February 2021

DOI: 10.1002/rnc.5472

RESEARCH ARTICLE

Anti-Swing control of the Pendubot using damper and


spring with positive or negative stiffness

Xin Xin1 Kazunori Makino1 Shinsaku Izumi1 Taiga Yamasaki1 Yannian Liu2

1
Faculty of Computer Science and
Systems Engineering, Okayama Abstract
Prefectural University, Okayama, Japan In this article, we study the anti-swing control of the Pendubot which is a
2
School of Automation, Southeast 2-link planar robot with a single actuator driving the first joint, whose control
University, Nanjing, China
objective is to stabilize the Pendubot to the downward equilibrium point (DEP)
Correspondence with the two links in the downward position for all its initial states with the
Xin Xin, Faculty of Computer Science and exception of a set of Lebesgue measure zero. To achieve this control objective,
Systems Engineering, Okayama
Prefectural University, Okayama
with respect to the angle of the first joint, we present a proportional-derivative
719-1197, Japan. (PD) controller using a damper and a linear spring with positive stiffness, and
Email: xxin@cse.oka-pu.ac.jp a sinusoidal-derivative (SD) controller using a damper and a nonlinear spring
Yannian Liu, School of Automation,
Southeast University, Nanjing, China. (force being sinusoidal function of the displacement) allowing even negative
Email: 108109095@seu.edu.cn stiffness. We prove that the control objective is achieved if some conditions on
the stiffness of linear or nonlinear spring are satisfied. We present analytical
Funding information
JSPS KAKENHI Grant, Grant/Award solutions to the optimal design of the derivative gain and the stiffness which
Number: 26420425 and 20K04554; minimizes the real parts of the dominant poles of the linearized model of the
National Natural Science Foundation of
corresponding closed-loop systems around the DEP for all Pendubots in terms
China, Grant/Award Number: 61973077
of their five physical parameters without any constraint. Our discussion shows
that the SD controller is superior to the PD controller. We provide simulation
results for a Pendubot to show that the presented SD controller can stabilize the
Pendubot to the DEP faster than the presented PD controller.

KEYWORDS
anti-swing control, linear or nonlinear spring, Lyapunov stability, negative stiffness, nonlinear
control, passivity, the Pendubot, underactuated mechanical systems

1 I N T RO DU CT ION

The Pendubot is a two-link planar robot with a single actuator at the base joint of the first link, and the joint between
two links is unactuated and allowed to swing freely.1,2 It, together with other mechanical systems such as the inverted
pendulum on a cart3 and the Acrobot,4 is used for control and robot education and for research as one of typical examples
of underactuated mechanical systems.5,6
Many researchers studied a particular problem of the set-point control called the swing up control for the Pendubot,
see, for example, References 1,7-10. Indeed, the swing up control is to swing the Pendubot to a small neighborhood of the
upright equilibrium point, where the two links are in the upright position, and balance the robot around that point.
In this article, we study a set-point control named as the anti-swing control of the Pendubot; that is, to stabilize the
Pendubot about the downward equilibrium point (denoted as DEP below) with two links in the downward position for all
Abbreviations: DEP, downward equilibrium point; PD, proportional-derivative; SD, sinusoidal-derivative.

Int J Robust Nonlinear Control. 2021;1–20. wileyonlinelibrary.com/journal/rnc © 2021 John Wiley & Sons Ltd. 1
2 XIN et al.

its initial states with the exception of a set of Lebesgue measure zero. The motivation of studying such a problem raised
from the anti-swing control of an overhead crane problem with a 1-DOF gantry, whose model is similar to a pendulum
on a cart with the pendulum being replaced by a rope with the payload at its end. To stabilize the rope in the downward
position with the cart being driven to a desired position, Reference 11 proposed a proportional-derivative (PD) controller
using the displacement of the cart, and Reference 12 proposed a PD controller using a variable which is the sum of the dis-
placement of the cart and a nonlinear function of the angle of the rope to improve transient performance of the anti-swing
control. The anti-swing control of an overhead crane is also investigated in Reference 13 by using a trajectory planning
approach.
In this article, we consider the case that the angle q1 and angular velocity q̇ 1 of the actuated first joint are measurable
by the encoder attached to the actuator, and the angle q2 and angular velocity q̇ 2 of the unactuated second joint are not
measurable. First, we prove that the PD controller using q1 can achieve the control objective when the derivative gain
is positive and the proportional gain (stiffness) is bigger than a positive constant or when the proportional gain is zero
which is the derivative (D) controller. Indeed, if the proportional gain is positive and is smaller than the positive constant,
we show that the Pendubot will be driven to other stable equilibrium points rather than the DEP for some initial states.
However, our simulation investigation for the Pendubot with its parameters taken from Reference 2 shows that it takes
a long time to drive the Pendubot close to the DEP under the above PD controller or D controller even with tuned gains.
Moreover, we find that for some initial states close to the DEP, the PD controller with a negative proportional gain has a
much better transient performance than that with the positive gain, but the stabilization of the DEP for some initial states
far away from the DEP is not guaranteed.
Although the DEP of the Pendubot is stable in the sense of Lyapunov,14 to achieve our control objective quickly is
not an easy task. To achieve our control objective with a better control performance, we present a controller using the
proportional term of sin q1 and the derivative term of q1 , which is called the sinusoidal-derivative (SD) controller in
this article, where S denotes sin. We analyze globally the motion of the Pendubot under the SD controller, and prove
that the SD controller can achieve the control objective when the derivative gain is positive and the gain for the sinu-
soidal term is greater than a negative constant. This indicates that the gain for the sinusoidal term can be chosen to be
negative.
It is known that the dominant poles of a linear stable system, whose real parts are closest to the imaginary axis, give
rise to the longest lasting terms in the transient response of the system.15 It is also mentioned in Reference 16 (p. 313)
that the farther to the left in the complex plane the system’s dominant poles are, the faster the system will respond and
the greater its bandwidth will be. Since the linearization of the SD controller around the DEP is the same as that of the
PD controller, we design the two control gains of the PD controller and SD controller by minimizing the real parts of
the dominant poles of the linearized model of the corresponding closed-loop systems around the DEP. Inspired by our
early conference paper17 for designing the gains of the PD controller for the TORA with its single physical parameter con-
strained in a certain range, in this article, by using the Routh–Hurwitz stability criterion, and studying the dominant poles
in a different way, we present analytical solutions to the above optimization problem for all Pendubots in terms of their
five physical parameters without any constraint. Indeed, first, we obtain the optimal control gains analytically by using
the Routh–Hurwitz criterion for the Pendubots with their closed-loop poles being the double complex conjugate roots.
Then, we obtain the optimal control gains analytically by using perturbation technique and employing Taylor expansion
for the class of the Pendubots with their closed-loop poles containing a triple real root. Our simulation investigation of
the Pendubot in Reference 2 illustrates the above analytical result and shows that the SD controller can achieve a better
performance than the PD controller.
Notations: Let R and S be the real number field and a unit circle, respectively.

2 PRELIMINARY KNOWLEDGE

First, consider the Pendubot shown in Figure 1, where for the ith (i = 1, 2) link, mi is its mass, li is its length, lci is the
distance from joint i to its center of mass (COM), and J i is the moment of inertia around its COM. The motion equation
of the Pendubot10 is:

M(q)q̈ + H(q, q)
̇ + G(q) = 𝜏, (1)

where q : = [ q1 , q2 ]T , 𝜏 ∶= [ 𝜏1 , 0 ]T with 𝜏1 being a single control torque applied to the first joint, and
XIN et al. 3

FIGURE 1 Pendubot (2-link planar robot)

[ ]
𝛼1 + 𝛼2 + 2𝛼3 cos q2 𝛼2 + 𝛼3 cos q2
M(q) ∶= , (2)
𝛼2 + 𝛼3 cos q2 𝛼2
[ ]
−2q̇ 1 q̇ 2 − q̇ 22
̇ ∶= 𝛼3
H(q, q) sin q2 , (3)
q̇ 21
[ ]
𝛽1 sin q1 + 𝛽2 sin(q1 + q2 )
G(q) ∶= , (4)
𝛽2 sin(q1 + q2 )

where

⎧𝛼1 ∶= J1 + m1 l2c1 + m2 l21 ,



⎨𝛼2 ∶= J2 + m2 l2c2 , 𝛼3 ∶= m2 l1 lc2 , (5)

⎩𝛽1 ∶= (m1 lc1 + m2 l1 )g, 𝛽2 ∶= m2 lc2 g,

with g being the acceleration of gravity. In this article, we treat qi (t) (i = 1, 2) in S unless we specify to treat in R.
Next, we recall the following results which are useful in this article.
Lemma 1. Let x = xe be an equilibrium point of
ẋ = f (x), (6)

where f (x) ∶ D → Rn is continuously differentiable and D is a neighborhood of xe . By linearizing (6) at x = xe to obtain the
linearized system
x̃̇ = Ãx, (7)

where x̃ ∶= x − xe , and
𝜕f (x) ||
A= (8)
𝜕x ||x=xe

is the Jacobian matrix of f (x) evaluated at x = xe . Then the following statements hold:

1) 14 (p. 139) The equilibrium point is asymptotically stable if all eigenvalues of A are in the open left-half plane;
2) 14 (p. 139) The equilibrium point is unstable if there exists at least one eigenvalue of A in the open right-half
plane;
3) 18 (p. 1225) If A has at least one eigenvalue in the open right-half plane and has at least one eigenvalue in the open left-half
plane, then associated to the latter there is a stable manifold, and the trajectories starting in this manifold will converge to
such an equilibrium point x = xe . Moreover, the set of initial conditions converging to such an unstable equilibrium x = xe
has Lebesgue zero measure.

The stable manifold of (6) at the equilibrium point x = xe is tangent to the eigenspace of eigenvalues with negative real
parts of A in (8), please refer Theorems 1.3.2 and 3.2.1 of Reference 19 for further information. According to the Statement
4 XIN et al.

3) of Lemma 1, since A has at least one eigenvalue in the open right-half plane and has at least one eigenvalue in the open
left-half plane, the dimension of the above eigenspace is greater than 0 and less than n. Thus, the set of initial conditions
converging to such an unstable equilibrium x = xe has Lebesgue zero measure in Rn . In other words, the solutions of (6)
starting for almost all its initial conditions do not converge to such an unstable equilibrium x = xe .

3 DESIGN OF ANTI- S WING CO NTRO LLERS AND MOTION ANALYSIS

In this article, our goal is to design 𝜏1 such that

lim q(t) = 0, ̇ =0
lim q(t) (9)
t→∞ t→∞

for all initial states of the Pendubot with the exception of a set of Lebesgue measure zero.

3.1 Controller using damping and linear spring

Consider the following PD controller:

𝜏1 ∶= −kD q̇ 1 − kP q1 , (10)

where kD > 0 and kP ≥ 0 are two constant gains. First, we present the following lemma with its proof given in Appendix A.
Lemma 2. Consider the closed-loop system consisting of (1) and (10). Suppose that kD > 0 and kP ≥ 0 hold. Then,
̇
starting from any initial state, the closed-loop solution (q(t), q(t)) approaches an equilibrium point of the closed-loop
system as t → ∞.
Second, we characterize the equilibrium points of the closed-loop system consisting of (1) and (10). Let
qe ∶= [qe1 , qe2 ]T be a closed-loop equilibrium configuration. Putting q̈ = 0, q̇ = 0, q = qe , and 𝜏1 = 𝜏1e ∶= −kP qe1 into (1),
we obtain

kP e
q = −sin qe1 , (11)
𝛽1 1
sin(qe1 + qe2 ) = 0. (12)

Define the following set which contains all equilibrium points of the closed-loop system:

Ω ∶= {(qe , 0) | qe satisfies (11) and (12)}. (13)

Regarding (11), as shown in Figure 2, we want to find kP such that y1 = (kP ∕𝛽1 )q1 is tangent to y2 = −sin q1 at q1 = ±𝜓1
(𝜋 < 𝜓1 < 3𝜋∕2). This with (11) yields

kP kP
= −cos 𝜓1 , 𝜓1 = −sin 𝜓1 , (14)
𝛽1 𝛽1

where the first equation means that kP ∕𝛽1 is the inclination of the tangent line of − sin q1 (the derivative of −sin q1 with
respect to q1 ) at q1 = ±𝜓1 . This shows
( )
3𝜋
𝜓1 = tan 𝜓1 𝜋 < 𝜓1 < ⇒ 𝜓1 = 4.4934, (15)
2
kP = −𝛽1 cos 𝜓1 ⇒ kP = 0.2172𝛽1 . (16)

We present the following lemma, whose proof is presented in Appendix B.


XIN et al. 5

k
FIGURE 2 Plots of y1 = 𝛽P q1 (with three kP ) and y2 = −sin q1 [Colour
1
figure can be viewed at wileyonlinelibrary.com]

Lemma 3. Consider the closed-loop system consisting of (1) and (10). The following statements hold:

1) The DEP (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 0, 0, 0) is asymptotically stable if

kD > 0, kP > −𝛽1 . (17)

2) If

kD > 0, kP > −𝛽1 cos 𝜓1 , (18)

where 𝜓1 is defined in (15), then Ω in (13) only contains the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 0, 0, 0), which is asymptotically
stable, and the down–up equilibrium point (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 𝜋, 0, 0), which is unstable with its Jacobian matrix having
eigenvalue(s) in both the open right-half plane and the open left-half plane.
3) If

kD > 0, 0 < kP < −𝛽1 cos 𝜓1 , (19)

then Ω in (13) contains at least two asymptotically stable equilibrium points (q1 , q2 , q̇ 1 , q̇ 2 ) = (𝜌12 , 2𝜋 − 𝜌12 , 0, 0),
(−𝜌12 , −2𝜋 + 𝜌12 , 0, 0) in addition to the DEP, where 𝜌12 ∈ ( 3𝜋∕2, 2𝜋 ) is the solution of (11) depicted in Figure 2.
4) If kD > 0 and kP = 0, then Ω in (13) satisfies

Ω = Ωs , (20)

where
{ [ ]T [ ]T [ ]T [ ]T }
Ωs ∶= (qe , 0) | qe = 0, 0 , 𝜋, 0 , 0, 𝜋 , 𝜋, 𝜋 (mod 2𝜋) , (21)

in which only the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 0, 0, 0) is asymptotically stable, and the other three equilibrium points in Ωs
are unstable with each of their Jacobian matrices having eigenvalue(s) in both the open right-half plane and the open
left-half plane.

We present the following proposition.


Proposition 1. Consider the closed-loop system consisting of (1) and (10). If

kD > 0, kP > −𝛽1 cos 𝜓1 or kP = 0 (22)

̇
with 𝜓1 in (15), then as t → ∞ the closed-loop solution (q(t), q(t)) approaches the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 0, 0, 0) for all initial
states with exception of a set of Lebesgue measure zero.
6 XIN et al.

Proof. From Lemmas 2 and 3, under kD > 0, if kP > −𝛽1 cos 𝜓1 or kP = 0, then the closed-loop solution approaches one of
equilibrium points in Ω with only the DEP being asymptotically stable, and the other equilibrium points being unstable
with each of their Jacobian matrices having eigenvalue(s) in both the open right-half plane and the open left-half plane.
This with the Statement 3) of Lemma 1 completes the proof of Proposition 1. ▪
Remark 1. Let kD > 0. From Lemma 3, when −𝛽1 < kP < 0 holds, although the DEP is asymptotically stable (locally),
under the PD controller (10), the achievement of the control objective cannot be guaranteed for the initial states of the
Pendubot far away from the DEP. When 0 < kP < −𝛽1 cos 𝜓1 holds, since the set Ω contains other stable equilibrium
points, the Pendubot cannot be stabilized to the DEP from some neighborhoods close to these other stable equilibrium
points.

3.2 Controller using damping and nonlinear spring

Consider the following SD controller:

𝜏1 ∶= −kD q̇ 1 − kP sin q1 , (23)

where kD > 0 and kP ∈ R are constant gains. To achieve a better performance of stabilization of the DEP, kP ∈ R in (23)
is different from kP ≥ 0 in (10).
First, we present the following lemma with its proof given in Appendix C.
Lemma 4. Consider the closed-loop system consisting of (1) and (23). Suppose that kD > 0 and kP ≠ −𝛽1 hold. Then, starting
̇
from any initial state, the closed-loop solution (q(t), q(t)) approaches an equilibrium point of the equilibrium set Ωs defined
in (21) as t → ∞.
Next, we present the following proposition of this article.
Proposition 2. Consider the closed-loop system consisting of (1) and (23). If (17) holds, then the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) =
(0, 0, 0, 0) in Ωs in (21) is asymptotically stable, and the other three equilibrium points in Ωs in (21) are unstable with each
of their Jacobian matrices having eigenvalue(s) in both the open right-half plane and the open left-half plane; moreover, as
̇
t → ∞ the closed-loop solution (q(t), q(t)) approaches the DEP for all initial states with exception of a set of Lebesgue measure
zero.

Proof. Since the linearization of (23) around the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 0, 0, 0) is (10), similar to the proof of the State-
ments 1) and 4) of Lemma 3, we can show that if (17) holds, then only the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 0, 0, 0) in Ωs in (21) is
asymptotically stable, and we can show that each of their Jacobian matrices has eigenvalue(s) in both the open right-half
plane and the open left-half plane. The detail is omitted. This with the Statement 3) of Lemma 1 completes the proof of
Proposition 2. ▪

4 DE S I G N OF T WO CON T ROL GAINS

In this section, we study how to design two control gains kD and kP in (10) and (23). Since the linearized model
of the closed-loop system consisting of (1) and (23) around the DEP is the same as that consisting of (1) and (10).
The characteristic polynomial determined by the Jacobian matrix of the above closed-loop systems evaluated at the
DEP is

P(s) ∶= s4 + a1 s3 + a2 s2 + a3 s + a4 , (24)

where

⎧ kD 𝛼2
⎪a1 ∶= Δ , a2 ∶= kP 𝛼2Δ + 𝛿 ,
⎨ (25)
⎪a3 ∶= kD 𝛽2 , a4 ∶=
(kP + 𝛽1 )𝛽2
,
⎩ Δ Δ
XIN et al. 7

where

Δ ∶= |M(0)| = 𝛼1 𝛼2 − 𝛼32 > 0, 𝛿 ∶= 𝛼2 𝛽1 + 𝛼1 𝛽2 . (26)

To improve the control performance of the controllers (10) and (23), we design the two control gains kD and kP to
minimize the real parts of the dominant poles (denoted as RDP); that is, to minimize

RDP(kD , kP , 𝛾) ∶= max Re [𝜆i (kD , kP , 𝛾)], (27)


1≤i≤4

where Re [𝜆i ] denotes the real part of 𝜆i , and 𝜆i (kD , kP , 𝛾) (1 ≤ i ≤ 4) are the roots of P(s) in (24), and 𝛾 ∶=
[ ]T
𝛼1 , 𝛼2 , 𝛼3 , 𝛽1 , 𝛽2 is a vector containing the positive parameters in (5) of the Pendubot.
The minimum of RDP(kD , kP , 𝛾) is defined as:

RDPm (𝛾) ∶= min RDP(kD , kP , 𝛾), (28)


kD >0, kP >−𝛽1

and the optimal control gains are defined as:

(kDm (𝛾), kPm (𝛾)) ∶= arg minkD >0, kP >−𝛽1 RDP(kD , kP , 𝛾). (29)

Although 𝜆i (kD , kP , 𝛾) (1 ≤ i ≤ 4) are complicated functions of kD , kP , and 𝛾, we present analytical solutions of (28) and
(29) in the following two theorems, which are the main results of this article.
Theorem 1. If the parameters of the Pendubot in (5) satisfy

4𝛼1 𝛼2 − 5𝛼32 ≥ 0, (30)

then the analytical solution of (28) and (29) is



𝛼3 𝛼2 𝛽2 Δ
RDPm (𝛾) = −𝜁a , 𝜁a ∶= > 0, (31)
2𝛼2 Δ

2𝛼3 𝛼2 𝛽2 Δ 𝛽2 Δ
kDm (𝛾) = , kPm (𝛾) = − 𝛽1 , (32)
𝛼22 𝛼22

where Δ is defined in (26). In this case, the roots of P(s) in (24) are −𝜁a ± j𝜔a and −𝜁a ± j𝜔a , where
√( )
4𝛼1 𝛼2 − 5𝛼32 𝛼2 𝛽2 Δ
𝜔a ∶= . (33)
2𝛼2 Δ

Theorem 2. If the parameters of the Pendubot in (5) satisfy

4𝛼1 𝛼2 − 5𝛼32 < 0, (34)

then the analytical solution of (28) and (29) is


√ ( √ )

√ 𝛽 2𝛼 Φ − Δ
√ 2 3
RDPm (𝛾) = −𝜁b , 𝜁b ∶= √ √ > 0, (35)
𝛼2 Δ

2Δ𝜁b 2𝛼 2 𝛽 2 𝜁b
kDm (𝛾) = + ( 3 2 ) , (36)
𝛼2 𝛼2 𝛼2 𝜁b2 + 𝛽2 2
8 XIN et al.

𝛿 6𝛼32 𝛽22 𝜁b2


kPm (𝛾) = − + ( ) , (37)
𝛼2 𝛼 𝛼 𝜁 2 + 𝛽 2
2 2 b 2

where Δ and 𝛿 are defined in (26), and


( ( ( √ )))
1 2 Δ
Φ ∶= sin 𝜋 − 2cos−1
. (38)
6 𝛼3

In this case, the roots of the P(s) in (24) are −𝜁b , −𝜁b , −𝜁b , and −𝜁b − 𝜁br , where
( )
2𝜁b 𝛼22 Δ𝜁b4 + 2𝛼2 𝛽2 Δ𝜁b2 + (Δ − 𝛼32 )𝛽22
𝜁br ∶= − ( )2 > 0. (39)
𝛼2 𝜁b2 + 𝛽2 Δ

After providing some common formulae for proving Theorems 1 and 2 in Appendix D, we present the proof of
Theorem 1 by using the Routh–Hurwitz criterion in Appendix E and present the proof of Theorem 2 by using perturba-
tion technique and employing Taylor expansion in Appendix F. It is worth pointing out that the proof of Theorem 2 is
more difficult than that of Theorem 1.
Finally, we present the following remark about (30).
Remark 2. There do exist two-link robots with their parameters 𝛼1 , 𝛼2 , 𝛼3 , 𝛽1 , and 𝛽2 satisfying (30). Here, we consider
all two-link robots with each link having uniform density which are often used as typical examples of robotic systems for
analytical analysis and simulations of their control systems. Using

li mi l2i
lci = , Ji = , 1≤i≤2
2 12

shows that (30) holds due to


l21 l22 m2 (16m1 + 3m2 )
4𝛼1 𝛼2 − 5𝛼32 = > 0.
36

5 DISCUSSION

First, we present Table 5 to compare three controllers including the D controller which is a special case of the PD or SD
controller with kP = 0. On the one hand, from Proposition 1, regarding the closed-loop system consisting of (1) and the PD
̇
controller (10), we know that if (22) holds, as t → ∞ the closed-loop solution (q(t), q(t)) approaches the DEP for all initial
states with exception of a set of Lebesgue measure zero. If the optimal gains in Theorems 1 or 2 also satisfy (22), they
are appropriate for almost all initial states including those far away from the DEP. If these optimal gains do not satisfy
(22), the convergence to the DEP is not guaranteed for the initial states far away from the DEP. On the other hand, from
Proposition 2, regarding the closed-loop system consisting of (1) and the SD controller (23), if (17) holds, as t → ∞ the
̇
closed-loop solution (q(t), q(t)) approaches the DEP for all initial states with exception of a set of Lebesgue measure zero.
Since the optimal gains in Theorem 1 or Theorem 2 satisfy (17), they are appropriate for almost all initial states including
those far away from the DEP.
Second, on the one hand, it is stated in Spong et al.20 (p. 285) that it is customary in robotics applications to take
the damping ratio being 1 for a second-order linear system so that the response is critically damped. Indeed, the critical
damping (the damping ratio being 1) is the result of designing damping coefficient to minimize the real part of its poles
close to imaginary axis. On the other hand, it is also quite often that the dominant poles occur in the form of a complex
conjugate pair, see Reference 15 (p. 182), for example. Since we have obtained the minimum of the real parts of the poles
closest to the imaginary axis; that is, −𝜁a in Theorem 1 or −𝜁b in Theorem 2, by using −𝜁a or −𝜁b , it is an interesting future
research subject to study how to determine a possible region of dominant poles in the form of complex conjugate pair for
designing two gains of our presented controllers.
XIN et al. 9

T A B L E 1 Comparison of three
Control PD D SD
controllers
Condition kD > 0 kD > 0 kD > 0
kP > −𝛽1 cos 𝜓1 kP = 0 kP > −𝛽1
Asymptotically stable q1 = 0 q1 = 0 (mod 2𝜋) q1 = 0 (mod 2𝜋)
equilibrium configuration q2 = 0 (mod 2𝜋) q2 = 0 (mod 2𝜋) q2 = 0 (mod 2𝜋)

Abbreviations: PD, proportional-derivative; SD, sinusoidal-derivative.

FIGURE 3 Time responses of q(t) and 𝜏1 (t) under the 3


sinusoidal-derivative controller (23) with gains (kD , kP ) = (kDm , kPm ) which are in 2
1
(42) (negative stiffness) [Colour figure can be viewed at wileyonlinelibrary.com] 0
-1
0 2 4 6 8 10
2
0
-2
0 2 4 6 8 10
3
2
1
0
-1
0 2 4 6 8 10
Time [s]

6 S I M UL AT ION R E SU LT S

To illustrate our obtained results, we studied two Pendubots.


Pendubot 1. We used the following parameters of the Pendubot, manufactured by Mechatronics Systems Inc.:2
𝛼1 ∶= 0.0308 Vs2 , 𝛼2 ∶= 0.0106 Vs2 , 𝛼3 ∶= 0.0095 Vs2 , 𝛽1 ∕g ∶= 0.2086 Vs2 ∕m, and 𝛽2 ∕g ∶= 0.0630 Vs2 ∕m. We took
g : = 9.81 m/s2 .
To achieve the control objective, from Propositions 1 and 2, the conditions on kD and kP are

kD > 0, kP > 0.4445 (40)

for the PD controller (10) due to (18), and

kD > 0, kP > −2.0464 (41)

for the SD controller (23) due to (17).


Since 4𝛼1 𝛼2 − 5𝛼32 = 0.0008547 > 0 holds, from (31) and (32), we obtain RDPm = −2.3598 and

kDm = 0.2104, kPm = −0.7470. (42)

The roots of the polynomial in (24) with kD = kDm and kP = kPm are −2.3598 ± 7.2620j and −2.3598 ± 7.2620j which are the
same as those expressed in Theorem 1.
For an initial state (q1 (0), q2 (0), q̇ 1 (0), q̇ 2 (0)) = (2.5, −0.5, 0, 0), the time responses of q(t) and 𝜏1 (t) of the Pendobot
under the SD and PD controllers with gains (kD , kP ) = (kDm , kPm ) which are in (42) are depicted in Figures 3 and 4. From
Figure 3, we can see that the Pendubot is driven quickly to the DEP. However, from Figure 4, the time responses of the
state of the Pendubot are not converge to the DEP. This is because the gain kP = kPm which is in (42) do not satisfy (40).
We use the PD controller (10) with

kD ∶= kDm , kP ∶= −kPm = 0.7470, (43)


10 XIN et al.

F I G U R E 4 Time responses of q(t) and 𝜏1 (t) under the


100 proportional-derivative controller (10) with gains (kD , kP ) = (kDm , kPm ) which
50 are in (42) (negative stiffness) [Colour figure can be viewed at
0 wileyonlinelibrary.com]
0 0.5 1 1.5 2
0

-5

-10
0 0.5 1 1.5 2
40
20
0
0 0.5 1 1.5 2
Time [s]

3 F I G U R E 5 Time responses of q(t) and 𝜏1 (t) under the


2 proportional-derivative controller (10) with the gains in (43) (positive
1
0 stiffness) [Colour figure can be viewed at wileyonlinelibrary.com]
-1
0 2 4 6 8 10
2
0
-2
0 2 4 6 8 10
2

-2
0 2 4 6 8 10
Time [s]

which satisfies (40). The time responses of q(t) and 𝜏1 (t) of the Pendubot are depicted in Figure 5. In comparison with
Figure 3, we can see that it takes longer time to stabilize the Pendubot to the DEP by the PD controller (10) than by the
SD controller (23). There is no much difference in the maximal values of the |𝜏1 (t)| of these two controllers. Note that
the roots of the polynomial in (24) with the gains in (43) are −0.4417 ± 6.9784j and −4.2780 ± 11.4530j, and the ratio is
(−0.4417)/RDPm = 0.1872.
Now we illustrate the Statement 3) of Lemma 3. If we choose kP ∶= 0.3 < −𝛽1 cos 𝜓1 = 0.4445, then (11) has solutions
qe1 = 0, ±3.71807, ±5.3755, which indicates from Figure 2 that 𝜌11 = 3.71807 and 𝜌12 = 5.3755. In addition to
the DEP, the equilibrium points (q1 , q2 , q̇ 1 , q̇ 2 ) = (5.3755, 0.9077, 0, 0) and (q1 , q2 , q̇ 1 , q̇ 2 ) = (−5.3755, −0.9077, 0, 0) are
stable.
For an initial state (q1 (0), q2 (0), q̇ 1 (0), q̇ 2 (0)) ∶= (4.5, 0, 0, 0), the time responses of q(t) and 𝜏1 (t) of the Pendobot under
the PD controller (10) and the SD controller (23) with gains

kD ∶= kDm , kP ∶= 0.3 (44)

are depicted in Figures 6 and 7, respectively. The closed-loop solution approaches the equilibrium point
(q1 , q2 , q̇ 1 , q̇ 2 ) = (5.3755, 0.9077, 0, 0) under the PD controller (10) as shown in Figure 6, while approaches the DEP (in
mod 2𝜋) under the SD controller (23) as shown in Figure 7.
Pendubot 2. We consider the Pendubot robot with m1 = 1 kg, m2 = 2 kg, l1 = 1 m, l2 = 2 m, lc1 = 0.5 m, lc2 = 1 m,
J1 = 1∕48 kg ⋅ m2 , J2 = 1∕6 kg ⋅ m2 . Since 4𝛼1 𝛼2 − 5𝛼32 = −(23∕72) < 0 holds, from Theorem 2, we obtain 𝜁b = 2.5731,
kDm = 5.3557, kPm = −20.6167, and the poles of the closed-loop system are −2.5731, −2.5731, −2.5731, and −4.8919. The
report of the simulation results for Pendubot 2 similar to those for Pendubot 1 is omitted.
XIN et al. 11

FIGURE 6 Time responses of q(t) and 𝜏1 (t) under the 6


proportional-derivative controller (10) with the gains in (44) (small positive
5
stiffness) [Colour figure can be viewed at wileyonlinelibrary.com]
4
0 2 4 6 8 10
2
1
0
0 2 4 6 8 10
-1

-2

-3
0 2 4 6 8 10
Time [s]

FIGURE 7 Time responses of q(t) and 𝜏1 (t) under the


2
sinusoidal-derivative controller (23) with the gains in (44) (small positive
stiffness) [Colour figure can be viewed at wileyonlinelibrary.com] 5
4
0 2 4 6 8 10
2

-2
0 2 4 6 8 10
1
0
-1
-2
0 2 4 6 8 10
Time [s]

7 CO N C LU S I O N

In this article, we studied the anti-swing control of the Pendubot, whose control objective is to stabilize the Pen-
dubot to the DEP for all its initial states with the exception of a set of Lebesgue measure zero. To achieve
this control objective, by using the angle and the angular velocity of the first joint, we presented a PD con-
troller (using a linear spring with positive stiffness), and an SD controller (using the nonlinear spring allowing
even negative stiffness). We analyzed the solutions and the equilibrium points of the closed-loop systems con-
sisting of the Pendubot and the two controllers, and presented the conditions on the two control gains of the
PD controller and the SD controller for achieving the control objective in Propositions 1 and 2 of this article,
respectively.
We designed the two control gains such that the real parts of the dominant poles of the linearized model of
the closed-loop systems around the DEP are minimized, and presented analytical solutions of the optimal control
gains for all Pendubots without any constraint on their physical parameters. Indeed, we obtained the optimal con-
trol gains analytically by using the Routh–Hurwitz criterion for the class of the Pendubots with their closed-loop
poles being the double complex conjugate roots, see Theorem 1 of this article. We obtained the optimal control
gains analytically by using perturbation technique and employing Taylor expansion for the rest of the Pendubots
with their closed-loop poles containing a triple real root, see Theorem 2 of this article. Our discussion and simula-
tion investigation showed that the SD controller is superior to the PD controller. We provided simulation results for
the Pendubot to show that the presented SD controller can stabilize the Pendubot to the DEP faster than the PD
controller.
The study for the Pendubot in this article paved a way for studying the anti-swing control for other underactuated
systems, for example, an overhead crane problem with a 1-DOF gantry in Reference 12 and for determining PD gains in
other control problems.
12 XIN et al.

ACKNOWLEDGMENTS
This work was supported in part by JSPS KAKENHI Grant Numbers 26420425 and 20K04554, by the National Natural
Science Foundation of China under grant number 61973077. The authors would like to express their gratitude to Mr. K.
Miyake for useful discussions of the content of this article.

CONFLICT OF INTEREST
The authors declare no potential conflict of interest.

DATA AVAILABILITY STATEMENT


Data available on request from the authors.

ORCID
Xin Xin https://orcid.org/0000-0002-0965-5930

REFERENCES
1. Spong MW, Block DJ. The Pendubot: a mechatronic system for control research and education. Paper presented at: Proceedings of the
34th IEEE Conference on Decision and Control, New Orleans, LA; 1995:555-556.
2. Mechatronic Systems Inc. The Pendubot User’s Manual. Mechatronic Systems Inc. Champaign, Illinois; 1997.
3. Åström KJ, Furuta K. Swinging up a pendulum by energy control. Automatica. 2000;36(2):287-295.
4. Spong MW. The swing up control problem for the acrobot. IEEE Control Syst Mag. 1995;15(1):49-55.
5. Li S, Moog CH, Respondek W. Maximal feedback linearization and its internal dynamics with applications to mechanical systems on R4 .
Int J Robust Nonlinear Control. 2019;29(1):2639-2659.
6. Ruiz-Duarte JE, Loukianov AG. Sliding mode output-feedback causal output tracking for a class of discrete-time nonlinear systems. Int
J Robust Nonlinear Control. 2019;29(6):1956-1975.
7. Fantoni I, Lozano R, Spong MW. Energy based control of the Pendubot. IEEE Trans Autom Control. 2000;45(4):725-729.
8. Ma X, Su CY. A new fuzzy approach for swing up control of Pendubot. Paper presented at: Proceedings of the 2002 American Control
Conference, Anchorage, AK; 2002:1001-1006.
9. Zhang M, Tarn T. Hybrid control of the Pendubot. IEEE/ASME Trans Mechatron. 2002;7(1):79-86.
10. Xin X, Tanaka S, She J, Yamasaki T. New analytical results of energy-based swing-up control for the Pendubot. Int J Non-Linear Mech.
2013;52:110-118.
11. Collado J, Lozano R, Fantoni I. Control of convey-crane based on passivity. Paper presented at: Proceedings of the 2000 American Control
Conference, Chicago, IL; 2000:1260-1264.
12. Sun N, Fang Y. New energy analytical results for the regulation of underactuated overhead cranes: an end-effector motion-based approach.
IEEE Trans Ind Electron. 2012;59(12):4723-4734.
13. Sun N, Fang Y. An efficient online trajectory generating method for underactuated crane systems. Int J Robust Nonlinear Control.
2014;24(11):1653-1663.
14. Khalil HK. Nonlinear Systems. 3rd ed. Upper Saddle River, NJ: Prentice-Hall; 2002.
15. Ogata K. Modern Control Engineering. 5th ed. Upper Saddle River, NJ: Prentice-Hall; 2010.
16. Golnaraghi F, Kuo BC. Automatic Control Systems. 9th ed. New York, NY: Wiley; 2009.
17. Sumida K, Xin X, Yamasaki T. Optimal PD control design via dominant pole assignment for a class of TORA systems. Paper presented at:
Proceedings of the 2014 IEEE/SICE International Symposium on System Integration, Tokyo, Japan; 2014:257-262.
18. Ortega R, Spong MW, Gomez-Estern F, Blankenstein G. Stabilization of a class of underactuated mechanical systems via interconnection
and damping assignment. IEEE Trans Autom Control. 2002;47(8):1218-1233.
19. Guckenheimer J, Holmes P. Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields. New York, NY: Springer; 2002.
20. Spong MW, Hutchinson S, Vidyasagar M. Robot Modeling and Control. 2nd ed. New York, NY: John Wiley & Sons; 2020.
21. Zwillinger D. CRC Standard Mathematical Tables and Formulas (Advances in Applied Mathematics). 33rd ed. Boca Raton, FL: Chapman
& Hall/CRC Press; 2018.

How to cite this article: Xin X, Makino K, Izumi S, Yamasaki T, Liu Y. Anti-Swing control of the Pendubot
using damper and spring with positive or negative stiffness. Int J Robust Nonlinear Control. 2021;1–20. https://doi.
org/10.1002/rnc.5472
XIN et al. 13

APPENDIX A. PROOF OF LEMMA 2

We design the following Lyapunov function candidate:

1
̇ = E(q, q)
V(q, q) ̇ + kP q21 , (A1)
2

̇ is the total energy of the Pendubot expressed as:


where E(q, q)

1 T
̇ ∶=
E(q, q) q̇ M(q)q̇ + P(q), (A2)
2

with P(q) being the potential energy of the Pendubot defined as:

P(q) ∶= 𝛽1 (1 − cos q1 ) + 𝛽2 (1 − cos(q1 + q2 )).

Taking the time-derivative of V along the trajectory of (1), and using

Ė = q̇ 1 𝜏1 , (A3)

we obtain

V̇ = q̇ 1 (𝜏1 + kP q1 ).

Using (10) yields

V̇ = −kD q̇ 21 ≤ 0. (A4)

Next, to use LaSalle’s invariance principle14 to proceed the proof, we need to define a compact (closed and bounded)
set that is positively invariant with respect to the closed-loop system consisting of (1) and (10). By treating q1 ∈ R and
q2 ∈ S, we define a state
[ ]T [ ]T
z = z1 , z2 , z3 , z4 , z5 ∶= q1 , sin q2 , cos q2 , q̇ 1 , q̇ 2 ,

and using

ż 2 = q̇ 2 cos q2 = z3 z5 , ż 3 = −q̇ 2 sin q2 = −z2 z5 , (A5)

we can rewrite the closed-loop system consisting of (1) and (10) as ż = F(z), where the explicit expression of F(z) is omitted.
Define

Γc = {z | V ≤ c}, (A6)

where c is a positive constant. Owing to V̇ ≤ 0 in (A4) and V ≥ 0, V is bounded. This shows that E(q, q),
̇ q1 , q̇ 1 , and q̇ 2 are
̇
bounded. This, together with z2 and z3 being bounded, proves that Γc is a compact set. Since V ≤ 0 in (A4), any solution
of ż = F(z) starting in Γc remains in Γc for all t ≥ 0. Let W be the largest invariant set in

S = {z ∈ Γc | V̇ = 0}. (A7)

Using LaSalle’s invariance principle14 yields that every z(t) starting in Γc approaches W as t → ∞. Since V̇ = 0 holds
identically in W, we conclude that q1 and 𝜏1 are constants in W, which are denoted as q∗1 and 𝜏1∗ , respectively.
Below we prove that q2 is a constant in W. The proof for this fact is similar to the corresponding statement in the proof
of lemma 2 in Reference 10, which is expressed below for a self-contained purpose. Putting q1 ≡ q∗1 and 𝜏1 ≡ 𝜏1∗ = −kP q∗1
(≡ means an equation holds for all time t) into (1) yields
14 XIN et al.

(𝛼2 + 𝛼3 cos q2 )q̈ 2 − 𝛼3 q̇ 22 sin q2 + 𝛽2 sin(q∗1 + q2 ) = 𝜂1 , (A8)

𝛼2 q̈ 2 = −𝛽2 sin(q∗1 + q2 ), (A9)

where

𝜂1 ∶= −kP q∗1 − 𝛽1 q∗1 . (A10)

Putting (A9) into (A8) yields

𝛼3 q̈ 2 cos q2 − 𝛼3 q̇ 22 sin q2 = 𝜂1 , (A11)

which follows that


d (q̇ 2 cos q2 )
𝛼3 = 𝜂1 . (A12)
dt

Since 𝜂1 is a constant, integrating the above equation with respect to time t yields

𝛼3 q̇ 2 cos q2 = 𝜂1 t + 𝜂2 , (A13)

where 𝜂2 is a constant. Since (A13) holds for any t and q̇ 2 is bounded (due to the boundedness of the Lyapunov function
V in W), 𝜂1 = 0 holds; otherwise, the right-hand side of (A13) will go infinity as t → ∞.
Rewriting (A13) with 𝜂1 = 0, we have

d(sin q2 ) 𝜂2
q̇ 2 cos q2 = = . (A14)
dt 𝛼3

Integrating the above equation with respect to time t yields

𝜂2
sin q2 = t + 𝜂3 , (A15)
𝛼3

where 𝜂3 is a constant. Similar to the proof of 𝜂1 = 0, we obtain 𝜂2 = 0 and sin q2 = 𝜂3 . Thus, q2 is a constant in W denoted
̇
as q∗2 . Thus, the closed-loop solution (q(t), q(t)), approaches an equilibrium point as t approaches ∞. ■

APPENDIX B. PROOF OF LEMMA 3

1) The characteristic polynomial of the Jacobian matrix of the closed-loop system evaluated at the DEP (q1 , q2 , q̇ 1 , q̇ 2 ) =
(0, 0, 0, 0) is given in (24). Using the Routh–Hurwitz stability criterion16 to (24), we know that P(s) in (24) is stable if
and only if
ai > 0, 1 ≤ i ≤ 4, (B1)

kD2 𝛼32 𝛽22


D3 ∶= a1 a2 a3 − a23 − a21 a4 = > 0, (B2)
Δ3

where each ai (1 ≤ i ≤ 4) is defined in (25) and D3 is a Hurwitz determinant of (24). Thus, (B1) and (B2) hold if and
only if (17) holds. If P(s) is stable, then the DEP is asymptotically stable from the Statement 1) of Lemma 1.
2) Under (18), qe1 = 0 is the only solution of (11) as shown in Figure 2. Thus, Ω in (13) contains only the DEP and the
down–up equilibrium point (q1 , q2 , q̇ 1 , q̇ 2 ) = (0, 𝜋, 0, 0). Under kD > 0, the DEP is asymptotically stable due to kP >
−𝛽1 cos 𝜓1 > −𝛽1 .
We prove that the down–up equilibrium point is unstable. Let Adu be its Jacobian matrix. Straightforward computation
yields
|sI − Adu | = s4 + a1 s3 + a2 s2 + a3 s + a4 , (B3)
XIN et al. 15

where
kD 𝛼2 (kP + 𝛽1 )𝛼2 − 𝛼1 𝛽2
a1 ∶= , a2 ∶= ,
Δ Δ
kD 𝛽2 (kP + 𝛽1 )𝛽2
a3 ∶= − , a4 ∶= − .
Δ Δ

Thus, Adu is unstable due to a3 < 0 and a4 < 0. Moreover, at least one eigenvalue of Adu is in the open-left-half
plane because a1 > 0. Indeed, if the real parts of all eigenvalues of Adu are nonnegative, then a1 ≤ 0 which raises a
contradiction.
3) Under (19), as depicted in Figure 2, (11) has at least solution qe1 = 0, ±𝜌11 , ±𝜌12 . Straightforward computation of the
characteristic equation of the Jacobian matrix evaluated at (q1 , q2 , q̇ 1 , q̇ 2 ) = (𝜌12 , 2𝜋 − 𝜌12 , 0, 0) yields

s4 + a1 s3 + a2 s2 + a3 s + a4 = 0, (B4)

where
kD 𝛼2 𝜌12 𝛼1 𝛽2 + 𝛼2 𝛽1 (𝜌12 cos 𝜌12 − sin 𝜌12 )
a1 ∶= , a2 ∶= ,
𝛼1 𝛼2 − 𝛼32 cos2 𝜌12 𝜌12 (𝛼1 𝛼2 − 𝛼32 cos2 𝜌12 )
kD 𝛽2 𝛽1 𝛽2 (𝜌12 cos 𝜌12 − sin 𝜌12 )
a3 ∶= , a4 ∶= .
𝛼1 𝛼2 − 𝛼32 cos2 𝜌12 𝜌12 (𝛼1 𝛼2 − 𝛼32 cos2 𝜌12 )

Using 3𝜋∕2 < 𝜌12 < 2𝜋 and kD > 0, we obtain ai > 0 (1 ≤ i ≤ 4) and

kD2 𝛼32 𝛽22 cos2 𝜌12


D3 ∶= a1 a2 a3 − a23 − a21 a4 = > 0.
(𝛼1 𝛼2 − 𝛼32 cos2 𝜌12 )3

Thus, (q1 , q2 , q̇ 1 , q̇ 2 ) = (𝜌12 , 2𝜋 − 𝜌12 , 0, 0) is asymptotically stable. The proof of the equilibrium point (q1 , q2 , q̇ 1 , q̇ 2 ) =
(−𝜌12 , −2𝜋 + 𝜌12 , 0, 0) being stable is similar, and is omitted.
4) Using the Statement 1) of Lemma 3, we can prove that the DEP is asymptotically stable. Similar to the proof in the
Statement 2) of Lemma 3, we can also prove that the down–up, or up–down, or up–up equilibrium point is unstable by
showing a coefficient of the characteristic polynomial of the linearized model of the closed-loop system around each
of the above equilibrium points is negative; and we can show that each of their Jacobian matrices has eigenvalue(s)
in both the open right-half plane and the open left-half plane. The detail is omitted. ■

APPENDIX C. PROOF OF LEMMA 4

We design the following Lyapunov function candidate:

̂ (q, q)
V ̇ = E(q, q)
̇ + kP (1 − cos q1 ). (C1)

̂ along the trajectories of (1) and using (A3), we obtain


Taking the time-derivative of V

̂̇ = q̇ 1 (𝜏1 + kP sin q1 ).
V

Using (23) yields

̂̇ = −kD q̇ 21 ≤ 0.
V (C2)

Next, to use LaSalle’s invariance principle14 to proceed the proof, we need to define a compact set that is positively
invariant with respect to the closed-loop system consisting of (1) and (23). By treating q1 ∈ S and q2 ∈ S, we define a state
[ ]T [ ]T
̂z = ̂z1 , ̂z2 , ̂z3 , ̂z4 , ̂z5 , ̂z6 ∶= sin q1 , cos q1 , sin q2 , cos q2 , q̇ 1 , q̇ 2 , (C3)
16 XIN et al.

and using

̂ż 1 = q̇ 1 cos q1 = ̂z2̂z5 , ̂ż 2 = −q̇ 1 sin q1 = −̂z1̂z5 , (C4)

̂ż 3 = q̇ 2 cos q2 = ̂z4̂z6 , ̂ż 4 = −q̇ 2 sin q2 = −̂z3̂z6 , (C5)

we can rewrite the closed-loop system consisting of (1) and (23) as ̂ż = F
̂ (̂z), where the explicit expression of F
̂ (̂z) is
omitted. Define
̂c ∶= {̂z | V
Γ ̂ ≤ c}, (C6)

where c is a positive constant. Owing to V ̂̇ ≤ 0 in (C2) and V ̂ ≥ E − 2|kP | ≥ −2|kP | for V


̂ in (C1), V
̂ is bounded. This shows
̇ q̇ 1 , and q̇ 2 are bounded. Since ̂z1 , ̂z2 , ̂z3 , and ̂z4 are bounded, Γ
that E(q, q), ̂c is a compact set. Again owing to V̂ ≤ 0 in (C2),
̇ ̂ ̂ ̂
we can see that any solution of ̂z = F(̂z) starting in Γc remains in Γc for all t ≥ 0. Let W be the largest invariant set in

̂ ̂̇ = 0}.
̂c | V
S ∶= {̂z ∈ Γ (C7)

̂c approaches W
Using LaSalle’s invariance principle14 yields that every ̂z(t) starting in Γ ̂̇ = 0 holds
̂ as t → ∞. Since V
̂, we conclude that q1 and 𝜏1 are constants in W
identically in W ̂, which are denoted as q and 𝜏 , respectively.
∗ ∗
1 1
̂ ∗
We can show that q2 is a constant in W denoted as q2 , whose proof is almost the same as that in the proof of Lemma 2,
except that 𝜂1 in (A10) is replaced by
𝜂1 ∶= −(𝛽1 + kP ) sin q∗1 . (C8)

Finally, by putting q1 ≡ q∗1 , q2 ≡ q∗2 , and 𝜏1 ≡ −kP sin q∗1 into (1), with kP ≠ 𝛽1 , we obtain sin q∗1 = 0 and sin q∗2 =
0. Thus, the closed-loop solution (q(t), q(t)) ̇ approaches an equilibrium point of the equilibrium set Ωs defined
in (21). ■

APPENDIX D. SOME FORMULAE FOR P(S) IN ( 23)

First, putting s = z − 𝜁 into P(s) in (24) yields:

P(s) = F(z, 𝜁, kD , kP ) ∶= z4 + b1 z3 + b2 z2 + b3 z + b4 , (D1)

where by using Δ and 𝛿 defined in (26), we have

kD 𝛼2 − 4Δ𝜁
b1 ∶= ,
Δ
𝛿 + 6Δ𝜁 2 + (−3𝜁kD + kP )𝛼2
b2 ∶= ,
Δ
−2𝜁(𝛿 + 2Δ𝜁 2 + kP 𝛼2 ) + kD (3𝜁 2 𝛼2 + 𝛽2 )
b3 ∶= ,
Δ
𝜁 2 (𝛿 + Δ𝜁 2 ) + 𝛽1 𝛽2 + (−𝜁kD + kP )(𝜁 2 𝛼2 + 𝛽2 )
b4 ∶= . (D2)
Δ

Second, we find kD and kP such that F(z, 𝜁, kD , kP ) has roots ±j𝜔 with 𝜔 ≥ 0. From Re[F(j𝜔, 𝜁, kD , kP )] = 0 and
Im[F(j𝜔, 𝜁, kD , kP )] = 0, we obtain
kD = hD (𝜁, 𝜔), kP = hP (𝜁, 𝜔), (D3)

where

2𝜁(p1 𝛼2 + p2 𝛽2 ) p3 𝛼2 + p4 𝛽2
hD (𝜁, 𝜔) ∶= , hP (𝜁, 𝜔) ∶= ,
Γ(𝜁, 𝜔) Γ(𝜁, 𝜔)
XIN et al. 17

with
Γ(𝜁, 𝜔) ∶= 𝛼22 (𝜁 2 + 𝜔2 )2 + 2𝛼2 𝛽2 (𝜁 2 − 𝜔2 ) + 𝛽22 ,
p1 ∶= Δ(𝜁 2 + 𝜔2 )2 , p2 ∶= 2Δ(𝜁 2 − 𝜔2 ) + 𝛼1 𝛽2 ,
p3 ∶= (𝜁 2 + 𝜔2 )2 (Δ(𝜁 2 + 𝜔2 ) − 𝛿) + (𝜔2 − 3𝜁 2 )𝛽1 𝛽2 ,
p4 ∶= (𝜁 2 + 𝜔2 )(𝛿 + 3Δ𝜁 2 − Δ𝜔2 ) − 𝛽1 𝛽2 . (D4)

We obtain
F(z, 𝜁, hD (𝜁, 𝜔), hP (𝜁, 𝜔)) = (z2 + 𝜔2 )(z2 + d1 (𝜁, 𝜔)z + d2 (𝜁, 𝜔)), (D5)

where di (𝜁, 𝜔) (i = 1, 2) will be presented later.

APPENDIX E. PROOF OF THEOREM 1

First, for F(z, 𝜁, kD , kP ) in (D1) with 𝜁 = 𝜁a > 0, we prove that it has at least one root in the closed right-half plane of
z for any kD > 0 and kP > −𝛽1 , and thus RDPm (𝛾) ≥ −𝜁a . On the contrary, suppose that there exist k̃ D and k̃ P such that
F(z, 𝜁a , k̃ D , k̃ P ) is stable, where b̃ i (1 ≤ i ≤ 4) are its coefficients. Thus,

b̃ 1 > 0, ̃ 3 ∶= b̃ 1 b̃ 2 b̃ 3 − b̃ 23 − b̃ 21 b̃ 4 > 0.
D (E1)

Direct computation yields

̃ 3 = w0 k̃ 2P + w1 k̃ P + w2 ,
Δ3 D (E2)

where by using v1 ∶= 2𝜁a (k̃ D 𝛼2 − 2Δ𝜁a ), we have

w0 ∶= −v1 𝛼22 , w1 ∶= −2v1 (𝛼2 (𝛿 − v1 ) − 2Δ𝛽2 ),


w2 ∶= −v1 (𝛿 − v1 )2 − Δk̃ D 𝛽22 + (k̃ D 𝛼2 (𝛿 − v1 ) − (k̃ D 𝛼2 − 4Δ𝜁a )2 𝛽1 )𝛽2 .
2 2

From b1 in (D2) and b̃ 1 > 0 in (E1), we obtain

k̃ D 𝛼2 > 4Δ𝜁a , v1 > 4Δ𝜁a2 > 0. (E3)

Thus, w0 < 0. From (E2), we obtain


( )2
4w0 w2 − w21
̃ 3 = w0 k̃ P + w1
Δ3 D + , (E4)
2w0 4w0

with
4w0 w2 − w21 4v2 (k̃ D 𝛼2 − 4Δ𝜁a )2 𝛽2
=− ,
4w0 𝛼22

where v2 ∶= 𝛼2 (v1 − 𝛼1 𝛽2 ) + Δ𝛽2 . Using (E3) yields v2 > 𝛼2 (4Δ𝜁a2 − 𝛼1 𝛽2 ) + Δ𝛽2 = 0. Thus, (4w0 w2 − w21 )∕(4w0 ) < 0. From
̃ 3 < 0 holds for any k̃ P ∈ R, which contradicts to (E1).
(E4), D
Second, we prove RDPm (𝛾) = −𝜁a by showing that there exist kP and kD such that some roots of F(z, 𝜁a , kD , kP ) are in
the imaginary axis of z and other roots (if there exist) are in the open left-half plane of z. Putting 𝜁 = 𝜁a into (D5), we obtain

𝛼3 𝛽2 Ξ2
d1 (𝜁a , 𝜔) = − √ , (E5)
Δ𝛼2 (Ξ2 + 16𝛼32 𝛽22 Δ)

where Ξ ∶= 4𝛼2 Δ𝜔2 − (4𝛼1 𝛼2 − 5𝛼32 )𝛽2 .


18 XIN et al.

Under (30), if Ξ ≠ 0, that is, 𝜔2 ≠ 𝜔2a , where 𝜔a is given in (33), then d1 (𝜁a , 𝜔) < 0 holds. Thus, F(z, 𝜁, hD (𝜁, 𝜔), hP (𝜁, 𝜔))
in (D5) with 𝜁 = 𝜁a has at least one root in the open right-half plane of z. If Ξ = 0, that is, 𝜔2 = 𝜔2a ,
direct computation of (D3) with (𝜁, 𝜔) = (𝜁a , 𝜔a ) yields kD = kDm (𝛾) and kP = kPm (𝛾), where kDm (𝛾) and kPm (𝛾)
are defined in (32). With these gains, (24) has double complex conjugate roots shown in Theorem 1. Thus,
RDPm (𝛾) = −𝜁a holds. ■

APPENDIX F. PROOF OF THEOREM 2

First, we show that P(s) in (24) with kD = kDm (𝛾) in (36) and kP = kPm (𝛾) in (37) has the roots shown in Theorem 2, and
thus RDPm (𝛾) ≤ −𝜁b . To begin with, from (34) and Δ > 0 in (26), we obtain

4
𝛼1 𝛼2 < 𝛼32 < 𝛼1 𝛼2 , (F1)
5

which is useful below. From (D5), we obtain

Γ(𝜁, 𝜔)d1 (𝜁, 𝜔) = −2𝜁(e0 (𝜁 2 ) + er (𝜁 2 , 𝜔2 )𝜔2 ), (F2)

Γ(𝜁, 𝜔)d2 (𝜁, 𝜔) = f0 (𝜁 2 ) + fr (𝜁 2 , 𝜔2 )𝜔2 , (F3)

where Γ(𝜁, 𝜔) > 0 is defined in (D4), and

e0 (x) ∶= 𝛼22 Δx2 + 2𝛼2 𝛽2 Δx + (Δ − 𝛼32 )𝛽22 , (F4)

er (x, y) ∶= 𝛼2 Δ(𝛼2 (2x + y) − 2𝛽2 ), (F5)

f0 (x) ∶= 𝛼22 Δx3 + 3𝛼2 𝛽2 Δx2 + 3𝛽22 (Δ − 𝛼32 )x + 𝛼1 𝛽23 , (F6)

fr (x, y) ∶= 𝛼2 (𝛼2 x + 𝛽2 )Δy + 2𝛼22 Δx2 − (2Δ + 𝛼32 )𝛽22 . (F7)

Note that e0 (x) in (F4) and f 0 (x) in (F6) satisfy

df0 (x)
= 3e0 (x). (F8)
dx

Solving e0 (x) = 0 with x ≥ 0 yields x = 𝜁c2 , where



(𝛼3 − Δ)𝛽2
𝜁c2 ∶= √ >0 (F9)
𝛼2 Δ

since (F1) guarantees 𝛼3 > 2 Δ. From (F8), we know that f 0 (x) is strictly decreasing for 0 ≤ x < 𝜁c2 due to e0 (x) < 0, and
f 0 (x) is strictly increasing for x > 𝜁c2 due to e0 (x) > 0. Since f0 (0) = 𝛼1 𝛽23 > 0,

√ √
f0 (𝜁c2 ) = −2𝛼32 𝛽23 (𝛼3 − 2 Δ)∕(𝛼2 Δ) < 0

(since (F1) guarantees 𝛼3 > 2 Δ), and f 0 (+ ∞) = +∞ hold, as shown in Figure F1, f 0 (x) has two distinct pos-
itive real roots, which are denoted as 𝜁b2 and 𝜁g2 with 𝜁b2 < 𝜁c2 < 𝜁g2 . Using the formulae of the roots of a
cubic polynomial in Reference 21 (p. 74), we can obtain 𝜁b > 0 in (35) with 𝜁b2 being the second biggest real
root of f 0 (x).
XIN et al. 19

F I G U R E F1 Plot of f 0 (x) in (F6) with x ≥ 0

Next, putting (𝜁, 𝜔) = (𝜁b , 0) into (D3), (D4), (F2), (F3), and (D5) and using f0 (𝜁b2 ) = 0, we obtain

hD (𝜁b , 0) = kDm (𝛾), hP (𝜁b , 0) = kPm (𝛾), (F10)


−2𝜁b e0 (𝜁b2 )
d1 (𝜁b , 0) = , d2 (𝜁b , 0) = 0, (F11)
(𝛼2 𝜁b2 + 𝛽2 )2
F(z, 𝜁b , kDm (𝛾), kPm (𝛾)) = z3 (z + d1 (𝜁b , 0)), (F12)

where kDm (𝛾) and kPm (𝛾) are in (36) and (37), respectively. Due to 𝜁b2 < 𝜁c2 as shown in Figure F1, we obtain e0 (𝜁b2 ) < 0.
Using (F11) and (F4) yields 𝜁br = d1 (𝜁b , 0) > 0 in (39). Thus, F(z, 𝜁b , kDm (𝛾), kPm (𝛾)) in (F12) has roots 0, 0, 0, and −𝜁br ; and
P(s) with s = z − 𝜁b has the roots shown in Theorem 2.
Second, we prove RDPm (𝛾) = −𝜁b . On the contrary, assume RDPm (𝛾) < −𝜁b . Since the roots of P(s) in (24) are
continuous functions of kD and kP , starting from the roots shown in Theorem 2, we can reach the roots corre-
sponding to the above assumed RDPm (𝛾) by varying kD and kP continuously. Thus, for sufficient small 𝜀1 > 0 and
𝜀2 > 0, there exist kD𝜀 and kP𝜀 such that −(𝜁b + 𝜀1 ) ± j𝜀2 are roots of P(s).
We take s = z − (𝜁b + 𝜀1 ). We can obtain kD𝜀 and kP𝜀 by using (D3) with 𝜁 = 𝜁b + 𝜀1 . Using (D5) with 𝜁 = 𝜁b + 𝜀1 yields

F(z, 𝜁b + 𝜀1 , hD (𝜁b + 𝜀1 , 𝜀2 ), hP (𝜁b + 𝜀1 , 𝜀2 )) = (z2 + 𝜀22 )(z2 + d1 (𝜁b + 𝜀1 , 𝜀2 )z + d2 (𝜁b + 𝜀1 , 𝜀2 )), (F13)

and di (𝜁b + 𝜀1 , 𝜀2 ) ≥ 0 (i = 1, 2) since the RDP of P(s) in (24) with obtained kD𝜀 and kP𝜀 is −(𝜁b + 𝜀1 ).
Below, we prove d2 (𝜁b + 𝜀1 , 𝜀2 ) < 0 for anysufficient small 𝜀1 > 0 and 𝜀2 > 0; this raises a contradiction, and thus,
RDPm (𝛾) = −𝜁b holds. From (F3) and Γ(𝜁b + 𝜀1 , 0) > 0 due to (D4), we know that the sign of d2 (𝜁b + 𝜀1 , 𝜀2 ) is the same as
that of f0 ((𝜁b + 𝜀1 )2 ) + fr ((𝜁b + 𝜀1 )2 , 𝜀22 )𝜀22 . Since the Taylor expansions of f0 (𝜁 2 ) at 𝜁 = 𝜁b and fr (𝜁 2 , 𝜔2 ) at (𝜁, 𝜔) = (𝜁b , 0) are

f0 (𝜁 2 ) = f0 (𝜁b2 ) + 6𝜁b e0 (𝜁b2 )(𝜁 − 𝜁b ) + ((𝜁 − 𝜁b )2 ),


fr (𝜁 2 , 𝜔2 ) = fr (𝜁b2 , 0) + (𝜁 − 𝜁b ) + (𝜔2 ),

respectively, where each (𝜅 i ) (i = 1, 2) denotes the term grows at the order of 𝜅 i , we obtain

f0 ((𝜁b + 𝜀1 )2 ) + fr ((𝜁b + 𝜀1 )2 , 𝜀22 )𝜀22 = f0 (𝜁b2 ) + 6𝜁b e0 (𝜁b2 )𝜀1 + fr (𝜁b2 , 0)𝜀22 + (𝜀21 ) + ((𝜀1 ) + (𝜀22 ))𝜀22 . (F14)

Since f0 (𝜁b2 ) = 0 and e0 (𝜁b2 ) < 0 hold, if we can show that under (F1) the following inequality holds:

fr (𝜁b2 , 0) = 2𝛼22 Δ𝜁b4 − (2Δ + 𝛼32 )𝛽22 < 0, (F15)

then for any sufficient small 𝜀1 > 0 and 𝜀2 > 0, we can neglect the higher order terms in (F14) for sufficient small 𝜀1 > 0
and 𝜀2 > 0 to obtain
f0 ((𝜁b + 𝜀1 )2 ) + fr ((𝜁b + 𝜀1 )2 , 𝜀22 )𝜀22 < 0 (F16)

Due to Φ in 𝜁b in (35), it is difficult to use 𝜁b to show directly (F15). Thus, below we prove fr (𝜁b2 , 0) < 0 for (𝛼1 , 𝛼2 , 𝛼3 )
belongs to the following two sets:
{ }
4 16
S1 ∶= (𝛼1 , 𝛼2 , 𝛼3 )| 𝛼1 𝛼2 < 𝛼32 < 𝛼1 𝛼2 , (F17)
5 17
S2 ∶= { (𝛼1 , 𝛼2 , 𝛼3 ) | 0.84𝛼1 𝛼2 < 𝛼32 < 𝛼1 𝛼2 }, (F18)
20 XIN et al.

whose union contains any (𝛼1 , 𝛼2 , 𝛼3 ) satisfying (F1).


For any (𝛼1 , 𝛼2 , 𝛼3 ) in set S1 and 𝜁c2 in (F9), if we can prove fr (𝜁c2 , 0) < 0, then fr (𝜁b2 , 0) < fr (𝜁c2 , 0) < 0 because 𝜁b < 𝜁c
holds (refer to Figure F1) and fr (𝜁 2 , 0) (refer to (F15)) is strictly increasing for 𝜁 ≥ 0. To this end, using

fr (𝜁c2 , 0) = 𝛼3 𝛽22 (𝛼3 − 4 Δ)

yields that fr (𝜁c2 , 0) < 0, that is, 𝛼3 < 4 Δ is equivalent to 𝛼32 < (16∕17)𝛼1 𝛼2 . This with (F1) shows fr (𝜁c2 , 0) < 0 for any
(𝛼1 , 𝛼2 , 𝛼3 ) in set S1 .
For any (𝛼1 , 𝛼2 , 𝛼3 ) in set S2 , as preparation, we decompose f0 (𝜁 2 ) using (F6) as

f0 (𝜁 2 ) = Q(𝜁 2 )fr (𝜁 2 , 0) + R(𝜁 2 ), (F19)

where

Q(𝜁 2 ) ∶= (𝛼2 𝜁 2 + 3𝛽2 )∕(2𝛼2 ),


((8𝛼1 𝛼2 − 3𝛼32 )𝛽2 − 𝛼2 (13𝛼32 − 8𝛼1 𝛼2 )𝜁 2 )𝛽22
R(𝜁 2 ) ∶= .
2𝛼2

Thus, fr (𝜁b2 , 0) < 0 is equivalent to R(𝜁b2 ) > 0 due to f0 (𝜁b2 ) = 0 and Q(𝜁b2 ) > 0. Since 8𝛼1 𝛼2 − 3𝛼32 > 0 and 13𝛼32 − 8𝛼1 𝛼2 >
0 hold due to (F1), by solving R(𝜁 2 ) = 0, we obtain R(𝜁d2 ) = 0 with

(8𝛼1 𝛼2 − 3𝛼32 )𝛽2


𝜁d2 ∶= > 0. (F20)
𝛼2 (13𝛼32 − 8𝛼1 𝛼2 )

Below if we can show f0 (𝜁d2 ) < 0, then we obtain 𝜁b2 < 𝜁d2 < 𝜁g2 from the f 0 (x) depicted in Figure F1, and
R(𝜁b2 )> R(𝜁d2 ) = 0 because R(𝜁 2 ) is strictly decreasing for 𝜁 ≥ 0. To this end, from

̂
2𝛼12 𝛼2 𝛼32 (9𝛼32 − 4𝛼1 𝛼2 )𝛽23 Δ
f0 (𝜁d2 ) = , (F21)
(13𝛼32 − 8𝛼1 𝛼2 )3

where Δ ̂ ∶= 151v2 − 432v + 256 with v ∶= 𝛼 2 ∕(𝛼1 𝛼2 ). Using 13𝛼 2 − 8𝛼1 𝛼2 > 0 and 9𝛼 2 − 4𝛼1 𝛼2 > 0 due to (F1) yields that
3 3 3
f0 (𝜁d ) < 0 holds if and only if Δ
2 ̂ < 0 holds, whose solution is 0.84 < v < 2.02. This with (F1) shows f0 (𝜁 2 ) < 0 for any
d
(𝛼1 , 𝛼2 , 𝛼3 ) in set S2 . In summary, we complete the proof of (F15) for any (𝛼1 , 𝛼2 , 𝛼3 ) satisfying (F1). ■

You might also like