2019 International Conference on Robotics and Automation (ICRA)

Palais des congres de Montreal, Montreal, Canada, May 20-24, 2019

A Robust Tracking Controller for Robot Manipulators:

Embedding Internal Model of Disturbances
Wonseok Ha and Juhoon Back

Abstract— This paper presents a robust controller for un- not considered explicitly. Extensions to the case with time-
certain robot manipulators subject to disturbances which are varying disturbances with known disturbance model [12]
composed of sinusoids. The controller employs the disturbance and n-DOF manipulators [13] have been reported later. It is
observer based controller which can effectively estimate and
compensate the effect of plant uncertainties and the distur- noted that the models for external disturbances or the lumped
bances. Assuming that the frequencies of sinusoids are known, disturbance considered in those works are linear and it means
we embed the internal model of disturbances into the proposed that plant uncertainties are not considered explicitly.
controller so that the design parameters of the controller can The main contribution of this paper is to embed the
be chosen without using the magnitude of disturbance or its
internal model of disturbance into the disturbance observer
time derivative. A rigorous stability analysis shows that the
closed-loop system under the proposed controller behaves like in order to reject the modeled disturbance and approximately
the nominal closed-loop system free of disturbances. Simulation cancel the effect of plant uncertainties, which has not been
results for a 2-DOF manipulator show the effectiveness of the covered in previous works. The idea of embedding the
proposed controller. disturbance model is adopted from the works [14], [15]
I. INTRODUCTION where uncertain linear systems are considered. Note that
the structure of disturbance observer used in this paper is
The development of robust control strategies for robot different from those used in [11] and its extensions, thus
manipulators has received considerable attention and a num- the current work can be regarded as an extension of [14] to
ber of solutions have been reported in the literature, for nonlinear systems under the assumption that state variables
example, adaptive control [1], [2], sliding mode control [3], are available for feedback.
H ∞ control [4], and disturbance observer [5].
The remainder of this paper is organized as follows.
Among many controllers developed so far, the disturbance
Section II formulates the tracking problem for the robot
observer is well known for its ability of estimating and
manipulator and Section III introduces the structure of
compensating the effect of plant uncertainties and external
proposed robust tracking controller with the disturbance
disturbances, and has been applied to a wide variety of robot
model embedded. In Section IV, the stability of the closed-
control problems including independent joint control [6],
loop system is analyzed based on Lyapunov stability. In
friction estimation and compensation [7], fault detection [8],
Section V, the proposed controller is applied to a 2-DOF
sensorless control [9], [10], and disturbance rejection [11],
manipulator and the performance of proposed controller is
etc.
verified by numerical simulations. Finally, some conclusions
The aim of this paper is to develop a robust tracking
and outlooks are presented in Section VI.
controller for robot manipulators subject to plant uncertain-  >
ties and external disturbances. Our approach is to design Notation: For a given vector x = x1 · · · xn ∈ Rn ,
 >
a disturbance observer explicitly considering the uncertain xfl is defined by xfl = xn · · · x1 . For two vectors
nonlinearity as well as the disturbance. One of most relevant x ∈ Rn1 and y ∈ Rn2 , the concatenated vector [x; y] is
>
defined by [x; y] = x> y > . Concatenation of multiple

works available would be the work [11] where the authors
introduced a trajectory tracking controller whose key compo- vectors (or scalars) are defined similarly when the operation
nent is an estimator for disturbance. It is noted that in [11], an makes sense.
auxiliary variable is used to implement the estimator without
using the angular acceleration signal, which is similar to the II. PROBLEM FORMULATION
case of unknown input observers. However, the disturbance
We consider an uncertain robot manipulator given by
should be constant in [11] and hence the plant uncertainties,
whose effect to the system is obviously time-varying, are M (q)q̈ + C(q, q̇)q̇ + G(q) = τ + τd (1)
This research was supported by the Human Resources Development of
the Korea Institute of Energy Technology Evaluation and Planning (KETEP) where q, q̇, τ , τd ∈ Rn denote joint angles, angular veloci-
grant funded by the Korea government Ministry of Knowledge Economy ties, joint torques, and disturbance torques, respectively, and
(No. 20184030201940) and Basic Science Research Program through the
National Research Foundation of Korea (NRF) funded by the Ministry of M (q) ∈ Rn×n , C(q, q̇) ∈ Rn×n , and G(q) ∈ Rn represent
Education (No. NRF-2018R1D1A1B07042833). inertia matrix, Coriolis and centrifugal force matrix, and
W. Ha and J. Back are with the School of robotics, Kwang- gravity vector, respectively [16], [17]. M̄ (q), C̄(q, q̇), and
woon University, Seoul, Korea. Emails: [email protected],
[email protected] Ḡ(q) denote the nominal counterparts of M (q), C(q, q̇), and
Corresponding author: J. Back G(q), respectively.

978-1-5386-6027-0/19/$31.00 ©2019 IEEE 1162

Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.
 Since the main source of uncertainties of M (q), C(q, q̇),  1Assumption  3: The disturbance torque τd =
and G(q) are length, mass, inertia, and gravity constant, it τd ; . . . ; τdn is of the form
is assumed that each parameter belongs to a known bounded ns
set, for example, the mass of ith link, denoted by mi , satisfies
X
τdi = ai0 + aij sin(ωj t + φij ) (6)
0 < m− + − +
i ≤ mi ≤ mi with mi , mi being known. It is also j=1
assumed that the nominal value for a parameter belongs to
the same set, e.g., the nominal value of mi , denoted by m̄i , where the constants aij ’s and φij are unknown constants while
satisfies 0 < m− + ωj ’s are known positive constants such that ωj 6= ωj 0 when
i ≤ m̄i ≤ mi . Based on this, we assume
the following. j 6= j 0 . 
Assumption 1: The inertia matrix M (q) is symmetric It is noted that τdi given in (6) can be viewed as an output
positive definite. kM (q)k is uniformly bounded with respect of a linear system, namely
to q and the bound is known. M̄ (q) has the same properties ẇi = Ewi , τdi = F wi , wi ∈ Rν (7)
as M (q). 
We would like to design a robust tracking controller such where ν = 2ns + 1, E ∈ Rν×ν , and F ∈ R1×ν are known,
that q(t) follows the desired trajectory qd (t) in a way that and (E, F ) is observable.
for any given  > 0, the controller can be designed so that III. PROPOSED CONTROLLER
lim sup kqd (t) − q(t)k ≤ . (2) In this section, the proposed controller is detailed adopting
t→∞ the idea of classical disturbance observer [5]. Firstly, using
The desired trajectory qd (t) is assumed to satisfy the follow- d, the system (1) is rewritten as
ing. M̄ (q)q̈ + N̄ (q, q̇) = τ + d (8)
Assumption
h 2: qd (t) is i a smooth function such that
qd (t) := qd (t); . . . ; qd(3) (t) is uniformly bounded and the where N̄ (q, q̇) = C̄(q, q̇)q̇ + Ḡ(q). For later use, we define
N (q, q̇) = C(q, q̇)q̇ +G(q). Note that d depends on q, q̇, and
bound is known, i.e., there exists q+ d > 0 such that kqd (t)k ≤ q̈, thus its bound cannot be predetermined.
q+
d , ∀t ≥ 0.  As explained in Section II, if we knew d, then we can make
It is well known that when the plant parameters are exactly the system (1) become the nominal one (3) by applying τ =
known and no disturbance torque is present, the computed −d + τ̄ . Since d cannot be computed directly, we construct
torque controller [17] guarantees equation (2) with  = 0. d̂, an estimate of d, as
Precisely, for the nominal plant free of disturbance given by
d̂ = τ̂p − Q(s)τ
M̄ (q)q̈ + C̄(q, q̇)q̇ + Ḡ(q) = τ̄ , (3)
where τ̂p is an estimate of τ + d, and Q(s) is a low-pass
the computed torque control τ̄ = τr where filter called Q-filter in the disturbance observer literature. It
is noted that Q(s) is introduced to make τ well defined.
τr = M̄ (q)(q̈d +Kd (q̇d −q̇)+Kp (qd −q))+C̄(q, q̇)q̇+Ḡ(q) (4) Replacing q̈ in (8) by an implementable signal, the esti-
mate τ̂p is obtained as
with Kp > 0, Kd > 0, ensures that limt→∞ kqd (t)−q(t)k =
0. τ̂p = M̄ (q)ζ̇ + N̄ (q, q̇) (9)
The main idea of this paper is to develop a dynamic where ζ̇ ∈ Rn is an estimate of q̈ and given by
controller of the form
1
ẋ = Γ(x, q, q̇, τ̄ ) ζ̇ = − Λ(q)(ζ − q̇) (10)
σ
d̂ = γ(x, q, q̇, τ̄ ) (5) where σ > 0 is a design parameter and Λ(q) ∈ Rn is a
τ = −d̂ + τ̄ symmetric positive definite matrix to be chosen later. The
idea behind using the filter (10) is that for sufficiently small
where d̂ is an estimate of the lumped disturbance σ, it holds that ζ ≈ q̇ and ζ̇ ≈ q̈ because the transfer function
Λi
from q̇i to ζi is σs+Λ i
.
d = τd − (M − M̄ )q̈ − (C q̇ + G − C̄ q̇ − Ḡ) The Q-filter Q(s) is designed in a way that the internal
model of disturbance is embedded in the proposed controller
so that the uncertain system (1) behaves similarly to the so that the disturbance is exactly canceled. The idea has
nominal system (3). The lumped disturbance d can be been successfully applied to the disturbance observer based
interpreted as the discrepancy between the actual dynamics controller design for uncertain linear systems [14], [15].
(1) and the nominal one (3) in the sense that the system (1) For the case of unit feedback system with plant P (s) and
with τ = −d + τ̄ becomes (3). As can be seen from (5), it controller C(s) where the relations of signals (control input
is assumed that both q and q̇ are available for feedback. u, reference r, output y, disturbance d with ẇ = Ew, d =
The disturbance torque τd considered in this paper is F w) are given by u = C(s)(r − y), y = P (s)(u + d),
composed of ns sinusoids and constants as stated below. this can be done by including the characteristic polynomial

1163
Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.
of E in the denominator of C(s) (or loop transfer function τd
τr τ q, q̇
C(s)P (s)). Following this principle, we derive a component P
which corresponds to C(s). From the relation −
τ = −d̂ + τ̄ = −τ̂p + Q(s)τ + τ̄ , χ̇ = (In ⊗(Aν −∆−1 fl
σ α Cν ))χ
−1 fl
+(In ⊗(∆σ α −δ fl ))τ
one has, with Q(s) = Qχ (s)In ,
1 (In ⊗ Cν )χ
τ= (−τ̂p + τ̄ ).
1 − Qχ (s) d̂ − τ̂p ζ̇ = − 1 Λ(ζ − q̇)
σ
Thus, it turns out that the component 1/(1 − Qχ (s)) corre- τ̂p = M̄ ζ̇ + N̄
sponds to C(s) from the unit feedback case, and from this
observation, we realize Qχ (s) as Fig. 1. Structure of the proposed controller.
i
χ̇ = (Aν −∆−1 fl i −1 fl fl i
σ α Cν )χ +(∆σ α −δ )τ (11)
i ν
where χ ∈ R , Suppose σ ∈ (0, 1), e(0) ∈ Se (δe ), χ(0) ∈ Sχ (δχ ), and
w(0) ∈ Sw (δw ), and define
 
0 Iν
Aν = ν 0>
 
, Cν = 1
0 0>
ν
ν
η̃ =(In ⊗ σ −1 ∆σ )(χ + (In ⊗ Σ)w)
2 ν
∆σ = diag{σ, σ , · · · , σ }, − (In ⊗ B̄)(M (q)(q̈d + Kp e1 + Kd e2 ) + N )
 
α = α1 ;· · · ; αν is chosen such that Aν −αfl Cν is Hurwitz, ξ = − Λ−1 (q)(q̈d + Kp e1 + Kd e2 ).
∗

and δ =Q δ1 ; · · · ; δν is a vector whose components are taken

from s i=1
ns
(s2 + ωi2 ) = sν + δν sν−1 + · · · + δ2 s + δ1 . Then, from Assumptions 1 and 2, there exists δη̃ > 0 and
It is noted that by the way δ is constructed, there exists a δξ̃ > 0 such that kη̃(0)k ∈ Sη̃ (δη̃ ) := {η̃ ∈ Rnν |kη̃k ≤
nonsingular matrix Σ such that ˜
δη̃ } and that ξ(0) = −ξ ∗ (0) ∈ Sξ̃ (δξ̃ ). It is noted that the
∗
variables η̃ and ξ will be determined in Section IV and
(Aν − δCν )Σ = ΣE, Cν Σ = F. (12) η̃(t) = η(t) − η ∗ (t) with η(0) = 0 will be used in the
From (9), (10), and (11), the proposed controller becomes stability analysis.
We define a Lyapunov function candidate
χ̇ = (In ⊗ (Aν −∆−1 fl −1 fl fl
σ α Cν ))χ+(In ⊗ (∆σ α −δ ))τ
˜ = e> Pe e + η̃ > ξ˜> Pψ η̃; ξ˜
   
1 V (e, η̃, ξ) (16)
ζ̇ = − Λ(q)(ζ − q̇), ζ(0) = q̇(0)
σ where Pe be the solution of Pe Ae + A>
e Pe = −In with
τ̂p = M̄ (q)ζ̇ + N̄ (q, q̇)  
0 In
d̂ = τ̂p − (In ⊗ Cν )χ Ae = n×n . (17)
−Kp −Kd
1
= − M̄ (q)Λ(q)(ζ − q̇) − (In ⊗ Cν )χ + N̄ (q, q̇) Given l > 0, a level set for the Lyapunov function V is
σ
τ = −d̂ + τr defined by
(13) ˜ (e, η̃, ξ)
Ωl = {(e, η̃, ξ)|V ˜ ≤ l}. (18)
where χ = χ ; · · · ; χn . Note that the proposed controller
 1 
Theorem 1: Suppose σ ∈ (0, 1) and positive constants
is of the form (5) with x = [χ; ζ]. δe , δχ , δw are given such that e(0) ∈ Se (δe ), χ(0) ∈ Sχ (δχ ),
We make a technical assumption related to the design and w(0) ∈ Sw (δw ). Let l be such that Se (δe ) × Sη̃ (δη̃ ) ×
parameters µ and α. Sδξ̃ ⊂ Ωl . Then, for any given  > 0, there exists σ ∗ > 0
Assumption 4: There exists α ∈ Rν whose components ∗
are all positive and a constant µ > 0 such that for the matrix
such that for any  0 < σ < σ ∗, the controller given by (13)
guarantees that e(t); η(t) − η (t); ξ(t) − ξ ∗ (t) ∈ Ωl , ∀t ≥


µ(In ⊗ αfl )
 
In ⊗ Aν 0 and that lim supt→∞ kqd (t) − q(t)k ≤ . 
Aψ = , (14)
−(M −1 ⊗ Cν ) −µ(M −1 + M̄ −1 ) We emphasize that the proposed controller can reject sinu-
soidal disturbances in the sense that the controller parameters
there exists a constant matrix Pψ = Pψ> > 0 such that
do not depend on the magnitude of disturbances or their time
Pψ A ψ + A >
ψ Pψ ≤ −I (15) derivatives explicitly, which can be seen in the stability proof
given in the following section.
for all m− In ≤ M ≤ m+ In . 
In order to state the stability of the closed-loop system, IV. STABILITY ANALYSIS
we define sets of initial conditions. Let δ > 0 and define We define η = η 1 ; · · · ; η n and ξ = ξ 1 ; · · · ; ξ n where
   

Se (δ) = {e ∈ R2n |kek ≤ δ} 1

nν ηi = ∆σ (χi + Σwi )
Sχ (δ) = {χ ∈ R |kχk ≤ δ} σ (19)
1
Sw (δ) = {w ∈ Rnν |kwk ≤ δ}. ξ i = (ζ i − q̇ i ).
σ

1164
Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.
With η, ξ defined above and applying the properties Similarly, one can compute
1 1 σ η̃˙ = (In ⊗ (Aν − ∆σ δ fl Cν ))η̃ + µ(In ⊗ (αfl − ∆σ δ fl ))ξ˜
∆σ Aν ∆σ−1 = Aν , ∆σ δ fl Cν ∆−1 σ = ∆σ δCν ,
σ σ
− (In ⊗ ∆σ δ fl )(M (q̈d + Kp e1 + Kd e2 ) + N ) − σ η̇ ∗
one can rewrite the closed-loop system as
˙
σ ξ˜ = −M −1 (In ⊗ Cν )η̃ − µ(M̄ −1 + M −1 )ξ˜ − σ ξ˙∗
M q̈ + N =τr + M̄ Λξ − N̄ +(In ⊗ Cν )η
(24)
σ η̇ =(In ⊗ (Aν − ∆σ δ fl Cν ))η
Define ψ = η̃; ξ˜ . Then, the closed-loop system given by
 
+ (In ⊗ (αfl − ∆σ δ fl ))(τr + M̄ Λξ − N̄ ) (20)
(23) and (24) becomes
σ ξ˙ = − (M −1 M̄ + In )Λξ
ė = Ae e + Be ψ
− M −1 (τr − N̄ − N + (In ⊗ Cν )η)   (25)
σ ψ̇ = Aψ ψ + Bψ1 ; Bψ2
which is in standard singular perturbation form with time
scale separation parameter σ [18]. The behavior of the where
overall system can roughly be explained as follows. When Bψ1 (e, ψ, qd )
the dynamics of η and ξ is exponentially stable and σ is
= −(In ⊗ ∆σ δ fl Cν ))η̃ − µ(In ⊗ ∆σ δ fl )ξ˜
sufficiently small, the fast variables η and ξ rapidly approach
to η ∗ and ξ ∗ which are functions of slow variables q, qd , − (In ⊗ ∆σ δ fl )(M (q̈d + Kp e1 + Kd e2 ) + N ) − σ η̇ ∗
τd , etc. As a result, the state trajectory (q(t), q̇(t)) of (20) Bψ2 (e, ψ, qd ) = −σ ξ˙∗ .
becomes very similar to that of reduced system which is
obtained by replacing (η, ξ) by (η ∗ , ξ ∗ ) in the q-dynamics It is noted that from (22) and the relation
of (20) provided that the reduced system is exponentially ξ ∗ = −Λ−1 (q)(q̈d + Kp e1 + Kp e2 ), (26)
stable. See for example [18] for details.
In order to analyze the stability of the closed-loop system, it holds that η̇ ∗ and ξ˙∗ are functions of e, ψ, and qd .
we find the equilibrium point of (η, ξ)-dynamics for the ex- Now, consider the Lyapunov function candidate (16) and
treme case σ = 0, denoted by (η ∗ , ξ ∗ ). (η ∗ , ξ ∗ ) corresponds the level set Ωl defined in (18). It will be shown that if we
to the quasi-steady-state in the singular perturbation theory take sufficiently small σ, then the set Ωl is forward invariant
and computed from with respect to the dynamics (25). Without loss of generality,
we take σ ≤ 1.
0 = (In ⊗ Aν )η ∗ + (In ⊗ αfl )[τr + M̄ Λξ ∗ − N̄ ] Compute the time derivative of V as follows.
0 = (M −1 M̄ + In )Λξ ∗ +M −1 (τr − N̄ − N +(In ⊗ Cν )η ∗ ).
V̇ =e> (Pe Ae + A> >
e Pe )e + 2e Pe Be ψ
∗ ∗
One can find an explicit form of (η , ξ ) as follows, whose 1 1 >
proof is omitted. + ψ > (Pψ Aψ + A> ψ Pψ )ψ + 2ψ Pψ Bψ
σ σ
Lemma 1: The quasi-steady-state of (η, ξ)-dynamics is 1
given by ≤ − kek2 − kψk2 + κ1 kekkψk + κ2 kψk
σ
η ∗ = (In ⊗ B̄)(M M̄ −1 (τr − N̄ ) + N ) where
(21)
ξ ∗ = −Λ−1 M̄ −1 (τr − N̄ ) κ1 = k2Pe Be k
1
 
where B̄ = 1; 0ν−1 .      κ2 = max 2 Pψ Bψ .
To proceed, define e = e1 ; e2 := qd − q; q̇d − q̇ , η̃ = 0≤σ≤1,kqd k≤q+
d ,(e,ψ)∈Ωl
σ
η − η ∗ , and ξ˜ = ξ − ξ ∗ , and rewrite the closed-loop system Applying Young’s inequality, we have
(20) in these coordinates. To derive the dynamics of e, we  
compute 1 1 2 1 1
V̇ ≤ − kek2 + κ1 − + kψk2 + ¯κ22
2 2 σ 4¯ 
M q̈ + N = (In ⊗ Cν )η̃ + M̄ Λξ˜ + (In ⊗ Cν )η ∗
where ¯ is an arbitrary positive constant.
+ M̄ Λξ ∗ + τr − N̄
Take σ ∗ ∈ (0, 1) such that
= (In ⊗ Cν )η̃ + M̄ Λξ˜ + M (q̈d + Kp e1 + Kd e2 )
1 2 1 1 λmax (Pψ )
κ − + ≤− .
where we used the relation (21) and 2 1 σ∗ 4¯
 2λmax (Pe )
(In ⊗ Cν )η ∗ = M (q)(q̈d + Kp e1 + Kd e2 ) + N (q, q̇). (22) Then, for any 0 < σ < σ ∗ , it holds that
Applying Λ = µM̄ −1 , we have 1 λmax (Pψ )
V̇ ≤ − kek2 − kη̃k2 + ¯κ22
2 2λmax (Pe )
ė = Ae e + Be η̃; ξ˜
 
(23)
≤ −λV + ¯κ22
where Ae is defined in (17) and where λ = 1
Thus, if ¯ ≤ λl
=: ¯1 , then one has
  2λmax (Pe ) . κ22
0nν×nν 0nν×n V̇ ≤ −λ(V − l) ≤ 0 on the boundary of Ωl , which shows
Be = .
−M −1 ⊗ Cν −µM −1 that Ωl is forward invariant.

1165
Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
˜
Since the initial condition of ζ is chosen so that ξ(0) =
S IMULATION PARAMETERS
∗ ˜
−ξ , it holds that [e(0); η̃(0); ξ(0)] belongs to the set Ωl
Parameter Value Parameter Value or Function
and forward invariance of Ωl ensures that the trajectory m̄1 3 [kg] q1 (0) 0.17 [rad]
[e(t); ψ(t)] remains in Ωl for all t ≥ 0. Using the comparison m̄2 2 [kg] q2 (0) 0.61 [rad]
lemma, we have l̄1 0.75 [m] qd1 π/6 sin π/5t [rad]
l̄2 0.75 [m] qd2 π/12 + π/6 cos π/5t [rad]
¯κ22 I¯c1 0.5625 [Nm2 ] τd11 5 sin(2t + π/5) [Nm]
V (t) ≤ e−λt V (0) + . I¯c2 0.375 [Nm2 ] τd12 3 sin(3t + π/6) [Nm]
λ m1 3.3 [kg] τd13 sin 5t [Nm]
Thus, m2 1.85 [kg] τd1c 3 [Nm]
l1 0.685 [m] τd21 2 sin(2t + π/12) [Nm]
¯κ22 l2 0.8 [m] τd22 3 sin(3t + π/3) [Nm]
lim sup V (t) ≤ Ic1 0.516 [Nm2 ] τd23 4.5 sin(5t + π/6) [Nm]
t→∞ λ Ic2 0.3947 [Nm2 ] τd2c 5 [Nm]
and kp 45 kd 10
s µ 50 σ 0.001
¯κ22
lim sup ke(t)k ≤ .
t→∞ λλmin (Pe )
values of angular displacement are also given in Table I and
Therefore, if we take ¯ ≤ min{¯ 1 , ¯2 } where ¯2 = it is assumed that the manipulator is at rest initially.
2
λλmin (Pe ) κ 2 , the trajectory of the closed-loop system re- Fig. 2 shows the tracking performance of the computed
2
mains bounded in the sense that [e(t); ψ(t)] ∈ Ωl , ∀t ≥ 0 torque controller when it is applied to the nominal system
and satisfies that lim supt→∞ kqd (t) − q(t)k ≤ . Thus, the free of disturbance. In the figure, qdi and q̄i represent the
proof is complete. desired trajectory and the nominal angular displacement of
joint i, respectively. As expected, the angular displacements
V. SIMULATION of the manipulator track the desired trajectory with no steady-
In this section, we evaluate the performance of the pro- state error. Fig. 3 shows the external disturbance torques
posed controller by numerical simulation on a 2-DOF ma- applied to the system, and they consist of a constant and
nipulator, i.e., n = 2 in the system (1). The inertia matrix three sinusoidal signals, respectively. It is well known that
can be expressed by the computed torque controller is not robust against plant
  uncertainties because it is essentially based on the feedback
d + 2d2 + d3 d2 + d3 linearization. As can be seen in Fig. 4, the tracking per-
M (q) = 1
d2 + d3 d3 formance under the computed torque controller is degraded
where significantly when it is applied to the actual system subject
to the effect of system uncertainties and the external distur-
2
d1 = m1 lc1 + m2 (l12 + lc2
2 2
+ 2l1 lc2 ) + Ic1 bances described in Fig. 3.
d2 = m2 l1 lc2 cos q2 Fig. 5 shows the simulation results when the proposed
2
d3 = m2 lc2 + Ic2 . controller is applied to the uncertain system and it is seen
that in spite of the exerted external disturbance and system
The Coriolis and centrifugal force matrix and the gravity uncertainties, the performance of the nominal closed-system
vector are given by shown in Fig. 2 is recovered. Fig. 6 shows the difference
    between the actual angular displacement and its nominal
−c1 −c1 − c2 g1 + g2
C(q, q̇) = , G(q) = counterpart of joint 1, q1 (t) − q̄1 (t), with various values of
c2 0 g2
σ. It turns out that as σ becomes smaller, the difference
where reduces towards zero (magnitude in rms: 5.34 × 10−4 [rad]
(σ = 0.1), 4.19 × 10−5 [rad] (σ = 0.01), 2.63 × 10−6 [rad]
c1 = m2 l1 lc2 q̇2 sin q2 (σ = 0.001)). This means that the controller can be tuned so
c2 = m2 l1 lc2 q̇1 sin q2 that the behavior of the closed-loop system can be made as
g1 = (m1 lc1 + m2 l1 )g cos q1 close as desired to that of the nominal one.
g2 = m2 lc2 g cos(q1 + q2 ). To verify the performance of the proposed controller, we
compared the proposed controller with that of [11]. We
The numerical values used in the simulation are sum- follow the design procedure described in [11] and the gain of
marized in Table I. The design parameters of the outer controller is chosen as c = 1000. Fig. 7 shows the tracking
loop controller, the computed torque controller given by errors of both controllers when external disturbances and un-
(4) designed for nominal system in our case, are taken as certainties are present. The proposed controller has approx-
Kp = kp I2 and Kd = kd I2 where I2 ∈ R2×2 is the imately 25 times smaller steady-state error (magnitude of
identity matrix. Assume that the external disturbances are steady-state error: 2 × 10−5 [rad] (proposed), 5 × 10−4 [rad]
given by τd1 = τd11 + τd12 + τd13 + τd1c and τd2 = (that of [11])), which shows that the proposed controller
τd21 +τd22 +τd23 +τd2c , and the numerical values associated effectively attenuates modeled disturbances and the effect of
with the disturbances can be found in Table I. The initial system uncertainties.

1166
Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.
 0.8 qd1 0.8 qd1
qd2 qd2
0.6 q̄1 0.6 q1
q̄2 q2
0.4 0.4
Angle [rad]

Angle [rad]
0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Time [s] Time [s]
Fig. 2. Trajectory of angular displacements: nominal closed-loop system. Fig. 5. Performance recovery under the proposed controller.

×10-5
14 τd1 5 σ = 0.1
12 τd2 4 σ = 0.01
Disturbance Torque [Nm]

10 3 σ = 0.001

Difference [rad]
8 2
1
6
0
4
-1
2
-2
0 -3
-2 -4
-4 -5
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Time [s] Time [s]
Fig. 3. Sinusoidal disturbance torques applied to joints. Fig. 6. Discrepancy between q̄1 and q1 for various σ.

×10-3
1.2 qd1 proposed controller
5 e1
1 qd2 4 e2 proposed controller
0.8 q1 3 e1 Chen’s controller
Tracking Error [rad]

q2 e2 Chen’s controller
0.6 2
Angle [rad]

0.4 1
0
0.2
-1
0 -2
-0.2 -3
-0.4 -4
-0.6 -5

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Time [s] Time [s]
Fig. 4. Performance degrade of computed torque controller in the presence Fig. 7. Comparison of the tracking error: proposed controller versus that
of external disturbances and system uncertainties. of [11].

VI. C ONCLUSION the bound of disturbance torque or its time derivative. For
This paper proposes a robust tracking controller for robot the case where the angular velocities are not available, it
manipulators under the assumption that the disturbance is expected that so-called peaking phenomenon takes place
model is known and that the angular displacements and their and one can still recover the nominal performance by using a
time derivatives are available for feedback. By embedding high gain observer and saturating the input. Future research
the internal model of disturbance in the disturbance observer topics include the extension of the proposed controller to
based controller, it is seen that the nominal performance can more general nonlinear systems and experiments on real
be recovered and that the controller design does not require robot manipulators.

1167
Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.
 R EFERENCES
[1] J. J. E. Slotine and W. Li, “Adaptive manipulator control: A case
study,” IEEE Trans. on Industrial Electronics, vol. 33, no. 11, pp.
995–1003, 1988.
[2] Y. Kim, J. Seok, I. Noh, and S. Won, “An adaptive disturbance observer
for a two-link robot manipulator,” IEEE Conf. on Control, Automation
and Systems, pp. 141–145, 2008.
[3] Y. Pi and X. Wang, “Trajectory tracking control of a 6-dof hydraulic
parallel robot manipulator with uncertain load disturbances,” Control
Engineering Practice, vol. 19, no. 2, pp. 185–193, 2011.
[4] A. F. Khelfi and A. Abdessameud, “Robust H-infinity trajectory
tracking controller for a 6 dof puma 560 robot manipulator,” Proc.
of IEEE int. conf. on electrical mashines and drives, pp. 88–94, 2007.
[5] K. Ohnishi, “A new servo method in mechatronics,” Trans. of Japanese
Society of Electrical Engineers, vol. 107-D, pp. 83–86, 1987.
[6] K. S. Eom, I. H. Suh, and W. K. Chung, “Disturbance observer
based path tracking control of robot manipulator considering torque
saturation,” Mechatronics, vol. 11, pp. 325–343, 2001.
[7] A. Mohammadi, H. J. Marquez, and M. Tavakoli, “Disturbance
observer-based trajectory following control of nonlinear robotic ma-
nipulators,” Proc. of the Canadian congress of applied mechanics, pp.
779–782, 2011.
[8] K. Ohnishi and H. Ohde, “Collision and force control for robot ma-
nipulator without force sensor,” Proc. of IEEE Int. Conf. on Industrial
Electronics, Control and Instrumentation, pp. 766–771, 1994.
[9] S. Komada and K. Ohnishi, “Force feedback control of robot manipu-
lator by the acceleration tracing orientation method,” IEEE Trans. on
Industrial Electronics, vol. 37, no. 1, pp. 6–12, 1990.
[10] T. Murakami, F. Yu, and K. Ohnishi, “Torque sensorless control
in multidegree-of-freedom manipulator,” IEEE Trans. on Industrial
Electronics, vol. 40, no. 2, pp. 259–265, 1993.
[11] W. H. Chen, D. J. Ballance, P. J. Gawthrop, and J. O’Reilly, “A
nonlinear disturbance observer for robotic manipulators,” IEEE Trans.
Ind. Electron., vol. 47, no. 4, pp. 932–938, 2000.
[12] W. H. Chen, “Disturbance observer based control for nonlinear sys-
tems,” IEEE/ASME Transactions on Mechatronics, vol. 9, no. 4, pp.
706–710, 2004.
[13] A. Mohammadi, M. Tavakoli, H. J. Marquez, and F. Hashemzadeh,
“Nonlinear disturbance observer design for robotic manipulators,”
Control Engineering Practice, vol. 21, no. 3, pp. 253–267, 2013.
[14] G. Park, Y. Joo, and H. Shim, “Asymptotic rejection of sinusoidal dis-
turbance with recovered nominal transient performance for uncertain
linear systems,” in Proc. of IEEE Conference on Decision and Control,
Dec. 15-17, Los Angeles, California, USA, 2014, pp. 4404–4409.
[15] Y. Joo, G. Park, J. Back, and H. Shim, “Embedding internal model
in disturbance observer with robust stability,” IEEE Transactions on
Automatic Control, vol. 61, no. 10, pp. 3128–3133, 2016.
[16] J. J. Craig, Introduction to robotics: mechanics and control, Third Ed.
Prentice-Hall, 2005.
[17] M. W. Spong, S. Hutchinson, and M. Vidyasagar, Robot modeling and
control. Wiley, 2006.
[18] H. K. Khalil, Nonlinear Systems, Third Ed. Prentice-Hall, Upper
Saddle River, NJ, 2002.

1168
Authorized licensed use limited to: Turun Yliopisto. Downloaded on July 09,2024 at 13:31:13 UTC from IEEE Xplore. Restrictions apply.