IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO.

2, APRIL 2015 331

Robust Swing-Up and Balancing Control Using a

Nonlinear Disturbance Observer for the Pendubot
System With Dynamic Friction
Myunghwan Eom and Dongkyoung Chwa

Abstract—This paper proposes the robust swing-up and bal- the swing-up motion from the stable initial position. However,
ancing control method using a nonlinear disturbance observer for even though the actuated variables are controlled to meet the
the pendubot system in the presence of dynamic friction. First, purpose, it is difficult to guarantee the stability of the unactu-
the uncertainties in the pendubot system are verified through the
experiment to take the form of a continuous and differentiable fric- ated variables [12]. As a result, various research works have
tion model. Then, the uncertainties present in both actuated and been studied on the pendubot system as one of the typical un-
unactuated dynamics are shown to have the coupled form between deractuated mechanical systems, since it is much challenging
them through the analysis, such that they can be compensated to guarantee the robustness against uncertainties due to its quite
by estimating only one of them. In order to compensate for the complicated characteristics compared with other underactuated
uncertainties, the robust swing-up control is proposed. Then, the systems [13]–[15].
limit-cycle phenomenon, which appears around the unstable equi-
librium point due to the uncertainties is compensated in the robust The swing-up and balancing control of the pendubot system
balancing control using the nonlinear disturbance observer, and has been studied in several previous works using the nonlinear
the disturbance estimation errors are shown through the stability control method. First, the partial feedback linearization control
analysis to converge to zero. Finally, the performance and the va- and, then, LQR or pole-placement method are applied in [1].
lidity of the proposed method are verified through both simulation In [11], the swing-up control method through the energy-based
and experimental results. Lyapunov stability analysis is proposed using the passive prop-
Index Terms—Disturbance observer, dynamic friction, limit- erty of the pendubot. In [16], a robust controller against the
cycle phenomenon, pendubot system, robust swing-up and bal- uncertainties and disturbances in the system is proposed us-
ancing control. ing the adaptive sliding-mode control method. In addition, [17]
proposed the fuzzy PI + PD control method for the underactu-
I. INTRODUCTION ated robot in the actual experimental environment. These works,
URING the past decades, the underactuated mechanical however, did not consider the influence of the friction, backlash,
D systems have been studied in many application system
areas. Underactuated systems are the systems, which have the
and modeling error that exist in the actual system, and further-
more, only the simulation results are provided. These uncertain
number of control inputs less than the degree of freedom for factors can result in the unexpected behaviors in the actual sys-
controlled actuators [1]. Control issues for these types of sys- tems, and can degrade the performance of the pendubot control
tems widely exist in several industries, such as mobile robots, system significantly. Consequently, it is necessary to compen-
ships, cranes, pendulum systems, vertical takeoff and landing sate for the disturbances and the uncertainties considering the
aircraft, etc. (see [2]–[11] and the references therein). Since the coupled phenomenon between actuated and unactuated dynam-
dynamic coupling between the unactuated state variables and ics to achieve the effective control of the pendubot system.
the actuated ones should be considered in the design of the Friction is one of the critical ones among the main factors of
controller for the underactuated systems, the control of the un- the disturbance and the uncertainties of the system, in that all the
deractuated systems becomes much challenging. The pendubot moving systems are always affected by friction. In particular,
system is a typical underactuated mechanical system, where the friction with highly nonlinear characteristics imposes con-
two rotating links are connected, whereas the first link operates siderable influence on the system response and, thus, has been a
by the external torque, and the second link is passive without popular research topic for a long time [18]–[26]. The nonlinear
the external torque. The control objective is to stabilize both of characteristics of the friction makes the design of controller con-
these two links toward the unstable vertical upright position by siderably difficult in maintaining the desired performance of the
system; thus, being considered as the uncertainties to be elim-
Manuscript received February 5, 2015; revised June 2, 2014; accepted January inated. Generally, the viscous friction, which is proportional
30, 2015. Date of publication March 6, 2015; date of current version April 2, to the relative velocity of the moving object and the contact
2015. This paper was recommended for publication by Associate Editor T. surface, and the Coulomb friction have been recognized as im-
Hamel and Editor B. J. Nelson upon evaluation of the reviewers’ comments.
This work was supported by the National Research Foundation of Korea under a
portant parts of the nonlinear friction. The friction model has
grant supported by the Korea government (MSIP) (2014R1A2A1A11053153). not only these properties, but several additional static character-
The authors are with the Department of Electrical and Computer Engineering, istics as well, such as static friction, break-away force, steady
Ajou University, Suwon 443-749, Korea (e-mail: dkchwa@[Link]). velocity friction, Stribeck effect, etc. However, the static friction
Color versions of one or more of the figures in this paper are available online
at [Link] model does not include the dynamic properties of the friction,
Digital Object Identifier 10.1109/TRO.2015.2402512 such as varying break-away force, frictional lag, presliding dis-

1552-3098 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] standards/publications/rights/[Link] for more information.
332 IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO. 2, APRIL 2015

placement, stick-slip motion, etc. All these dynamic character- well) through the simulation and experimental results. Then,
istics can result in the significantly degraded performance in we develop the robust swing-up control law of the pendubot
the tracking or precise position control. Among the various dy- system with the uncertainties. The robust swing-up controller
namic friction models, the LuGre model [18] describes well is designed considering the uncertainties through the modifi-
the friction characteristics of the moving object in low velocity, cation of the conventional partial feedback linearization [11].
and can also include both static and dynamic friction models. The uncertainties of the overall system exist in each of the dy-
Nevertheless, a LuGre model and other dynamic friction models namic equation of the first link and the second link. Through the
include partially continuous or discontinuous functions. derivation of the coupled form between them, it is shown that
The LuGre model-based adaptive friction observer is em- both uncertainties can be compensated, even when just one of
ployed to compensate for the friction in [19]. Since friction them can be estimated. Based on this coupled property between
is a function that changes with the variation of temperature them, the robust balancing control using the nonlinear distur-
and position on the contact surface, the method to estimate and bance observer is designed to compensate for the uncertainties
compensate for the parameters of the friction model is proposed. of the pendubot system, and its performance is verified through
However, the parameters in the friction model are highly non- the simulation and experimental results.
linear, and the estimation of them requires to know the internal The main contributions of the proposed robust swing-up and
friction variable, which is practically impossible to measure. balancing control can be summarized as follows. First, the non-
As a consequence, it is assumed that static parameters are time linear dynamic friction model has been verified to include the
varying, and dynamic parameters are time invariant. As these properties of both static and dynamic friction through actual
assumptions are not valid in practice, adaptive friction observers experiments. Second, the proposed robust swing-up control
may not be appropriate for the actual systems. In [20], the fric- scheme can maintain the expected performance compared with
tion in Furuta pendulum is compensated using the LuGre friction the previous swing-up control methods in the pendubot sys-
observer. Since the previous friction compensation problems are tem, which is affected by both uncertainties and friction force.
limited to the simple servo and mechanical systems [21], [22], Third, unlike the previous nonlinear disturbance observers, the
the performance of the control system deteriorates, even with proposed robust nonlinear disturbance observer for more com-
the compensation of the friction. The information of the struc- plicated underactuated system, such as pendubot can estimate
ture in the friction model is necessary in the studies, where the time-varying disturbance with the estimation error converg-
nonlinear or adaptive observers are designed based on the fric- ing asymptotically to zero with time. In addition, the robust
tion model. Thus, the performance of the disturbance observers nonlinear observer is developed to cope with system uncertain-
can no longer be guaranteed if the parameters of the model are ties, dynamic friction, and external disturbance in more practical
not accurate, or if the friction model to be estimated is different systems. Finally, the stability of the pendubot system using the
from the actual one. In addition, even if the accurate mathe- proposed robust observer-based balancing control has been an-
matical friction model can be obtained, it is nearly impossible alyzed to show that it can be maintained in the region around
to perfectly realize the nonlinear characteristics of the actual balancing position.
friction. This paper is organized in the following way. First, the dy-
On the other hand, the use of the nonlinear disturbance ob- namic equations of the pendubot system, which include the
server, which is not based on the exact structure of the friction friction model as the system uncertainties are described in
in contrast with the model-based observers can be much bene- Section II. In Section III, a robust swing-up control method is
ficial. In [23], Lyapunov-based nonlinear disturbance observer proposed for the pendubot system with friction. Then, a nonlin-
is developed for the coefficient of the unknown input applied ear disturbance observer is designed to estimate and compensate
to two-link robot arms. In [24], the sliding-mode-based fric- for the uncertainties to achieve the satisfactory balancing con-
tion observer is designed, and utilized in the adaptive controller trol in Section IV. In Section V, the feasibility and effectiveness
such that the performance can be improved by estimating un- of the proposed method is verified through the simulation and
known parameters of the system and friction model. In [25], the experimental results. Finally, Section VI presents the con-
the disturbance observer can compensate for not only the con- clusion.
stant but the ramp disturbances as well. However, the type of
disturbances is generally restricted to the constant disturbance II. DYNAMICS OF THE PENDUBOT SYSTEM WITH FRICTION
or a simple form in these disturbance observers. In particular,
if the parameters of the disturbance model is time varying, it is A. Dynamics of the Pendubot System
impossible to estimate them exactly. In addition, the application In this section, consider a pendubot system shown in Fig. 1,
of the disturbance observers is deficient in the case of the un- where the actuator exists for the actuated link and does not exist
deractuated system, which has a highly nonlinear and coupled for the unactuated link. The dynamics of the pendubot can be
characteristics between the actuated and unactuated dynamics. derived from the Euler–Lagrange as follows [27]:
This paper proposes a robust swing-up and balancing control  
method for the pendubot system in the presence of the friction. d ∂L ∂L
− =u (1)
We first verify that the friction model in the pendubot system dt ∂ q̇ ∂q
can be a particular form of continuous and differential function
where q = [qa , qu ]T ∈ 2 and q̇ = [q̇a , q̇u ]T ∈ 2 are vectors
(which is verified to include not only the static characteristics
of generalized coordinates and velocities, respectively; qa and
but the dynamic characteristics of the friction model in [26] as
qu are actuated and unactuated variables; u = [τ, 0]T ∈ 2 is a
EOM AND CHWA: ROBUST SWING-UP AND BALANCING CONTROL USING NONLINEAR DISTURBANCE OBSERVER FOR THE PENDUBOT 333

sin qu (q̇a + q̇u )2

− (θ1 + θ3 cos qu )θ3 sin qu q̇a2
+ (θ2 + θ3 cos qu )(θ4 cos qa − τ )
− (θ1 + θ3 cos qu )θ5 cos(qa + qu )]. (4b)

B. Equilibrium Configuration
The pendubot system has an equilibrium configuration when
a control input and state variables satisfy the particular condi-
tion for which the pendubot is at rest, i.e, q̇ = 0. Solving the
equation of the pendubot system in (3) under this condition, the
equilibrium points are given as
θ4 cos qa + θ5 cos(qa + qu ) = τ
θ5 cos(qa + qu ) = 0. (5)
Fig. 1. Dynamics of the pendubot system.
As long as the torque input τ is constrained to satisfy τ ≤ θ4 ,
control input variable; the function the solution of (5) for the equilibrium states can be obtained as
 
1 ˙T τ
L(q, q̇) = q M (q)q̇ − V (q) (2) qa = arccos
θ4
2
is a Lagrangian; M (q) is a symmetric positive-definite inertia π
qu = n − q a (6)
matrix; and V (q) is a potential energy of the system. If the 2
friction present in the system is not considered, then the nominal where n = 1, 3, 5, . . .. This means that the pendubot system will
pendubot dynamics in (1) can be rewritten in matrix form as balance at the state (qa , qu , 0, 0), if we apply a constant torque
τ . In this paper, we are interested in the natural equilibria of the
M (q)q̈ + C(q, q̇)q̇ + G(q) = u (3) pendubot when τ = 0. As a result, from the solutions for (6), we
where can obtain the following four equilibrium configurations [28].
  1) (qa , qu , q̇a , q̇u ) = ( π2 , 0, 0, 0): (both actuated and unactu-
θ1 + θ2 + 2θ3 cos qu θ2 + θ3 cos qu ated links are in their top positions).
M (q) =
θ2 + θ3 cos qu θ2 2) (qa , qu , q̇a , q̇u ) = (− π2 , 0, 0, 0): (both actuated and unac-
  tuated links are in their bottom positions).
−θ3 sin qu q˙u −θ3 sin qu q˙u − θ3 sin qu q˙a 3) (qa , qu , q̇a , q̇u ) = ( π2 , π, 0, 0): (actuated link is in its top
C(q, q̇) = position, and unactuated link is in its bottom position).
θ3 sin qu q˙a 0
4) (qa , qu , q̇a , q̇u ) = (− π2 , π, 0, 0): (actuated link is in its
 
θ4 cos qa + θ5 cos(qa + qu ) bottom position, and unactuated link is in its top posi-
G(q) = . tion).
θ5 cos(qa + qu ) Among the above four equilibrium states described above,
(qa , qu , q̇a , q̇u ) = (− π2 , 0, 0, 0) is the stable equilibrium point,
Note that C(q, q̇)q̇ is the Coriolis and centripetal loading vec-
and the other three equilibrium points are unstable. The pur-
tor, and G(q) is the gravitational loading vector. The param-
2 pose of the swing-up control and the balancing control is
eters θ1 = m1 lc1 + m2 l12 + I1 , θ2 = m2 lc2
2
+ I2 , θ3 = m2 l1 lc2 ,
to make the pendubot system swing from the initial sta-
θ4 = (m1 lc1 + m2 l1 )g, and θ5 = m2 lc2 g are dependent on the
ble state (qa , qu , q̇a , q̇u ) = (− π2 , 0, 0, 0) to the unstable state
following variables: l1 is the length of the first link, lc1 (respec-
(qa , qu , q̇a , q̇u ) = ( π2 , 0, 0, 0) and, then, to maintain the bal-
tively, lc2 ) is the distance to the center of the mass of the first
anced state, which will be designed in Sections III and IV,
(respectively, second) link, m1 (respectively, m2 ) is the mass of
respectively.
the first (resp., second) link, I1 (respectively, I2 ) is the moment
of inertia of the first (resp., second) link, and g is acceleration
due to gravity. Equation (3), if expressed as the angular accel- C. Modeling of Nonlinear Uncertainties—Friction
eration for each link, can be derived as In the actual mechanical systems, friction causes much in-
1 fluence on the precise position control, which is the same with
q̈a = [θ2 θ3 sin qu (q̇a + q̇u )2 the considered pendubot system and, thus, friction can be as-
θ1 θ2 − θ32 cos2 qu
sumed to be uncertainties. Among the conventional research
+ θ32 cos qu sin qu q̇a2 + θ3 θ5 cos qu works, [18] proposed a LuGre model, which is a dynamic fric-
tion model in such a way that it can accurately represent the
cos(qa + qu ) − θ1 θ2 cos qa + θ2 τ ] (4a) actual friction phenomenon. This friction model describes the
1 characteristics of the friction, such as Stribeck effect, hysteretic
q̈u = [−θ3 (θ2 + θ3 cos qu )
θ1 θ2 − θ32 cos2 qu behavior, stick-slip limit cycling, presliding displacement, and
334 IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO. 2, APRIL 2015

rising static friction. On the other hand, since this model includes ifying the method in [11], where the uncertainties Ωa and Ωu
discontinuous and piecewise continuous terms, it becomes diffi- were ignored in the system. Thus, the stability of the pendubot
cult to design analytic and rigorous controller for the nonlinear system may not be maintained depending on the magnitudes of
system model in (4a) and (4b). Therefore, we employed a non- the uncertainties, as can be seen through the simulations. To this
linear friction model in [26], which is given as end, we need to obtain the control input using the properties of
the passivity of the energy of the pendubot system.
f (q̇) = γ1 (tanh γ2 q̇ − tanh γ3 q̇) + γ4 tanh γ5 q̇ + γ6 q̇ (7)
First, a symmetric matrix M (q) can be shown using (3) to
where γi ∈  ∀i = 1, 2, . . . , 6 are unknown positive con- satisfy the positive definiteness for all q, i.e.,
stants. The friction model in (7) has the following M11 = θ1 + θ2 + 2θ3 cos qu
properties [26].
2
1) It is continuously differentiable. = m1 lc1 + m2 l12 + I1 + m2 lc2
2

2) It is symmetric at the origin. + I2 + 2m2 l1 lc2 cosq2

3) The static coefficient of friction can be approximated in
terms of γ1 + γ4 . ≥ m1 lc1
2
+ m2 l12 + I1 + m2 lc2
2

4) The term tanh γ2 q̇ − tanh γ3 q̇ captures the Stribeck ef-

+ I2 − 2m2 l1 lc2
fect.
5) A viscous dissipation is presented as γ6 q̇. ≥ m1 lc1
2
+ I1 + I2 + m2 (l1 − lc2
2 2
) > 0 (11)
6) The Coulomb friction is modeled as γ4 tanh γ5 q̇.
det(M (q)) = θ1 θ2 − θ32 cos2q u
Therefore, considering the presence of friction in the pen-
2
dubot system (the validity of which will be verified in Section V), = (m1 lc1 + I1 )(m2 l22 + I2 )
the nominal dynamics in (3) should be replaced by the dynamics 2
including the friction model in (7) as + m2 l11 I2 + m22 l12 lc2
2
sin2 q2 > 0. (12)
Then, a skew-symmetric matrix property can be derived again
M (q)q̈ + C(q, q̇)q̇ + G(q) + f (q̇) = u (8)
from (3) as
where f (q̇) := (fa (q̇), 0) ∈ 2 is the vector of the actuated link  
0 1
in the presence of friction. Then, using the Legendre transfor- Ṁ (q) − 2C(q, q̇) = θ3 sin q2 (2q̇a + q̇u ) (13)
mation in [29], the pendubot dynamics in (8) can be expressed −1 0
as
which leads to
q̈ = Λ(q, q̇, τ ) + Ω(q, q̇) (9)
z T (Ṁ (q) − 2C(q, q̇))z = 0 ∀z. (14)
or, equivalently
To show the passivity of the pendubot, the total energy of the
     
q̈a Fa (q, q̇) + Ga (q)τ Ωa (q, q̇) pendubot is chosen as
= + (10) 1 T
q̈u Fu (q, q̇) + Gu (q)τ Ωu (q, q̇) E= q̇ M (q)q̇ + θ4 sin qa + θ5 sin(qa + qu ). (15)
2
where Λ := [Fa + Ga τ, Fu + Gu τ ]T , |Ωa (q, q̇)| < Ω̄a and Its time derivative can be derived from (8), (14), and (15) as
|Ωu (q, q̇)| < Ω̄u for known constants Ω̄a and Ω̄u , respectively,
1
and each nonlinear functions are defined from (8) as Ė = q̇ T Ṁ (q)q̈ + q̇ T Ṁ (q)q̇ + q̇ T G(q)
2
Fa = Π/Ξ, Fu = Φ/Ξ, Ga = Θ/Ξ, Gu = Ψ/Ξ 1
2
= q̇ T (Bτ −C(q, q̇)q̇−G(q)−f (q̇))+ q̇ Ṁ (q)q̇ + q̇ T G(q)
Π = θ2 θ3 sin qu (q̇a + q̇u ) + θ32 cos qu sin qu q̇a2 2
+ θ3 θ5 cos qu cos(qa + qu ) − θ2 θ4 cos qa = q̇ T (Bτ − f (q̇))

Φ = − θ3 (θ2 + θ3 cos qu ) sin qu (q̇a + q̇u )2 = q̇a (τ − fa ). (16)

− (θ1 + θ3 cos qu )θ3 sin qu q̇a2 + (θ2 + θ3 cos qu ) By integrating both sides of (16), we obtain
 t
θ4 cos qa − (θ1 + θ3 cos qu )θ5 cos(qa + qu ) q̇a (τ − fa )dt = E(t) − E(t0 ). (17)
t0
Θ = θ2 , Ψ = −(θ2 + θ3 cos qu ), Ξ = θ1 θ2 − θ32 cos2 qu
Therefore, the pendubot system is passive from the input with
Ωa = −θ2 fa (q̇)/Ξ, Ωu = (θ2 + θ3 cos qu )fa (q̇)/Ξ. the friction τ − fa and q̇a . To achieve the control objective,
Here, note that Ga > 0 since (θ1 θ2 − θ32 cos2 qu ) > 0 for all qu . the energy of the total system for the equilibrium point is set
In the sequel, Ω, Ωa , and Ωu will be used instead of Ω(q, q̇), to be E(q, q̇) = ETop ( π2 , 0, 0, 0) = θ4 + θ5 . To make the link
Ωa (q, q̇), and Ωu (q, q̇) for the notational brevity. reach the target state (qa , qu , q̇a , q̇u ) = ( π2 , 0, 0, 0) (i.e., Ẽ 
E − ETOP becomes zero), the robust swing-up controller and
III. ROBUST SWING-UP CONTROL TOWARD THE TOP POSITION parameter adaptation law with dead zone can be designed as
In this section, we will design the robust swing-up control −KD Fa − KP q̃a − q̇a − D̂sat(q̇a /dw )
τ = (18a)
method based on the energy of the pendubot system by mod- KE Ẽ + KD Ga
EOM AND CHWA: ROBUST SWING-UP AND BALANCING CONTROL USING NONLINEAR DISTURBANCE OBSERVER FOR THE PENDUBOT 335

˙ ˙
D̂ = δ(|ew | + dw ) (18b) D̃D̃
≤ − q̇a2 + D̃(|ew | + dw ) + . (21)
δ
where KE , KD , KP are positive constants and should sat-
isfy K By applying the parameter adaptation law in (18b), (21) becomes
K E > 2θ1 (θ4 + θ5 ); D and D̂ are the upper bound of
D

| − q̇a KE Ẽfa + q̇a KD Ωa | and its estimate, respectively; δ V̇ ≤ − q̇a2 . (22)

and dw are positive constants determining the rate of adapta- Integrating both sides of the above equation shows that
tion and the width of the dead zone, respectively; and ew :=  t
q̇a − dw sat(q̇a /dw ) is defined as the dead-zoned error. To re-
q̇a2 dτ ≤ V (0) − V (t) ≤ V (0) < ∞ ∀t ≥ 0. (23)
duce the chattering phenomenon, the signum function needs to 0
be replaced by the saturation function such as D̂sat(q̇a /dw ) in
(18a). The introduction of the dead zone to the parameter adap- Therefore, q̇a ∈ L2 ∩ L∞ and D̃ ∈ L∞ , which implies that
tation law as in (18b) is necessary in the sense that just using Fa , Ga , τ ∈ L∞ and, thus, q̈a ∈ L∞ . Therefore, from Barbalat’s
q̇a instead of ew in (18b) results in the constant increase of D̂ lemma q̇a → 0 as t → ∞. It can be shown as in [11] that the
since q̇a may not be identically zero all the time. Thus, this solution of the closed-loop system with the control law in (18a)
problem could be solved using the proposed method involving converges to the invariant set given by the homoclinic orbit
the parameter adaptation law with dead zone as in the following with (qa , q̇a ) = ( π2 , 0), and the rest of the stability analysis fol-
theorem. lows in the same procedure in [11], which is therefore omitted
Theorem 1 (Robust Swing-Up Control). Consider the pen- here. 
dubot system including the friction in (8). If the robust swing- Remark 1: In the case of the swing-up control, the time for
up control law τ in (18a) and the parameter adaptation law the estimation of friction is so short and, thus, it is of little use.
in (18b) are employed for the initial conditions satisfying Therefore, a robust controller is designed as in (18a) and (18b).
|Ẽ| ≤ min(2θ4 , 2θ5 ), then the solution of the pendubot sys- On the other hand, in the case of the balancing controller, a
tem with the friction model in (8) converges to the invariant nonlinear disturbance observer will be designed and used for
set given by the homoclinic orbit 12 θ2 q̇u2 = θ5 (1 − cos qu ) with the friction compensation in the next section.
(qa , q̇a ) = ( π2 , 0).
Proof: The Lyapunov candidate function can be defined to IV. ROBUST BALANCING CONTROL USING THE
make Ẽ, q̃a , and q̇a converge to 0 for q̃a := (qa − π2 ) as NONLINEAR DISTURBANCE OBSERVER
A. Coupled Uncertainties
KE KD 2 KP 2 D̃2
V (q, q̇) = Ẽ(q, q̇)2 + q̇ + q̃ + (19) The pendubot system consists of two separate systems as
2 2 a 2 a 2δ
in (10); each actuated and unactuated subsystem includes the
where D̃ := D − D̂ denotes the parameter adaptation error. uncertainties. When these uncertainties can be expressed in a
Then, the time derivative of (19) can be given as coupled form as in the following lemma, the estimation of only
one of them can make it possible to compensate for both uncer-
˙
D̃D̃ tainties of the overall system at the same time.
V̇ = KE Ẽ Ė + KD q̇a q̈a + KP q̃a q̇a + Lemma 1 (Coupled Uncertainties): Consider the pendubot
δ
˙ system with friction in (8). For each servo system, unknown
D̃D̃ uncertainties Ωa and Ωu are coupled and satisfy
= KE Ẽ q̇a (τ − fa ) + KD q̇a q̈a + KP q̃a q̇a +
δ
Ωa = −ΓΩu (24)
= q̇a {KE Ẽ(τ − fa ) + KD (Fa + Ga τ + Ωa )
where Γ = θ2 /(θ2 + θ3 cos qu ).
˙
D̃D̃ Proof: First, we can assume Ωa and Ωu to be uncertain in
+ KP q̃a } + regards to each angular acceleration in (10), i.e., Ωa := −Δq̈a
δ
and Ωu := −Δq̈u . Then, (10) can be rewritten in the following
= q̇a {τ (KE Ẽ + KD Ga ) − KE Ẽfa + KD Fa form including uncertainties:
˙
D̃D̃
+ KD Ωa + KP q̃a } + . (20) q̈a + Δq̈a = Fa (q, q̇) + Ga (q)τ
δ
q̈u + Δq̈u = Fu (q, q̇) + Gu (q)τ. (25)
When |q̇a | > dw , it follows that sat(q̇a /dw ) = sgn(ew ) and,
Thus, by substituting q̈a − Ωa and q̈u − Ωu for q̈a and q̈u into
thus, q̇a = ew + dw sgn(ew ). Furthermore, when |q̇a | ≤ dw , it
the second row of (8), we can have
can be concluded that ew becomes zero. Substitution of the
control input in (18a) into (20) gives (θ2 + θ3 cos qu )(q̈a − Ωa ) + θ2 (q̈u − Ωu )
˙
D̃D̃ + θ3 sin qu q̇a2 + θ5 cos(qa + qu ) = 0. (26)
V̇ = q̇a {−q̇a − D̂sgn(ew ) − KE Ẽfa + KD Ωa } +
δ Therefore, this equation can be simplified by referring to the
second row of (8) as
≤− q̇a2 − {ew + dw sgn(ew )} · D̂sgn(ew )
˙ (θ2 + θ3 cos qu )Ωa + θ2 Ωu = 0 (27)
D̃D̃
+ D{ew + dw sgn(ew )} + which results in (24). 
δ
336 IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO. 2, APRIL 2015

Fig. 2. Structure of the proposed pendubot control system.

B. Nonlinear Disturbance Observer TABLE I

PARAMETERS OF THE PROPOSED CONTROL LAWS AND IDENTIFIED
In this section, we assume that the disturbance exists in the FRICTION MODEL
pendubot system as friction, and explain the algorithm, which
can compensate for the disturbance through the nonlinear distur- Robust swing-up Robust balancing Friction
control law control law model
bance observer. The overall system structure using the proposed in (18a) and (18b) in (30), (39), and (40) in (7)
nonlinear disturbance observer is shown in Fig. 2. First, (9) can
Q = diag(100, 100, 1, 1) γ1 = 4.180
be expressed by introducing X = (q̇a , q̇u )T ∈ 2 as the follow- K P = 2.50 R=1 γ2 = 1.590
ing simple matrix form: K D = 0.10 K = [−63.5, −49.7, −9.7, −4.6] γ3 = 3.150
K E = 9.00 k s = diag(500, 500) γ4 = 0.090
Ẋ = Λ + Ω. (28) d w = 0.01 α = diag(250, 250) γ5 = 3.520
δ = 1.5 μ = diag(10 −5 , 10 −5 ) γ6 = 0.021
When Ω, Ω̇, and Ω̈ are assumed to be bounded for the known
constants ζ1 , ζ2 , and ζ3 , a nonlinear disturbance observer can
be designed based on (28) as
˙ Theorem 2 (Nonlinear Disturbance Estimation): If the fol-
X̂ = Λ + Ω̂ (29) lowing condition (34) is satisfied by each factor of a diagonal
where Ω̂ is the estimate of the disturbance vector. From [30], matrix μ for a nonlinear disturbance observer in (29) and (30),
we can obtain then the estimation error converges to zero with time
˙
Ω̂ = −(ks + α)Ω̂ + μ · sgn(eobs ) + αks eobs (30) 1
μi ≥ ζ4 + ζ5 . (34)
where ks := diag(ks1 , ks2 ), α := diag(α1 , α2 ), and μ := αi
diag(μ1 , μ2 ) ∈ 2×2 are diagonal matrices that have posi-
tive constants; sgn(·) is a general signum function; eobs := Since Ω and Ω̂ are bounded, ||η(·)|| ≤ ζ4 and ||η̇(·)|| ≤ ζ5 are
X − X̂ ∈ 2 is an estimation error or residual. In addition, also bounded for the positive constants ζ4 and ζ5 . Therefore, the
when the filtered estimation error is introduced as estimation error Ω̃ converges to zero. The proof is omitted and
can be referred to [30], [31].
r = ėobs + αeobs (31) Remark 2: In the case of many existing disturbance ob-
servers, the class of the friction model (which can be compen-
the following error equation can be expressed using (28)
sated) is much limited. In particular, the friction model can be
and (29):
both matched and unmatched in the case of the pendubot system.
ėobs = Ω̃ (32) On the other hand, the validity of the friction model proposed
in (7) can be verified in Section V-A. Due to the coupled form
where Ω̃ := Ω − Ω̂. Therefore, if η(t) := (η1 , η2 )T ∈ 2×1 is derived in Lemma 1, the friction and the model uncertainties in
defined as η := Ω̇ + (ks + α)Ω, then the filtered error equation the pendubot system can be compensated using the nonlinear
can be derived as disturbance observer in (29) and (30), which will be verified by
ṙ = ëobs + αėobs simulation and experimental results in Sections V-B and V-C,
respectively.
˙
= Ω̇ − Ω̂ + αėobs
= Ω̇ + (ks + α)Ω̂ − μ · sgn(eobs ) − αks eobs + αėobs C. Robust Balancing Control Using the Nonlinear
= η+(ks +α)(−ėobs ) − (μ · sgn(eobs ) + αks eobs ) + αėobs Disturbance Observer
The pendubot control scheme is switched from the swing-up
= η − ks r − μ · sgn(eobs ). (33)
control to the balancing control to stabilize the unactuated link
The performance of the nonlinear disturbance observer is sum- at the top position. The nonlinear model can be linearized about
marized in the following theorem. the equilibrium point to apply the linear controller as in [11] as
EOM AND CHWA: ROBUST SWING-UP AND BALANCING CONTROL USING NONLINEAR DISTURBANCE OBSERVER FOR THE PENDUBOT 337

the following linearized state equation: error ē of the closed-loop system is guaranteed to be bounded
in the region around the top position. In particular, balancing
ē˙ = Aē + Bτ + Ω
error ē converges to zero exponentially in the region around top
y = C ē (35) position as time goes on.
Proof: The error dynamics of the pendubot system with fric-
where ē = [q −
T
qdT , q̇ −
T
q̇dT ]T ∈  , qd =
4
[ π2 T
, 0] ,
tion in (8) take the form
⎡ ⎤
0 0 1 0
M (q)ë + C(q, q̇)ė + f (q̇) − G(e) = u (41)
⎢ ⎥
⎢ 0 0 0 1⎥
⎢ ⎥ where e := q − qd ∈  . Since we are interested in the region
2
⎢ ⎥
A= ⎢
⎢ θ2 θ4 − θ3 θ5 −θ3 θ5 ⎥, around the top position (i.e., balancing position), G(q) can be
⎢ 0 0⎥⎥ expressed as
⎢ Y Y ⎥
⎣ (θ + θ )θ − (θ2 + θ3 )θ4 (θ1 + θ3 )θ5 ⎦
1 3 5 G(e) = Gl e + Gn l (e, qd ) (42)
0 0
Y Y
⎡ ⎤ where Gl is a linear constant matrix, and Gn l (e, qd ) is the re-
0 maining nonlinear term of G(e), which is, in fact, sufficiently
⎢ ⎥ small in the interested region and converges to zero at the top
⎢ 0 ⎥
⎢ ⎥ position. Therefore, (41) is expressed as
⎢ ⎥
⎢
B= ⎢ θ2 ⎥,
⎥ M (q)ë + C(q, q̇)ė + f (q̇) − Gl e − Gn l (e, qd ) = u (43)
⎢ ⎥
⎢ Y ⎥
⎣ −(θ + θ ) ⎦ T
where u = [τ, 0] . Substitution of the robust control law in (40)
2 3
Y into (43) gives
 
C = [1, 1, 0, 0], Ω = [0, 0, Ωa , Ωu ]T , −K ē
˜
M (q)ë + C(q, q̇)ė + f (q̇) − Gl e − Gn l (e, qd ) =
and Y = θ1 θ2 − θ32 . Then, the balancing controller can be given 0
as (44)
where f˜(q̇) := f (q̇) − fˆ(q̇). Since the stability of the nonlinear
τ = −K ē (36)
error dynamics in (44) around the top equilibrium point can be
where a control gain K is obtained using the LQR control analyzed by that of the linearized error dynamics, the lineariza-
method. The cost function, which is the sum of the squares tion of the error dynamics in (44) should be performed to obtain
of the weighted state variables and the weighted control input for a stable matrix Ac := A − BK as
signal, is expressed as
 ∞ ē˙ = Ac ē + Ω̃ (45)
J= (xT Qx + τ T Rτ )dt (37) which is equivalent to the first equation of (35).
0
Since Ω̃ is bounded and independent of ē, it can be easily
and the LQR control method can be used to obtain the control shown that ē ∈ L∞ . Here, it should be noted that the peaking
law that minimizes the cost function. Here, Q and R are symmet- phenomenon [32] does not happen even with the combination
ric matrices, each of which is positive semidefinite (Q ≥ 0) and of the nonlinear observer and balancing controller due to the
positive definite (R > 0), respectively. For any initial values, fact that f˜ (respectively, Ω̃) is not multiplied by the state vari-
the stabilization gain matrix K that minimizes the cost function able in (44) [respectively, (45)]. Furthermore, Ω̃ decreases to
(37) is given as zero with time from Theorem 2, such that for a stable matrix
K = R−1 B T P (38) Ac (which satisfies ||eA c (t−τ̄ ) || ≤ α1 e−α 0 (t−τ̄ ) for appropriate
positive constants α0 and α1 where || · || denotes the norm of
where P is a symmetric positive-definite matrix, and can be the matrix), γ(:= α0 − α1 ||Ω̃||) can become positive after some
obtained from the algebraic Riccati equation time. To show the exponential stability of (45), the solution of
AT P + P A − P BR−1 B T P + Q = 0. (39) the error dynamics in (45) needs to be obtained as
 t
The previous balancing controller in (36) does not consider the
ē(t) = eA c t ē(0) + eĀ c (t−τ̄ ) Ω̃(t)dτ̄ . (46)
friction of the pendubot and, therefore, it shows much degraded 0
performance due to the friction [i.e., Ω in (35)]. Therefore, we
Hence, we have
need to introduce the disturbance compensation term using its  t
estimate in the balancing control law instead of employing (36).
|ē(t)| ≤ α1 e−α 0 t |ē(0)| + α1 e−α 0 t ||Ω̃|| eα 0 τ̄ |ē(τ̄ )|dτ̄ .
To this end, we design the torque input as 0
(47)
Ω̂a Applying the Bellman–Gronwall Lemma [33] to (47), we have
τ = −K ē − . (40)
Ga
|ē(t)| ≤ α1 e−α 0 t |ē(0)| + α1 |ē(0)|e−γ t . (48)
Theorem 3 (Robust Observer-Based Balancing Control):
For the nonlinear pendubot systems with friction in (8), the Therefore, ē converges exponentially to zero around the top
robust observer-based control law in (40) ensures that balancing equilibrium point. 
338 IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO. 2, APRIL 2015

Fig. 3. Verification of the friction model through the simulation and experimental results. (a) Actuated link angle (dashed line: simulation, solid line: experiment).
(b) Unactuated link angle (dashed line: simulation, solid line: experiment).

Remark 3: The proposed robust balancing control law in (40) linear disturbance observer. As for the allowable magnitude of
provides the local exponential stability of the top equilibrium friction in the proposed method, it can be said that as long as
point in the presence of the friction. Furthermore, the overall the friction is not large enough to break the balanced states
disturbances and uncertainties of the pendubot system can be in the switching region from swing-up to balancing motion (i.e.,
effectively compensated by the robust balancing control law the region, where the robust observer-based balancing control
according to Lemma 1, which will be demonstrated in numerical remains effective), and also does not cause the saturation of the
simulations and experiments. control input, the proposed method can be effectively employed.
Therefore, its validity will be verified through the simulation and
V. SIMULATION AND EXPERIMENTAL RESULTS experimental results as in the following sections.

In this section, the friction model is first verified and, then, B. Simulation Results
both simulation and experimental results are provided to demon-
strate the validity of the proposed method. The performance of the proposed observer in (29) and (30),
which can effectively estimate and compensate for the distur-
bances in the pendubot system is demonstrated by the numeri-
A. Verification of the Friction Model
cal simulation results. The overall parameters of the pendubot
In this section, the validity of the friction model proposed in model including the friction in (7) are the same as Section V-A
(7) is verified through the comparison between the numerical and Table I. When the pendubot completes the swing-up motion
simulation and experimental results. First, in the swing-up according to (18a) in the overall system structure described in
control experiment of the pendubot system, we performed Fig. 2, the swing-up controller switches to the robust balanc-
the stabilization through the robust swing-up and the balanc- ing controller in (40) according to the switching conditions
ing controllers proposed in Sections III and IV. The nomi- (|qa − π/2| ≤ 0.35, |qu | ≤ 0.4, |q̇a | ≤ 0.3, |q̇a | ≤ 0.3) around
nal parameters of the pendubot system in (3) are given as the top position. The gains of the proposed nonlinear disturbance
m1 = 0.09 kg, m2 = 0.73 kg, l1 = 0.15 m, lc1 = 0.09 m, lc2 = observer are chosen as in Table I. First, the system becomes un-
0.08 m, I1 = 0.0073 kg · m2 , I2 = 0.0001 kg · m2 , and g = stable and diverges due to the friction effect when the swing-up
9.8 m/s2 . In addition, the initial position is (qa , qu , q̇a , q̇u ) = controller in [11] is applied to the pendubot system in (3) in
( 1 + π2 , 1 , 2 , 2 ) where 1 = 0.05 rad and 2 = 0.01 rad/s. the presence of the friction, which are omitted here for brevity.
In addition, the design parameters of the robust swing-up con- In the case of the proposed robust swing-up controller method,
troller in (18) and the balancing controller in (36), and the phys- we can see that the performance of the swing-up control can be
ical parameters of the nonlinear friction model are given in maintained well, while suppressing the influences of the fric-
Table I. The comparison between the simulation and experi- tion and the uncertainties. In other words, it is verified that the
mental results is performed for the verification of the friction friction effect can be quite significant during the swing-up and
model, which is shown in Fig. 3. The solid (respectively, dashed) that the robust swing-up controller has to be employed. Fig. 4(a)
lines correspond to the trajectories of each link in the actual ex- and (b) shows that friction of both two links of the pendubot
periments (respectively, simulations). When we applied the bal- system are completely compensated to make both actuated and
ancing controller in (36) to the pendubot system with the actual unactuated links achieve each control goal (i.e., 90◦ and 0◦ , re-
friction based on the linear system without considering the fric- spectively) during the swing-up motion from the initial position.
tion, the limit-cycle phenomenon appears around the unstable In addition, it can be shown from Fig. 4(c) that the estimates of
equilibrium point at the top position. Based on this, the friction the uncertainties quickly converge to the actual ones. Fig. 4(d)
model parameters are determined as in Table I. Although there and (e) shows the estimate D̂ of the upper bound of the distur-
is a slight difference in the form of the state trajectory between bance D and the corresponding control input, respectively. The
the simulation and experimental results, the employed friction performance is shown in Fig. 5 when both uncertainties of the
model can be identified well in terms of both frequency and system and the external disturbance in Fig. 5(e) are present in
magnitude. The small error arises due to many other disturbing the links. As shown in Fig. 5(a) and (b), even when the external
factors in actual environment or the modeling error. However, input is applied to the links at 3 and 6 s as in Fig. 5(e), friction
these errors can be effectively compensated through the non- of two links can be quickly compensated.
EOM AND CHWA: ROBUST SWING-UP AND BALANCING CONTROL USING NONLINEAR DISTURBANCE OBSERVER FOR THE PENDUBOT 339

Fig. 4. Simulation results of the swing-up performance in the presence of fric-

tion. (a) Actuated link angle qa (dashed: reference, solid: proposed method). Fig. 5. Simulation results of the swing-up performance in the presence of both
(b) Unactuated link angle qu (dashed: reference, solid: proposed method). friction and external disturbance. (a) Actuated link angle qa (dashed: reference,
solid: proposed method). (b) Unactuated link angle qu (dashed: reference, solid:
(c) Disturbance and its estimate (dashed: Ω a , solid: Ω̂ a ). (d) Estimate D̂ of
the upper bound of the disturbance D. (e) Control input u. proposed method). (c) Disturbance and its estimate (dashed: Ω a , solid: Ω̂ a ).
(d) Control input u. (e) Time history of external disturbance.

C. Experimental Results
system, the planar SIP model (produced by Quanser Consult-
This section shows the feasibility and effectiveness of the ing Inc.) is used, and its physical parameters are obtained by
proposed algorithm when applied to the actual pendubot sys- Quanser manual of the pendubot system. Of course, we have
tem. The experimental environment to verify the proposed al- verified the validity of these parameters by using several pa-
gorithm in this paper is shown in Fig. 6(a). For the pendubot rameter identification methods in [34] and, thus, we have used
340 IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO. 2, APRIL 2015

Fig. 6. Experimental environment. (a) Apparatus used for the experiment of a pendubot system. (b) Block diagram of the servo system.

Fig. 7. Experimental results of the swing-up performance in the presence of friction. (a) Actuated link angle qa (dashed: reference, solid: proposed method).
(b) Unactuated link angle qu (dashed: reference, solid: proposed method). (c) Disturbance estimate. (d) Control input u.

Fig. 8. Experimental results of the swing-up performance in the presence of both friction and external disturbance. (a) Actuated link angle qa (dashed: reference,
solid: proposed method). (b) Unactuated link angle qu (dashed: reference, solid: proposed method).
EOM AND CHWA: ROBUST SWING-UP AND BALANCING CONTROL USING NONLINEAR DISTURBANCE OBSERVER FOR THE PENDUBOT 341

Fig. 9. Image sequences (in seconds) of the previous swing-up and balancing control [11] of the actual pendubot system.

them as the system parameters in the proposed method. All the pendubot is equipped with two optical encoders for the measure-
physical parameters and control design variables used in the ex- ment of both angular positions of the actuated and unactuated
periment are the same as those in the simulation, and the control links. In the measurement of the velocity, the numerical differen-
torque input τ applied to the actuator link is given by the servo tiation is sensitive to the high-frequency noise, which can cause
system as shown in Fig. 6(b). In this figure, each constant is serious problem in the applications. As a result, we estimate the
−1 −1
given as follows: Kv e  Km Kg rm p , Kcf  ηg Kg ηm Kt rm p , corresponding angular velocities of the actuated and unactuated
−2
and J  M + ηg Kg2 Jm rm p ; ηg and ηm are a planetary gearbox
links from the angular information by using a derivative filter
efficiency constant and a motor efficiency constant, respectively; designed as
Kt , Kg , and Km are a motor torque constant, a planetary gear- ωc2 s
box gear ratio, and a back EMF constant, respectively; Rm and DF(s) = (50)
s2 + 2ζωc s + ωc2
rm p are a motor armature resistance and a motor pinion radius,
respectively; Jm is the rotor inertia of the motor, and Beq is where ζ = 0.7 and ωc = 10 rad/s are the damping ratio and the
the equivalent viscous damping coefficient of the servo system. natural frequency, respectively. The encoders connected to two
Then, the actuator dynamics can be deduced as links can produce the relative angle data in the pendubot system
as shown in Fig. 6(a). Fig. 7 shows the experimental results
τ = −Kx q̇a + Kv vm (49) when the proposed algorithm is applied in the environment of
the actual pendubot system, where there exist friction and un-
where vm and τ are the input voltage and the ideal output certainties. Fig. 7(a) and (b) shows the state trajectories for each
−1
torque of the servo system, and Kx  Kv e Kcf Rm + Beq and link. We can see that they become stabilized toward the top po-
−1
Kcf  Kcf Rm are constants, respectively. The experiments sition rapidly from the initial position, and the influence of the
can be performed in the MATLAB/Simulink (real-time win- friction in the system is effectively compensated. Fig. 8 shows
dows target). Although the theoretical analysis is performed for the envelopes of two state variables when applied by a similar
the continuous-time system model, the proposed control and external disturbance like Fig. 5(e). These results demonstrate
estimation algorithm can work well in the discrete experimen- that the proposed method can maintain the expected perfor-
tal environment due to the fact that the employed experimental mance of the pendubot system in the presence of uncertainties
equipment can operate with very short sampling time of 1 ms, and disturbances. The actual experimental image sequences are
and the proposed control and estimation algorithm is computed shown in Figs. 9 and 10, which can compare more clearly the
by using the ode45 solver. In these experiments, the employed previous swing-up and balancing control method in [11] and
342 IEEE TRANSACTIONS ON ROBOTICS, VOL. 31, NO. 2, APRIL 2015

Fig. 10. Image sequences (in seconds) of the proposed robust swing-up and balancing control of the actual pendubot system.

the proposed robust swing-up and balancing control method. of the proposed robust swing-up controller and balancing con-
In the case of the previous swing-up and balancing controller, troller using the nonlinear disturbance observer has been veri-
there exists limit-cycle phenomenon at the unstable top equi- fied by the stability analysis, numerical simulations, and exper-
librium point due to the friction in Fig. 9. On the other hand, iments.
Fig. 10 shows that the pendubot system is quickly stabilized at
the top position using the robust swing-up and the balancing
controller. REFERENCES
[1] M. Spong, Underactuated Mechanical Systems. Berlin, Germany:
Springer, 1998.
VI. CONCLUSION [2] R. Olfati-Saber, “Nonlinear control of underactuated mechanical systems
with application to robotics and aerospace vehicles,” Ph.D. dissertation,
This paper proposes the robust swing-up and balancing con- Dept. Electr. Eng. Comput. Sci., Massachusetts Inst. Technol., Cambridge,
trol method using the nonlinear disturbance observer for the MA, USA, Feb. 2001.
pendubot system. In the proposed methods, we first identified [3] D. Chwa, “Tracking control of differential-drive wheeled mobile robots
using backstepping-like feedback linearization,” IEEE Trans. Syst., Man,
the friction model and its parameters. Through the comparison Cybern. A, Syst., Humans, vol. 40, no. 9, pp. 1285–1295, Nov. 2010.
of simulation and experimental results, the identified friction [4] P. Batista, C. Silvestre, and P. Oliveira, “A sensor-based controller for
model was verified to be similar to actual friction in the pen- homing of underactuated AUVs,” IEEE Trans. Robot., vol. 25, no. 3,
dubot system. Then, both uncertainties in actuated and unac- pp. 701–716, Jun. 2009.
[5] M.-S. Park and D. Chwa, “Swing-up and stabilization control of inverted-
tuated dynamics are analyzed considering the coupling effect. pendulum systems via coupled sliding-mode control method,” IEEE Trans.
Based on these results, when the previous swing-up and bal- Ind. Electron., vol. 56, no. 9, pp. 3541–3555, Sep. 2009.
ancing controller is applied to the pendubot system, we could [6] H. Fukushima, M. Kakue, K. Kon, and F. Matsuno, “Transformation con-
trol to an inverted pendulum for a mobile robot with wheel-arms us-
see that the significant effect of the friction on the pendubot ing partial linearization and polytopic model set,” IEEE Trans. Robot.,
system is present in the case of the considered actual system. vol. 29, no. 3, pp. 774–783, Jun. 2013.
Unlike most of the previous swing-up and balancing controllers [7] V. Sankaranarayanan and A. Mahindrakar, “Control of a class of under-
actuated mechanical systems using sliding modes,” IEEE Trans. Robot.,
based on the overall energy of the pendubot system, which were vol. 25, no. 2, pp. 459–467, Apr. 2009.
designed without considering the friction and the uncertainties [8] M.-S. Park and D. Chwa, “Orbital stabilization of inverted-pendulum
in the system, the proposed robust swing-up and balancing con- systems via coupled sliding-mode control,” IEEE Trans. Ind. Electron.,
troller incorporates a nonlinear disturbance observer for the un- vol. 56, no. 9, pp. 3556–3570, Sep. 2009.
[9] C. Zhang, J. Franch, and S. Agrawal, “Differentially flat design of a closed-
certain pendubot system to compensate for the friction, system chain planar underactuated 2-DOF system,” IEEE Trans. Robot., vol. 29,
uncertainties, and external disturbance. Finally, the effectiveness no. 1, pp. 277–282, Feb. 2013.
EOM AND CHWA: ROBUST SWING-UP AND BALANCING CONTROL USING NONLINEAR DISTURBANCE OBSERVER FOR THE PENDUBOT 343

[10] M.-S. Park and D. Chwa, “Antisway tracking control of overhead cranes [29] J. E. Marsden and T. S. Ratiu, Introduction to Mechanics and Symmetry:
with system uncertainty and actuator nonlinearity using an adaptive A Basic Exposition of Classical Mechanical Systems, vol. 17. New York,
fuzzy sliding-mode control,” IEEE Trans. Ind. Electron., vol. 55, no. 11, NY, USA: Springer, 1999.
pp. 3972–3984, Nov. 2008. [30] G. Hu, D. Aiken, S. Gupta, and W. Dixon, “Lyapunov-based range iden-
[11] I. Fantoni and R. Lozano, Control of Underactuated Mechanical Systems. tification for paracatadioptric systems,” IEEE Trans. Autom. Control,
New York, NY, USA: Springer, 2001. vol. 53, no. 7, pp. 1775–1781, Aug. 2008.
[12] M. Reyhanoglu, A. van der Schaft, N. McClamroch, and I. Kolmanovsky, [31] D. Chwa, A. Dani, H. Kim, and W. Dixon, “Camera motion estimation for
“Dynamics and control of a class of underactuated mechanical systems,” 3-D structure reconstruction of moving objects,” in Proc. Int. Conf. Syst.
IEEE Trans. Autom. Control, vol. 44, no. 9, pp. 1663–1671, Sep. 1999. Man, Cybern., Seoul, Korea, Oct. 2012, pp. 1788–1793.
[13] T. Albahkali, R. Mukherjee, and T. Das, “Swing-up control of the pen- [32] M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic, Nonlinear and Adap-
dubot: An impulse momentum approach,” IEEE Trans. Robot., vol. 25, tive Control Design. New York, NY, USA: Wiley, 1995.
no. 4, pp. 975–982, Aug. 2009. [33] P. A. Ioannou and J. Sun, Robust Adaptive Control. Chelmsford, MA,
[14] T. Shoji, S. Katsumata, S. Nakaura, and M. Sampei, “Throwing motion USA: Courier Corporation, 2012.
control of the springed pendubot,” IEEE Trans. Control Syst. Technol., [34] K. J. Åström and B. Wittenmark, Adaptive Control. Chelmsford, MA,
vol. 21, no. 3, pp. 950–957, May 2013. USA: Courier Corporation, 2013.
[15] L. Consolini, and M. Maggiore, “On the swing-up of the pendubot us-
ing virtual holonomic constrains,” in Proc. Eur. Control Conf. Decision
Control, Orlando, FL, USA, Dec. 2011, pp. 4803–4808.
[16] W. Wang, J.-Q. Yi, D.-B. Zhao, and X.-J. Liu, “Adaptive sliding mode
controller for an underactuated manipulator,” in Proc. Int. Conf. Mach.
Learn. Cybern., Shanghai, China, Aug. 2004, vol. 2, pp. 882–887.
[17] E. N. Sanchez and V. Flores, “Real-time fuzzy PI + PD control for
an underactuated robot,” in Proc. Int. Symp. Intell. Control, Vancouver,
Canada, Oct. 2002, pp. 137–141. Myunghwan Eom received the B.S. and M.S. de-
[18] C. C. De Wit, H. Olsson, K. J. Åström, and P. Lischinsky, “A new model grees in electrical and computer engineering from
for control of systems with friction,” IEEE Trans. Autom. Control, vol. 40, Ajou University, Suwon, Korea, in 2009 and 2011,
no. 3, pp. 419–425, Mar. 1995. respectively, where he is currently working toward
[19] C. C. De Wit and P. Lischinsky, “Adaptive friction compensation with the Ph.D. degree.
partially known dynamic friction model,” Int. J. Adaptive Control Signal His research interests include robust nonlinear
Process., no. 11, pp. 65–80, Feb. 1997. control and disturbance observer theories, and their
[20] M. Gafvert, “Dynamic model based friction compensation on the furuta applications to robotics and underactuated mechani-
pendulum,” in Proc. Int. Conf. Control Appl., Kohala Coast, HI, USA, cal systems including pendubot and cranes.
Aug. 1999, vol. 2, pp. 1260–1265.
[21] T. H. Lee, K. K. Tan, and S. Huang, “Adaptive friction compensation with
a dynamical friction model,” IEEE/ASME Trans. Mechatronics, vol. 16,
no. 1, pp. 133–140, Feb. 2011.
[22] L. Freidovich, A. Robertsson, A. Shiriaev, and R. Johansson, “LuGre-
model-based friction compensation,” IEEE Trans. Control Syst. Technol.,
vol. 18, no. 1, pp. 194–200, Jan. 2010.
[23] W.-H. Chen, D. J. Ballance, P. J. Gawthrop, and J. O’Reilly, “A non-
linear disturbance observer for robotic manipulators,” IEEE Trans. Ind. Dongkyoung Chwa received the B.S. and M.S. de-
Electron., vol. 47, no. 4, pp. 932–938, Aug. 2000. grees in control and instrumentation engineering and
[24] W.-F. Xie, “Sliding-mode-observer-based adaptive control for servo actua- the Ph.D. degree in electrical and computer engineer-
tor with friction,” IEEE Trans. Ind. Electron., vol. 54, no. 3, pp. 1517–1527, ing from Seoul National University, Seoul, Korea, in
Jun. 2007. 1995, 1997, and 2001, respectively.
[25] K.-S. Kim, K.-H. Rew, and S. Kim, “Disturbance observer for estimating From 2001 to 2003, he was a Postdoctoral Re-
higher order disturbances in time series expansion,” IEEE Trans. Autom. searcher with Seoul National University, where he
Control, vol. 55, no. 8, pp. 1905–1911, Aug. 2010. was also a BK21 Assistant Professor in 2004. He
[26] C. Makkar, G. Hu, W. G. Sawyer, and W. E. Dixon, “Lyapunov-based was a Visiting Scholar with University of New South
tracking control in the presence of uncertain nonlinear parameterizable Wales; the Australian Defence Force Academy; and
friction,” IEEE Trans. Autom. Control, vol. 52, no. 10, pp. 1988–1994, University of Melbourne, Melbourne, VIC, Australia,
Oct. 2007. in 2003, and University of Florida, Gainesville, FL, USA, in 2011. Since 2005,
[27] M. W. Spong and M. Vidyasagar, Robot Dynamics and Control. New he has been with the Department of Electrical and Computer Engineering,
York, NY, USA: Wiley, 2008. Ajou University, Suwon, Korea, where he is currently a Professor. His research
[28] I. Fantoni, R. Lozano, M. W. Spong, “Energy based control of the pen- interests include nonlinear, robust, and adaptive control theories, and their ap-
dubot,” IEEE Trans. Autom. Control, vol. 45, no. 4, pp. 725–729, Apr. plications to robotics, underactuated systems, including wheeled mobile robots,
2000. underactuated ships, cranes, and guidance and control of flight systems.