Robust Control of Robot Manipulators Using Inclusive and Enhanced Time Delay Control

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO.

5, OCTOBER 2017 2141

Robust Control of Robot Manipulators Using


Inclusive and Enhanced Time Delay Control
Maolin Jin, Member, IEEE, Sang Hoon Kang , Member, IEEE, Pyung Hun Chang, Member, IEEE,
and Jinoh Lee, Member, IEEE

Abstract—Thanks to its simplicity and robustness, time sliding surface, robot manipulators, time delay control
delay control (TDC) has been recognized as a simple and (TDC), time delay estimation (TDE).
yet effective alternative to robot model-based controls
and/or intelligent controls. An inclusive and enhanced for-
mulation of TDC for robust control of robot manipulators is I. INTRODUCTION
presented in this paper. The proposed formulation consists
OTION control of robot manipulators is a highly chal-
of three intuitive terms: 1) time delay estimation (TDE),
inherited from the original TDC, for cancellation of mostly
continuous nonlinearities; 2) nonlinear desired error
M lenging task. Robot dynamics is highly nonlinear and
strongly coupled because of the nonlinear terms, including grav-
dynamics (DED) (i.e., a “mass”–“nonlinear damper”– ity, Coriolis and centrifugal torque, and nonlinear friction and
“nonlinear spring” system) injection term; and 3) a TDE
other disturbances. Identifying the parameters of the dynamics
error correction term based on a nonlinear sliding surface.
The proposed TDC formulation has an inclusive structure. could be a cumbersome task, and the number of parameters
Depending on the gain/parameter set chosen, the proposed of the model (e.g., the inertia and mass of each segment) that
formulation can become Hsia’s formulation, Jin’s formu- need to be estimated generally increases with the increase of the
lations including a type of terminal sliding mode control degree-of-freedom (DOF) of the robot. Most controllers based
(SMC), an SMC with a switching signum function, or a novel
on robot dynamics [4] are highly complicated, because of the
enhanced formulation. Experimental comparisons were
made using a programmable universal manipulator for necessity of computation of the nonlinear terms of the robot
assembly-type robot manipulator with various parameter dynamics, and may be sensitive to changes in the parameters of
sets for the proposed control. Among them, the highest po- the robot dynamics. To name a few, computed torque control,
sition tracking accuracy was obtained by using a terminal computed-torque-like control, sliding mode control (SMC), and
sliding DED with a terminal sliding correction term.
adaptive control with robot nominal model are the ones based on
robot dynamics model. To alleviate the need of a robot dynam-
Index Terms—Inclusive formulation, model-free control ics model, intelligent control techniques (e.g., fuzzy control and
(MFC), nonlinear desired error dynamics (DED), nonlinear
neural networks) have been developed, as a model-free control
(MFC), by obtaining “black box” models of robot manipula-
Manuscript received July 21, 2016; revised November 1, 2016, Febru- tor dynamics [5]–[9]. To use intelligent controls, however, one
ary 15, 2017, and April 26, 2017; accepted June 11, 2017. Date of needs to tune a number of gains/parameters of its own that may
publication June 20, 2017; date of current version October 13, 2017. affect the control performance. Thus, for practicality, there is a
Recommended by Technical Editor Y.-J. Pan. This work was supported
in part by the 2015 Research Fund of the Ulsan National Institute of need for a simple yet effective control scheme.
Science and Technology under Grant 1.150048.01; in part by the Trans- An alternative to the aforementioned controllers could be the
lational Research Center for Rehabilitation Robots, National Rehabilita- time delay control (TDC) [9]–[16] using time delay estimation
tion Center, Ministry of Health and Welfare, Korea under Grant NRCTR-
EX17012; and in part by the Ministry of Trade, Industry, and Energy (TDE) to compensate for complex robot dynamics without a
(MOTIE, Korea) under Industrial Technology Innovation Program (Devel- robot dynamics model and its parameters [13]. TDE requires
opment of armored robot systems for personal protections of rescuemen only the most recent past control input to the robot and the
and emergency management operations in the composite disaster site)
under Grant 10067184. (Corresponding Author: Sang Hoon Kang.) acceleration for estimation and compensation of the robot dy-
M. Jin is with the Korea Institute of Robot and Convergence, Pohang namics (including uncertainties and unknown dynamics, and
37666, South Korea (e-mail: mulimkim@kiro.re.kr). disturbances [9]–[15], [17]), based on the assumption that the
S. H. Kang is with the Robotics and Rehabilitation Engineering Lab-
oratory, Department of System Design and Control Engineering, Ulsan dynamics and disturbances are at least piecewise continuous.
National Institute of Science and Technology (UNIST), Ulsan 44919, Of course, TDE needs the order and structure of the robot dy-
South Korea (e-mail: sanghkang@unist.ac.kr). namics but not the parameters or a dynamics equation. With
P. H. Chang is with the Department of Robotics Engineering, Daegu
Gyeongbuk Institute of Science and Technology, Daegu 42988, South compensation/cancellation of robot dynamics using TDE, it is
Korea (e-mail: phchang@dgist.ac.kr). possible to specify the desired error dynamics (DED) explic-
J. Lee is with the Department of Advanced Robotics, Istituto Italiano itly and inject them as closed-loop dynamics behavior [10],
di Tecnologia, Genoa 16163, Italy (e-mail: jinoh.lee@iit.it).
Color versions of one or more of the figures in this paper are available [11], [13], [14], which were mostly linear dynamics. Because of
online at http://ieeexplore.ieee.org. the simplicity of its structure, model-freeness (independence),
Digital Object Identifier 10.1109/TMECH.2017.2718108 numerical efficiency, and robustness against unknown and

1083-4435 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
2142 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO. 5, OCTOBER 2017

uncertain dynamics and disturbances, TDC has been utilized TABLE I


NOMENCLATURE
to control various systems, including robot manipulators [10]–
[16], [18]–[21].
Overall, TDE estimates the robot dynamics and disturbance Symbol Variable Name
closely. However, discontinuous and/or abruptly changing dy- q, q̇, q̈ ∈ n joint position, velocity, and acceleration
namics (e.g., Coulomb friction and stiction) can degrade TDE M(q) ∈ n ×n generalized inertia matrix
performance. From the view point of TDE, continuous nonlin- C(q, q̇) ∈ n ×n Coriolis/centripetal matrix
G(q), F(q, q̇) ∈ gravitational torque and friction disturbance torque
ear terms are classified as soft nonlinearities, and discontinuous n
terms such as Coulomb friction and stiction are classified as τ d , τ ∈ n disturbance torque and joint torque
hard nonlinearities [9], [17]. Friction disturbance accounts for M̄ ∈ n ×n positive constant diagonal matrix
•(t ) , •(t −L ) values of • at times t and t-L
as much as 30% of the maximum motor torque and behaves non- L∈ estimation time delay
linearly even at low velocities [22]–[25]. It has been reported q d ∈ n desired trajectory
that pulse-type TDE errors due to hard nonlinearity (e.g., sudden e ∈ n position error
K D , K P ∈ n ×n diagonal coefficients matrices
changes in the magnitude and polarity of Coulomb friction at u h , u n ∈ n auxiliary control signal [1] (also known as the new
velocity reversal) can increase position tracking error [26]. To control input [2], [3])
improve TDC, in conjunction with TDE, various schemes (e.g., ε ∈ n TDE error
q̇ id e a l ∈ n ideal velocity
SMC [27], ideal velocity feedback [17], [28], internal model α, β ∈ n exponent vectors
control [29], terminal sliding mode (TSM) [9], and nonlinear ei , ėi ∈  ith joint position error and its time derivative
damping [30], [31]) have been exploited. sgn(•) signum function
s ∈ n integral sliding variable
Recently, nonlinear sliding surfaces have been adopted as the λ ∈ n ×n constant diagonal gain matrix whose diagonal
DED of TDC, in place of the commonly used linear dynamics. elements are non-negative
A TSM was used for the TDC’s DED injection term to design (λi i ≥ 0; i = 1, 2, . . . , n)
γ ∈ n exponent vector whose elements are non-negative
a fast and terminal sliding DED [26], [30]. Because the TSM (γi ≥ 0; i = 1, 2, . . . , n)
speeds up the convergence rate near the equilibrium point, a ξ∈ finite positive constant
properly designed terminal sliding DED has been shown to im- λmin ∈  λ’s smallest diagonal element
ω∈ cutoff frequency
prove the position tracking performance of TDC in comparison
to TDC with linear DED [26], [32]. The adoption of a hyper-
bolic tangential sliding mode (SM) as a DED has been shown
to alleviate the negative effects of velocity saturation and noise II. BRIEF REVIEW OF TDC
due to numerical differentiation [33]. In short, advanced control TDC will be briefly reviewed with the introduction of the
schemes that combine TDE with various SMs [26], [30], [32], TDC with linear DED [11], [12].
[33] have shown promising results. Thus, properly designed
nonlinear DED may further improve the tracking performance A. Robot Manipulator Dynamics
of TDC.
However, during the reaching phase before the SM occurs, n-DOF robot manipulator dynamics can be described as fol-
controllers with an SM [26], [30], [32], [33] (and many other lows:
conventional SMCs) may not guarantee the robustness to param- M(q)q̈ + C(q, q̇)q̇ + G(q) + F(q, q̇) + τ d = τ . (1)
eter variations and disturbances [34]–[37]. Furthermore, those
utilizing an SM [26], [27] may have discontinuous chattering. The dynamics in (1) can be rewritten as follows with the
The goals of this paper are to propose an enhanced TDC for- introduction of M̄ (see Table I):
mulation that can include many other previously proposed TDC
M̄q̈ + N(q, q̇, q̈) = τ (2)
formulations; and to examine the performance and properties of
the novel TDC formulation in comparison to other controllers, where
including Hsia’s TDC formulation [12], Jin’s TDC formulation Δ
N(q, q̇, q̈) = [M(q) − M̄]q̈ + C(q, q̇)q̇ + G(q)
[17], an SMC with a switching signum function, and another
TDC formulation by Jin et al. [9], which is a type of TSM + F(q, q̇) + τ d . (3)
control, with experiments using a Programmable Universal Ma-
nipulator for Assembly(PUMA)-type robot manipulator. The
Note that N(q, q̇, q̈) includes all the uncertain and un-
enhancement was made with the adoption of a novel nonlin-
known dynamics including the friction and other disturbance
ear DED based on a “mass”–“nonlinear damper”–“nonlinear
torques.
spring” system, which makes it possible to utilize not only full
second-order linear dynamics but also certain types of nonlinear
second-order dynamics as a DED, unlike the controllers used in B. Approximation of Uncertainties by the Delayed
[26], [30], [32], [33], and a nonlinear residual energy absorber Dynamics
for correction of the TDE error based on a nonlinear integral The idea of TDE can be summarized as follows. Assuming
SM (full-order SM) expecting no reaching phase [34], [35], [37], that N(q, q̇, q̈) is piecewise continuous, for a sufficiently small
[38] and no chattering. time L, the value of N(q, q̇, q̈) at time t may be close to that of
JIN et al.: ROBUST CONTROL OF ROBOT MANIPULATORS USING INCLUSIVE AND ENHANCED TIME DELAY CONTROL 2143

N(q, q̇, q̈) at time t − L [10]–[12], [14], [15]: for uh − q̈ yields


 
N(q, q̇, q̈)(t) ∼
= N(q, q̇, q̈)(t−L ) . (4) uh(t) − q̈(t) = M̄−1 N(q, q̇, q̈)(t) − (τ (t−L ) − M̄q̈(t−L ) ) .
L is usually set to be the sampling period in digital implementa- (11)
tion. From the robot dynamics in (2), one can have the following Substitution of (6) into (11) yields
equality:  
uh(t) − q̈(t) = M̄−1 N(q, q̇, q̈)(t) − N(q, q̇, q̈)(t−L ) . (12)
N(q, q̇, q̈)(t−L ) = τ (t−L ) − M̄q̈(t−L ) . (5)
Δ
From (4) and (5), one can obtain the TDE of N(q, q̇, q̈), The RHS of (12) is defined as TDE error ε(=
N̂(q, q̇, q̈) as follows: M̄−1 [N(q, q̇, q̈)(t) . − N(q, q̇, q̈)(t−L ) ]). With the definition of
the TDE error, from (12), one can easily obtain the following
N̂(q, q̇, q̈)(t) = N(q, q̇, q̈)(t−L ) equation:
= τ (t−L ) − M̄q̈(t−L ) . (6) uh (t) = q̈(t) + ε(t) (13)
It is clear from (6) that, to obtain N̂(q, q̇, q̈), the most recent implying, with the application of TDE in (9), the robot dynamics
torque τ (t−L ) and acceleration q̈(t−L ) are all that needed in- (2) became a linear decoupled second-order dynamic system
stead of the complex robot dynamics, (3). As the estimation subject to ε, and the system can be controlled by designing the
time delay (L) becomes smaller, TDE in (6) becomes more auxiliary control signal uh [e.g.,(10)] [1]–[3], [41]. From (8)
accurate. and (10), it can be easily derived that

C. TDC With Linear DED (Hsia’s Formulation) uh (t) − q̈(t) = ë(t) + KD ė(t) + KP e(t) . (14)
1) Desired Error Dynamics (DED): The following linear Substituting (14) into (12) yields
second-order DED was suggested by Hsia and his colleagues 
and utilized for TDC [11], [12] ë(t) + KD ė(t) + KP e(t) = M̄−1 N(q, q̇, q̈)(t)
ë(t) + KD ė(t) + KP e(t) = 0 (7) 
− N(q, q̇, q̈)(t−L ) . (15)
where
Δ Because RHS of (15) is the TDE error, (15) becomes
e(t) = qd (t) − q(t) . (8)
ë(t) + KD ė(t) + KP e(t) = ε(t) (16)
KD (= diag(KD 11 , KD 22 , . . . , KD n n )) and KP (= diag
(KP 11 , KP 22 , . . . , KP n n )) can be designed to meet the which is the closed-loop resulting error dynamics of the robot
desired performance (e.g., settling time, damping ratio). Note manipulator under the Hsia’s formulation. The TDE error ε
that the DED, (7), is a decoupled dynamics. causes the resulting dynamics (16) to deviate from the DED in
2) TDC formulation: The TDC with the linear DED in (7) (7). With a sufficient small sampling time L, the ε is close to 0,
can be designed as follows: if N(q, q̇, q̈) is continuous; however, if N(q, q̇, q̈) is discon-
tinuous, ε cannot be ignored. For instance, at the velocity rever-
τ (t) = M̄uh(t) + τ (t−L ) − M̄q̈(t−L ) with (9)
   sal, a pulse-type sudden increase in position tracking error was
Time Delay Estimation found, because ε increased due to the sudden polarity change
of discontinuous joint friction (i.e., in this case, N(q, q̇, q̈) is
uh(t) = q̈d(t) + KD ė(t) + KP e(t) . (10)
discontinuous at the velocity reversal) [26].
The last two terms (τ (t−L ) − M̄q̈(t−L ) ), the TDE term, in
the right-hand side (RHS) of (9) compensate for the nonlinear III. INCLUSIVE AND ENHANCED FORMULATION OF TDC
robot dynamics; and the first term, M̄uh , in the RHS of (9)
injects DED to the closed-loop system (i.e., the robot manipu- A novel formulation of TDC will be proposed to enhance
lator under TDC). Thus, M̄uh is called the injection term. One the position tracking performance of TDC with a nonlinear
can see from (9) and (10) that TDC is simple in that it does DED and a correcting term reducing the TDE error due to
not require robot dynamics model and its parameters. Instead, discontinuous dynamics. It will be proved that the proposed
it utilizes recent past control input, τ (t−L ) , and acceleration, TDC formulation is inclusive by showing that each of the
q̈(t−L ) , to compensate for the robot dynamics (3). In this con- previous TDC formulations is a special case of the proposed
text, the rather recent MFC [39], [40] may be associated with the formulation.
Hsia’s TDC formulation (see Appendix C). Hereafter, the TDC
with linear DED, given in (9) and (10), will be called Hsia’s A. Proposed Nonlinear DED
formulation. We propose a nonlinear DED as below
3) Closed-Loop Dynamics and TDE Error: Subtracting
the RHS of (9) from the left-hand side (LHS) of (2), and solving ë(t) + KD sig(ė(t) )α + KP sig(e(t) )β = 0 (17)
2144 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO. 5, OCTOBER 2017

Fig. 1. Phase portrait of the proposed nonlinear error dynamics, ë + 20sig(ė)α + 100sig(e)β = 0, with various α and β. If α = 1.0 and β = 1.0 the
error dynamics becomes linear error dynamics. Dashed lines shown in the center plot showing the phase portrait with α = 95/100 and β = 95/105.

where The time derivative of V can be obtained as follows:


n  
sig(ė)α = [|ė1 |α 1 sgn(ė1 ), . . . , |ėn |α n sgn(ėn )]T (18) V̇(t) = ėi(t) ëi(t) + ėi(t) KP ii eβi(t)
i
sgn(ei(t) )
i=1
sig(e)β = [|e1 |β 1 sgn(e1 ), . . . , |en |β n sgn(en )]T . (19) 
n  
= ėi(t) ëi(t) + KP ii sig(ei(t) )β i
All elements of α (= [α1 , α2 , . . . , αn ]T ) and β (= i=1

[β1 , β2 , . . . , βn ]T ) are positive values [i.e., αi > 0, βi > 


n

= ėi(t) −KD ii sig(ėi(t) )α i
0 (i = 1, 2, . . . , n)]. Similar to (7), the novel DED in (17) is
i=1
also a decoupled dynamics. n  

If α = 1n and β = 1n , the proposed DED is identical to (7), = −KD ii ėi(t)
α i +1
. (21)
the conventional linear DED. Here 1n = [1, 1, . . . , 1]T ∈ Rn . i=1
If not, the DED represents an equivalent “mass”–“nonlinear
damper”– “nonlinear spring” model. Phase portraits of the DED Because V̇ ≡ 0 with (17) implies e ≡ 0 and ė ≡ 0, with
for several α and β show that the equilibrium points are either positive KD ii s and KP ii s, the equilibrium (e, ė) = (0, 0) of
stable node or stable focus, depending on α and β (see Fig. 1). (17) is globally asymptotically stable by LaSalle’s invariant set
One can check the stability of the DED given in (17) with the theorem.
construction of a Lyapunov function (V), the total energy of the
DED, as follows: B. Control Law Design
1) TDC With Nonlinear DED: Similar to the Hsia’s for-
n 
  mulation, control torque can be designed as follows:
V(t) = 0.5ė2i(t) + KP ii (βi + 1)−1 ei(t)
β i +1
. (20)
i=1 τ (t) = M̄un (t) + N̂(t) (22)
JIN et al.: ROBUST CONTROL OF ROBOT MANIPULATORS USING INCLUSIVE AND ENHANCED TIME DELAY CONTROL 2145

where Based on this idea, to suppress the TDE error (ε), a correcting
term M̄λ · sig(s)γ is added to the control input in (22), as
un (t) = q̈d(t) + KD sig(ė(t) )α + KP sig(e(t) )β . (23) 
τ (t) = M̄ un (t) + λ · sig(s(t) )γ + τ (t−L ) − M̄q̈(t−L )

In comparison with uh of the Hsia’s formulation in (10), = M̄ q̈d(t) + KD sig(ė(t) )α + KP sig(e(t) )β
the proposed control uses un to realize the proposed nonlinear

DED. Thus, with the combination of (6), (22), and (23), the
+ λ · sig(s(t) )γ
control law is expressed as follows:
  + τ (t−L ) − M̄q̈(t−L ) . (31)
τ (t) = M̄ q̈d(t) + KD sig(ė(t) )α + KP sig(e(t) )β + N̂(t) .
The λ is the coefficient of nonlinear damping for the ideal
(24)
velocity feedback [see (30)]. The damping term λ · sig(s)γ ab-
Substitution of the control input (24) into robot dynamics (2),
sorbs the residual energy due to the TDE error as a counteracting
and simple manipulations yield the closed-loop dynamics
force while reducing the difference between the real and the ideal
velocity (i.e., making the behavior of the closed-loop dynamics
ë(t) + KD sig(ė(t) )α + KP sig(e(t) )β = ε(t) (25)
close to that of the DED).
Substituting the control input (31) into the robot dynamics
indicating that the closed-loop dynamics is perturbed by ε. As
(2), we obtain the closed-loop dynamics as
mentioned in Section II-C-3, the TDE error ε can be increased
with discontinuity in N(q, q̇, q̈). Thus, this control may need a ṡ(t) + λ · sig(s(t) )γ = ε(t) . (32)
term suppressing the effect of ε on position error.
Because, in (32), ṡ – the LHS of the DED [see (17) and
2) Addition of TDE Error Correction: If we define
(26)] – can be regarded as the output of a stable first-order filter
   with input ε [42], it is expected that the effect of ε on ṡ (and
ë + KD sig(ė)α + KP sig(e)β dt
Δ
s= (26) consequently on tracking error, e) will be attenuated compared
with (27).
having a zero initial value (i.e., s(t = 0) = 0), then, one The ε is bounded (i.e., ||ε|| < ξ), if M̄ satisfies the well-
can easily find that the closed-loop dynamics (25) can be known stability condition of TDC [11], [17], [26] given below
 
rewritten as I − M−1 M̄ < 1 (33)

ṡ(t) = ε(t) . (27) because N(q, q̇, q̈) is the sum of continuous terms and bounded
discontinuous terms [9], [26]. The stability criterion can be
Again, one can see that the closed-loop error dynamics is obtained easily by following the stability analysis in [9], [17],
disturbed by the TDE error (ε). and [26].
If the DED in (17) is perfectly realized, one can obtain the
following relation from (17): C. Discussion of the Proposed TDC Formulation
1) Properties of DED: We can further examine the non-
q̈ideal(t) = q̈d(t) + KD sig(ė(t) )α + KP sig(e(t) )β . (28) linear DED (17), which includes the linear DED (7).
Finite-time stable DED can be obtained, if the follow-
Note that because (28) holds only when ε is 0, the LHS angu- ing conditions (proof in Appendix A) are met for αi and
lar acceleration term is not the same as the real acceleration (q̈), βi (i = 1, 2, . . . , n) [43]
but it is an idealized case acceleration. Thus, to avoid confu-
sion, for the idealized case acceleration, we use q̈ideal denoting 0 < αi < 1, and (34)
ideal acceleration. By integrating both sides of (28), one can βi = αi / (2 − αi ) . (35)
obtain the ideal velocity (q̇ideal ), representing the velocity of the
closed-loop system that perfectly follows the DED, as follows: For instance, if we set αi = 95/100, β i that satisfies (35) is
95/105. With these values, one can see that the finite-time con-
   vergence of the nonlinear DED is achieved (see Fig. 2) [43].
q̇ideal = q̈d + KD sig(ė)α + KP sig(e)β dt. (29) The convergence speed of the finite-time stable nonlinear DED
is faster than that of the asymptotically stable linear DED re-
By comparing (29) with (26), one can see that the following quiring infinite time for the convergence of error (e) to 0 (see
holds: Fig. 2). The phase portrait of the DED with αi = 95/100 and
βi = 95/105 is slightly twisted compared with that of the lin-
s(t) = q̇ideal(t) − q̇(t) . (30) ear DED (center plot of Fig. 1). Moreover, for αi < 1 and
βi < 1 (i = 1, 2, . . . , n), the nonlinear damping and spring
Thus, from (30), one implication of the sliding variable s terms (KD sig(ė)α and KP sig(e)β ) are more effective than
is the difference between the ideal velocity and the real robot the linear damping and spring terms – in the linear DED –
manipulator velocity. around the neighborhood of the equilibrium point (i.e., |ėi | < 1
2146 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO. 5, OCTOBER 2017

theoretically, si (i = 1, 2, . . . , n) converges to the sliding sur-


face (see Appendix B), provided that there is no chattering [9].
If all γ i are positive constants, with Vs , each element of s is
bounded as follows (see Appendix B)
1
|si | < (ξ/ λm in ) γ i , (i = 1, 2, . . . , n) (38)

indicating that increasing λmin can reduce the effect of the TDE
error, ξ. In particular, if (37) – the sliding condition when γ = 0
– holds, the bound for si in (38) becomes smaller than unity.
If all elements of γ are in between zero and unity (0 < γi <
1; i = 1, 2, . . . , n), the s dynamics becomes a TSM [44]. If
(37) holds and 0 < γi < 1, as shown in (38), decreasing γi can
make the bound of |si | smaller without increasing the coefficient
of the nonlinear correcting term (i.e., λmin ). In other words,
ξ can be further attenuated with the decrease of γi s, making
Fig. 2. Comparisons of behaviors of DED. the correcting term practically valuable. In addition, the sig(•)γ
can effectively increase the effect of the correcting term near the
sliding surface (|si | < 1) because the fractional power of |si |
substantially increases near the region. For instance, if γi = 0.4
and |ei | < 1). The fractional power terms (|ėi |α i and |ei |β i ) and si = 0.0001, sig(si )γ i = 0.0251. Thus, λ needs to satisfy
amplify ėi and ei significantly in that neighborhood without (37) to take full advantage of the TSM.
increasing KD and KP compared with the linear feedback in If γ = 1n , the s dynamics becomes linear. If all elements
(10). This property can avoid sticking of the system states near of γ are larger than unity (γi > 1; i = 1, 2, . . . , n), the s dy-
the ei = 0 and/or ėi = 0, helping the system reach the sliding namics converges fast when |si | > 1, and converges slowly
surface. when |si | < 1 (i = 1, 2, . . . , n). Therefore, various reaching
If αi > 1 and βi > 1 (i = 1, 2, . . . , n), the convergence conditions can be made by changing γ.
speed of the DED becomes slower than that of the linear DED Note that, because si (i = 1, 2, . . . , n) is bounded with the
(see dash-dot line in Fig. 2). bound given in (38) and has zero initial condition, si will stay
Note that, even with α = 0 and β = 0, the DED is stable (see in the boundary from the initial time instance. In other words,
Section III-A). However, the “nonlinear damping” and “non- practically, the initial deviation of the s from the sliding surface
linear spring” can induce chattering, because those two terms (s = 0) does not depend on the initial condition unlike other
become KD sgn(ė) and KP sgn(e), respectively. controllers [26], [30], [32], [33] that have a reaching phase and
2) Terms of the Proposed TDC: Similarly to the Hsia’s may not be insensitive to parameter variations and disturbances
formulation [in Section II-C-2)], in the proposed control in (31), during the phase [34]–[37].
τ(t−L ) − M̄q̈(t−L ) is the TDE term; M̄un (t) is the injection 4) Inclusiveness of the Proposed TDC: The proposed
term [i.e., injecting DED in (17)]. For the proposed control, formulation of TDC can include many other existing TDC for-
there is one more term: the correcting term (i.e., M̄λ · sig(s)γ ) mulations.
that corrects/suppresses the TDE error by utilizing the ideal 1) If λ = 0, and α = β = 1n , the proposed control be-
velocity feedback. comes the Hsia’s formulation [12] in (9) and (10).
The elements of the proposed controller, given in (23) and 2) If all diagonal elements of λ have positive values (λii >
(31), have their own clear physical meaning. The DED is de- 0; i = 1, 2, . . . , n) and α = β = γ = 1n , the proposed
scribed by a “mass”–”nonlinear damper”–“nonlinear spring” control becomes the Jin’s position control formulation of
system with the nonlinear damping KD sig(ė)α and the nonlin- TDC in [17]
ear spring KP sig(e)β . The TDE error is corrected by another
nonlinear damping term (λ · sig(s)γ (= λ · sig(q̇ideal − q̇)γ ) 
τ (t) = M̄ q̈d(t) + KD ė(t) + KP e(t) + λs(t) + N̂(t) .
to achieve DED. (39)
3) Closed-Loop Dynamics Change With γ: One can ex- 3) If all diagonal elements of λ have positive values (λii >
amine the resulting closed-loop dynamics (32) with various γ 0; i = 1, 2, . . . , n), α = β = 1n , and γ = 0, the pro-
values. First, because sig(s)γ is continuous for γ = 0, there is posed control becomes an SMC with integral sliding sur-
no chattering unlike the sgn(s). face and TDE, most probably having chattering due to
If we let γ = 0, the closed-loop dynamics becomes the signum function, as
ṡ(t) + λ · sgn(s(t) ) = ε(t) . (36) 
τ (t) = M̄ q̈d(t) + KD ė(t) + KP e(t)
For (36), with the Lyapunov function Vs (= 0.5sT s), if 
+ λ · sgn(s(t) ) + N̂(t) . (40)
λmin > ξ (37)
JIN et al.: ROBUST CONTROL OF ROBOT MANIPULATORS USING INCLUSIVE AND ENHANCED TIME DELAY CONTROL 2147

4) If all diagonal elements of λ have positive values (λii >


0; i = 1, 2, . . . , n), α = β = 1n , and all elements of γ
are in between 0 and 1 (0 < γi < 1; i = 1, 2, . . . , n),
the proposed control becomes another TDC formulation
of Jin (a type of TSM control) [9]:

τ (t) = M̄ q̈d(t) + KD ė(t) + KP e(t)

+ λ · sig(s(t) )γ + N̂(t) . (41)

Thus, the proposed TDC formulation is indeed inclusive.

D. Low-Pass Filtering With Lowering of M̄


To reduce the noise due to acceleration and other signals, a Fig. 3. PUMA-type robot manipulator (Samsung Faraman-AT2).
low-pass filter may be needed. A possible candidate is a first-
order digital low-pass filter with the cutoff frequency ω. With
the application of the filter, (24) changes as follows:

τ ∗(t) = M̄ un (t) + λ · sig(s(t) )γ − q̈(t−L ) + τ (t−L ) , and
(42)
τ (t) = [ω /(1 + ω )] τ ∗(t) + [1/(1 + ω )] τ (t−L ) (ω = ωL)
(43)
where τ ∗ denotes input to the filter. Substituting (42) into (43)
yields

τ (t) = [ω / (1 + ω )] M̄ un (t) + λ · sig(s(t) )γ

− q̈(t−L ) + τ (t−L ) . (44)

Thus, lowering of M̄ is equivalent to the use of a low-pass Fig. 4. Desired trajectory (four fifth-order polynomial functions of time
in series) of joint 1, 2, and 3.
filter. Further, there is no additional dynamics introduced with
the low-pass filter.
compare the experimental results of this study (see Section IV)
E. Tuning Procedure with many previous studies using the same KD and KP at least
qualitatively [9], [11], [13], [26], [29].
To implement the proposed controller, we need to tune M̄,
KD , KP , λ, α, β, and γ. Note that L is usually set to be the
sampling time for discrete time domain implementation. The IV. EXPERIMENTAL STUDIES
gains can be tuned through the following procedure. A. Experimental Setup
1) Set α = β = γ = 1n and λ = 0, design a stable linear
A Samsung Faraman-AT2 robot manipulator is used in the ex-
DED by specifying KD and KP , and tune the M̄ only
periment (see Fig. 3). The maximum continuous torques were
(i.e., tune Hsia’s formulation [12]) by increasing diagonal
0.637, 0.637, and 0.319 Nm for joint 1, 2, and 3, respectively.
elements of the M̄ from a small value until there is no
The gear-reduction ratio and the encoder resolution of all those
decrease in error with vibrations (and noisy sound).
three joints were 120:1 and 2048 pulses/rev, respectively. Res-
2) Maintain γ = 1n and tune λ by increasing its diagonal
olution of each robot joint was 3.66 × 10−4◦ . Angular velocity
elements from 0, while checking the control performance
and acceleration of the three joints were computed by a simple
(e.g., error magnitude and unwanted vibrations).
numerical differentiation [12]. The controller was operated in
3) Change the linear reaching condition to the TSM by de-
Linux-RTAI, a real-time operating system environment. Sam-
creasing positive γi from 1 (i = 1, 2, . . . , n) with the
pling time was 0.001 s, and, accordingly, the time delay L was
observation of error magnitude and unwanted vibrations.
set to be 0.001 s.
4) Twist the DED by decreasing αi and βi from 1 (i =
1, 2, . . . , n). For the sake of simplicity, βi can be calcu-
B. Comparison of Different Parameter Sets
lated from (35). If vibrations are observed, then one stops
decreasing elements of the exponent vectors. Experimental comparisons were made using the robot with
Remark: For the experimental studies in Section IV, KD = various parameter sets of the proposed control. The desired
20 • I and KP = 100 • I – locating double pole at –10 for each trajectory of joint 1, 2, and 3 consists of four fifth-order poly-
DOF for linear DED – were used. This selection allows us to nomial functions of time (see Fig. 4). The parameters of the
2148 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO. 5, OCTOBER 2017

TABLE II
RMS ERROR WITH VARIOUS PARAMETER SETS

No. KD i i KP i i λi i αi βi γi RMS Error (%RMS Differencea ) Remark

1st joint 2nd joint 3rd joint

Exp01 20 100 0 1 1 – 0.0139° (1290%) 0.0377° (2118%) 0.0478° (2712%) [12]


Exp02 20 100 20 1 1 1 0.0045° (350%) 0.0100° (488%) 0.0135° (694%) [17]
Exp03 20 100 20 1 1 0.8 0.0029° (190%) 0.0065° (282%) 0.0082° (382%)
Exp04 20 100 20 1 1 0.6 0.0019° (90%) 0.0038° (124%) 0.0048° (182%)
Exp05 20 100 20 1 1 0.4 0.0017° (70%) 0.0026° (53%) 0.0027° (59%) [9]
Exp06 20 100 20 98/100 98/102 0.4 0.0014° (40%) 0.0021° (24%) 0.0023° (35%)
Exp07 20 100 20 98/100 98/102 1 0.0043° (330%) 0.0105° (518%) 0.0131° (671%)
Exp08 20 100 20 95/100 95/105 0.4 0.0010° (0%) 0.0017° (0%) 0.0017° (0%) ∗ smallest
Exp09 20 100 0 95/100 95/105 – 0.0082° (720%) 0.0238° (1300%) 0.0294° (1629%)
Exp10 20 100 20 95/100 95/105 0.8 0.0023° (130%) 0.0055° (227%) 0.0072° (324%)
Exp11 20 100 20 1.2 1.2 1 0.0094° (840%) 0.0219° (1188%) 0.0271° (1494%)
Exp12 20 100 0 1.2 1.2 – 0.0478° (4680%) 0.1171° (6788%) 0.1287° (7471%)
Exp13 20 100 0 1.5 1.5 – 0.2073° (1973%) 0.4018° (23535%) 0.4129° (24188%)

a
% RMS Difference = ((RMS Error of ith EXP) − (RMS Error of 8th EXP))/(RMS Error of 8th EXP) × 100

three joints, and the root-mean-square (RMS) value of those


three joints’ position tracking error and the %RMS difference
(defined in Table II) are listed in Table II.
By following the tuning procedure described in Section III-E,
the proposed controller was tuned experimentally as fol-
lows. First, M̄ was tuned and set to be diag(0.3818,
0.30545, 0.15296) kg · m2 (dotted line in Fig. 5).
Second, the diagonal elements of λ increased (dashed line
in Fig. 5). Comparison of the RMS error of Exp01 [Hsia’s
formulation in (9) and (10)] and that of Exp02 [Jin’s formulation
in (39)] shows the effectiveness of the correcting term, and the
tendency agrees with the free space motion control performance
reported in [17].
Third, γi s decreased from 1. Examination of RMS errors
obtained from Exp02 to Exp05 (see Table II) shows that, with
the gradual reduction of γi s from 1, the RMS error decreased.
If γi < 0.4 (i = 1, 2, . . . , n), however, vibration with noisy
sound was observed without further decrease of the RMS error
(dash-dotted line showing the results of Exp05 in Fig. 5).
Fourth, we slightly twisted the DED with the tuning of α
and β, and obtained better tracking performance (by com-
paring Exp05, Exp06, and Exp08). With the parameters (λ =
diag(20, 20, 20), α = 95/100 • 13 , β = 95/105 • 13 , and γ =
0.4 • 13 ) guaranteeing the finite-time stable terminal sliding
DED and terminal sliding correcting term (Exp08; solid line
in Fig. 5), the smallest RMS error was observed. A severe vi-
bration was observed with too small αi and βi (see Fig. 6). Thus,
the exponents α, β, and γ need careful tuning.
Additional experiments Exp11–13 show that, if αi > 1 and
βi > 1 (i = 1, 2, . . . , n), the RMS error increased substan-
tially with the small increase of exponents. Fig. 5. Comparison of tracking errors with various parameters. Exp01,
One of the Jin’s formulation [9] in (41), which used lin- the Hsia’s formulation (dotted line): λ = 0, α = 13 and β = 13 ; Exp02
(dashed line): λ = diag(20, 20, 20), α = 13 , β = 13 , and γ = 13 ; Exp05
ear DED with the nonlinear correcting term, achieved the 2nd (dash-dotted line): λ = diag(20, 20, 20), α = 13 , β = 13 , and γ = 0.4 •
smallest RMS error. Even compared with the RMS error of one 13 ; and Exp08 (solid line): λ = diag(20, 20, 20), α = 95/100 • 13 , β =
of the Jin’s formulation [9], a special case of the proposed for- 95/105 • 13 , and γ = 0.4 • 13 . The proposed formulation is a superset
of the existing controls [9], [12], [17]. If we properly twist the error dy-
mulation, the proposed control achieved significant decrease of namics and the correcting term with tuning of α, β, and γ, the tracking
RMS error (more than 53%; Table II), indicating substantial performance can be substantially improved.
JIN et al.: ROBUST CONTROL OF ROBOT MANIPULATORS USING INCLUSIVE AND ENHANCED TIME DELAY CONTROL 2149

tions [9]–[12], [14], [15], [17], [27] can be similarly applied


to the proposed formulation, because the proposed formulation
includes all those controllers.
Seemingly, there are two more parameters (α and β) to
tune for the proposed control compared with the formulation
achieved the 2nd smallest RMS error, but, with the finite-time
stable DED condition [(34) and (35)] relating α and β, the
number of additional parameters becomes one (e.g., α) with the
limited parameter space, (0, 1).
Although the tuning procedure given in Section III-E may
seem not very solid, it took less than an hour for the gain tuning
of the PUMA-type robot, showing the practicality of the pro-
cedure. It would, however, be worth to develop a more solid
procedure as a further study.
The RMS errors of many different TDC formulations are
given in Table II. Thus, depending on the required precision/
accuracy to complete a task given to a robot that one wants to
use, one may be able to choose a suitable controller for the task
with Table II. In other words, these promising results with the
Fig. 6. Tracking error with too small α and β (dashed line) and with proposed control do not necessarily mean we always have to
those achieved the smallest RMS error (solid line). Decreasing α and β use the proposed control with twisted nonlinear DED. In stiff-
below a certain value could induce vibration.
ness control or in impedance control (though these are not the
position control), the stiffness terms are desired variables of ref-
erence model (typically chosen as linear and decoupled ones)
improvement. At the velocity reversal (3 and 9 s), position error and are not gains of the controller [17]. The TDE error cor-
increased due to the rapidly changing friction (see Fig. 5). recting term may not be required if discontinuities of nonlinear
One can easily see that the control performance with the terms are not dominant (e.g., with the use of magnetic bearing).
nonlinear DED (Exp09) was improved in comparison with that Because of its inclusiveness and simplicity, the proposed for-
with the linear DED (Exp01), showing the usefulness of the mulation may be used for the practical precise position control
nonlinear DED. of robot manipulators. Further, the proposed formulation might
be extended to stiffness control, impedance control, and force
C. Robustness Against the Increase of Speed control for robot manipulators.
Robustness against the increase of speed was tested without V. CONCLUSION
any further fine tuning of gains from those used in Exp08. With
the 50% increase in speed (reaching the robot’s limit), RMS The inclusive and enhanced TDC formulation proposed in
error of each joint was maintained at a reasonable magnitude (1st this paper inherits the advantages of TDC: it is simple, intuitive,
joint: 0.0017◦ ; 2nd joint: 0.0025◦ ; 3rd joint: 0.0030◦ ), indicating and model-free.
the robustness of the proposed control. It also shows that the fine The proposed control consists of three terms: a continuous
tuning of gains for each different trajectory is not necessary. nonlinearity canceling term using TDE, a DED injecting term,
and a TDE error correcting term. The nonlinear DED describes
a ”mass”–“nonlinear damper”–“nonlinear spring” system. The
D. Discussion residual energy due to the perturbation of the TDE error can be
As shown in (31), the proposed inclusive and enhanced TDC absorbed by a nonlinear damper utilizing ideal velocity feedback
does not use the knowledge of M(q) and N(q, q̇, q̈) – which based on an integral SM.
includes uncertainties in inertial parameters (M(q) − M̄), Cori- The formulation of proposed control is inclusive: with certain
olis and centrifugal torque terms (C(q, q̇)q̇), and nonlinear joint parameters, the proposed formulation becomes Hsia’s formula-
friction disturbance (F(q, q̇)) – of the robot being controlled. tion, Jin’s formulations including a type of TSM control, and
The robot we used is a PUMA-type robot having substantial an SMC with switching signum function. The position tracking
amount of highly nonlinear (even at low velocity), uncertain, performance is improved significantly by properly twisting the
and time-, velocity-, and temperature-varying joint friction dis- DED and the correcting dynamics by changing α, β, and γ.
turbance [22]–[25], and coupled and highly nonlinear dynamics. Thus, a near future work might be adaptive tuning of α, β, and
Thus, the aforementioned experimental results proved the ro- γ to increase its practicality. It might be possible to extend the
bustness against parameter variation and unknown disturbances, proposed TDC formulation to the stiffness control, impedance
including one of the most often encountered friction distur- control, and force control considering the simplicity and in-
bance, to some extent. The results of analysis, simulations, and clusiveness of the proposed formulation, which is an on-going
experiments about the robustness of previous TDC formula- research in authors’ laboratory.
2150 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO. 5, OCTOBER 2017

APPENDIX B. Analysis of the s Dynamics in (32)


A. Condition for Finite-Time DED For (32), the time derivative of Vs can be obtained as follows:
Lemma 2 of [43] may be summarized as follows, the following V̇s = sT [−λ · sig(s)γ + ε]
control law

n
= [−λi si sig(si )γ i + si εi ]
u = − [l1 sig(x1 )ρ 1 + l2 sig(x2 )ρ 2 ] with (45)
i=1
0 < ρ1 < 1, and (46) 
n
= [−λi |si | |si |γ i + si εi ]
ρ2 = 2ρ1 / (1 + ρ1 ) . (47)
i=1

Finite-time stabilizes the following double integrator 


n

system ≤ [−λm in |si | |si |γ i + |si | ξ]


 i=1
ẋ1 = x2 
n
(48)
ẋ2 = u. = |si | [−λm in |si |γ i + ξ]. (57)
i=1
Substituting u in (45) into (48) yield
Thus, if

ẋ1 = x2 |si |γ i > ξ/ λm in , (i = 1, 2, . . . , n) (58)
ρ1 ρ2
. (49)
ẋ2 = − [l1 sig(x1 ) + l2 sig(x2 ) ] then V̇s < 0.
By using the lemma, we will show that, if (34) and (35) are If γ = 0, (57) becomes

met, the desired decoupled error dynamics, (17), is finite-time V̇s ≤ |si | [−λm in + ξ]. (59)
stable. The DED in (17) can be rewritten as below:
⎧ Therefore, if (37) is satisfied, V̇s < 0 outside of s = 0.
⎨ė1 = e2 , (e1 = e) If all γi are positive constants, from (58), the bound of si ,
  (50)
⎩ė2 = − KP sig(e1 )β + KD sig(e2 )α (38), can be directly obtained.

where e1 denotes the position error e; and e2 the time derivative C. Hsia’s TDC Formulation and MFC
of e1 (i.e., velocity error). Because (50) is a decoupled dynamics, MFC employs a simple and effective estimation scheme to
for each DOF, one can rewrite (50) as follows: estimate poorly known parts of the plant as well as of the vari-
 ous possible disturbances [39], [40]. Because the objective and
ė1i = e2i , (e1i = ei ) function of the TDE of TDC is similar to those of the MFC’s
ė2i = −[KP ii sig(e1i )β i + KD ii sig(e2i )α i ], (i = 1, 2, . . . , n) estimation method, the two schemes are compared. We restrict
(51) ourselves, for simplicity, to single-input single-output robot ma-
where e1i and e2i denote ith element of e1 and e2 , respectively. nipulators.
By comparing (49) with (51), one can see that, if KP ii = l1 , MFC may be summarized as follows. Instead of the plant
KD ii = l2 , then model, MFC adopts the continuously updated ultralocal model,
which is valid only for very short time interval, as below
βi = ρ1 , and (52)
q (ν ) = F + στ (60)
αi = ρ2 (53)
where σ denotes a nonphysical constant chosen by the practi-
and (51) is finite-time stable. tioner; and F a time-varying function that includes poorly known
With simple mathematical manipulations, (46) and (47) can parts of the plant as well as of the various possible disturbances.
be rewritten as follows: Integer ν ≥ 1 may always be chosen as 1 or 2 based on the
selection of the practitioner [39], [40]. With an estimate of F, φ,
0 < 2ρ1 / (1 + ρ1 ) < 1, and (54) MFC may be designed as follows:
  
ρ1 = ρ2 /(2 − ρ2 ). (55)
τ = −σ −1 φ − qd − KD M ė − KP M e − KI M
(ν )
edt .
From (54), it is obvious that (61)
Here KD M , KP M , and KI M denote gains that are tuned
0 < ρ2 < 1. (56)
based on the desired convergent behavior of error, and the choice
Thus, (46) and (47) are equivalent to (55) and (56). Therefore, of ν [39], [40]. With (61), the closed-loop dynamics becomes

from (52), (53), (55), and (56), for the exponents αi and βi , the
condition that (51) is finite-time stable is 0 < αi < 1, and e(ν ) + KD M ė + KP M e + KI M edt = φ − F. (62)
βi = αi / (2 − αi ) as given in (34) and (35). 
JIN et al.: ROBUST CONTROL OF ROBOT MANIPULATORS USING INCLUSIVE AND ENHANCED TIME DELAY CONTROL 2151

From (60), in discrete time domain, with the assumption of [5] A. Ishiguro, T. Furuhashi, S. Okuma, and Y. Uchikawa, “A neural network
sufficiently small time interval, φ is estimated as follows [39]: compensator for uncertainties of robotics manipulators,” IEEE Trans. Ind.
Electron., vol. 39, no. 6, pp. 565–570, Dec. 1992.
φ(k) = q̂ (ν ) (k) − στ (k − 1) (63) [6] M. J. Er and Y. Gao, “Robust adaptive control of robot manipulators using
generalized fuzzy neural networks,” IEEE Trans. Ind. Electron., vol. 50,
no. 3, pp. 620–628, Jun. 2003.
where q̂ (ν ) (k) denotes the estimate of q (ν ) at the time instant [7] C.-K. Lin, “Nonsingular terminal sliding mode control of robot manipu-
k. Later, in [40], to avoid the use of noisy q̂ (ν ) (k), an algebraic lators using fuzzy wavelet networks,” IEEE Trans. Fuzzy Syst., vol. 14,
identification technique (iterated time integrals during short time no. 6, pp. 849–859, Dec. 2006.
[8] S. Han and J. M. Lee, “Fuzzy echo state neural networks and funnel
interval) is proposed. For instance, if ν = 1, for sufficiently dynamic surface control for prescribed performance of a nonlinear dy-
small te , φ was computed as follows: namic system,” IEEE Trans. Ind. Electron., vol. 61, no. 2, pp. 1099–1112,
 t Feb. 2014.
[9] M. Jin, Y. Jin, P. H. Chang, and C. Choi, “High-accuracy tracking control
φ(t) = −6t−3 e [(te − 2η) q (η) + ση (te − η) τ (η)] dη. of robot manipulators using time delay estimation and terminal sliding
t−t e mode,” Int. J. Adv. Robot. Syst., vol. 8, pp. 65–78, 2011.
(64) [10] T. C. S. Hsia, “A new technique for robust control of servo systems,” IEEE
From this summary, on the one hand, one can see that MFC Trans. Ind. Electron., vol. 36, no. 1, pp. 1–7, Feb. 1989.
differs from TDC in that MFC does not need system order [40]. [11] T. C. S. Hsia and L. S. Gao, “Robot manipulator control using
decentralized linear time-invariant time-delayed joint controllers,” in
On the other hand, TDC is similar to MFC when ν = 2. One Proc. Int. Conf. IEEE Robot. Autom., Cincinnati, OH, USA, 1990,
can rewrite (2) as follows: pp. 2070–2075.
[12] T. C. S. Hsia, T. A. Lasky, and Z. Guo, “Robust independent joint controller
q̈ = −M̄ −1 N (q, q̇, q̈) + M̄ −1 τ. (65) design for industrial robot manipulators,” IEEE Trans. Ind. Electron.,
vol. 38, no. 1, pp. 21–25, Feb. 1991.
If ν = 2 in (60), from (60) and (65), one can see that F and [13] T. C. S. Hsia and S. Jung, “A simple alternative to neural network control
α in (60) amount to −M̄ −1 N (q, q̇, q̈) and M̄ −1 , respectively. scheme for robot manipulators,” IEEE Trans. Ind. Electron., vol. 42, no. 4,
pp. 414–416, Aug. 1995.
Multiplying −M̄ −1 on both sides of (6) yields [14] K. Youcef-Toumi and O. Ito, “A time delay controller for systems with
unknown dynamics,” in Proc. Amer. Conf. Control, Atlanta, GA, USA,
−M̄ −1 N̂ (q, q̇, q̈)(t) = q̈(t−L ) − M̄ −1 τ(t−L ) (66) 1988, pp. 904–913.
[15] K. Youcef-Toumi and O. Ito, “A time delay controller for systems with
implying similarity between the MFC’s estimation scheme (63) unknown dynamics,” ASME J. Dyn. Syst. Meas. Control, vol. 112, pp. 133–
and TDE in (66) when ν = 2 with a difference on the use of 142, 1990.
[16] J. Kim, H. Joe, S. C. Yu, J. S. Lee, and M. Kim, “Time-delay controller
acceleration: q̈(t−L ) in TDE, q̈(t) in MFC’s estimation method. design for position control of autonomous underwater vehicle under dis-
Hsia’s formula, (9) and (10), can be rewritten as turbances,” IEEE Trans. Ind. Electron., vol. 63, no. 2, pp. 1052–1061,
Feb. 2016.
τ(t) = − (M̄ −1 )−1 [(q̈(t−L ) − M̄ −1 τ(t−L ) ) − q̈d(t) [17] M. Jin, S. H. Kang, and P. H. Chang, “Robust compliant motion control
of robot with nonlinear friction using time-delay estimation,” IEEE Trans.
− KD ė(t) − KP e(t) ]. (67) Ind. Electron., vol. 55, no. 1, pp. 258–269, Feb. 2008.
[18] K.-H. Kim, H.-S. Kim, and M.-J. Youn, “An improved stationary-frame-
When ν = 2, by comparing (61) and (67), one can again find based current control scheme for a permanent-magnet synchronous mo-
a similarity: q̈(t−L ) − M̄ −1 τ(t−L ) corresponds to φ, the esti- tor,” IEEE Trans. Ind. Electron., vol. 50, no. 5, pp. 1065–1068, Oct. 2003.
[19] Y.-X. Wang, D.-H. Yu, and Y.-B. Kim, “Robust time-delay control for
mate of F, and −q̈d(t) − KD ė(t) − KP e(t) to −q̈d − KD M ė − the DC–DC boost converter,” IEEE Trans. Ind. Electron., vol. 61, no. 9,
KP M e − KI M ∫ edt. If KI M = 0, the similarity becomes more pp. 4829–4837, Sep. 2014.
[20] S. Talole, A. Ghosh, and S. Phadke, “Proportional navigation guidance
evident. Furthermore, an alternative TDE method [45] utilizes using predictive and time delay control,” Control Eng. Practice, vol. 14,
convolution integrals instead of q̈(t−L ) to reduce noise, and, in pp. 1445–1453, 2006.
this sense, the alternative TDE method is similar to the algebraic [21] M. Jin and P. H. Chang, “Simple robust technique using time delay es-
timation for the control and synchronization of Lorenz systems,” Chaos
method (64). Soliton Fract., vol. 41, pp. 2672–2680, 2009.
Because MFC estimates and compensates for the unknown [22] B. Armstrong, “Friction: Experimental determination, modeling and com-
and uncertain dynamics and disturbance, such as the TDE error, pensation,” in Proc. Int. Conf. IEEE Robot. Autom., Philadelphia, PA,
USA, 1988, pp. 1422–1427.
MFC may have estimation error (i.e., φ − F ) that perturbs the [23] H. Schempf and D. R. Yoerger, “Study of dominant performance character-
error dynamics (62), unless φ ≡ F . Thus, it may be possible to istics in robot transmissions,” ASME J. Mech. Des., vol. 115, pp. 472–482,
reduce the effect of the estimation error (φ − F ) by applying the 1993.
[24] B. Armstrong-Hélouvry, P. Dupont, and C. C. De Wit, “A survey of mod-
well-developed TDE error correcting methods introduced in this els, analysis tools and compensation methods for the control of machines
study. Meanwhile, the idea of the MFC’s algebraic identification with friction,” Automatica, vol. 30, pp. 1083–1138, 1994.
method may benefit TDE in reducing noise. [25] M. Ruderman and M. Iwasaki, “Observer of nonlinear friction dynamics
for motion control,” IEEE Trans. Ind. Electron., vol. 62, no. 9, pp. 5941–
5949, Sep. 2015.
REFERENCES [26] M. Jin, J. Lee, P. H. Chang, and C. Choi, “Practical nonsingular termi-
nal sliding-mode control of robot manipulators for high-accuracy track-
[1] F. L. Lewis, D. M. Dawson, and C. T. Abdallah, Robot Manipulator ing control,” IEEE Trans. Ind. Electron., vol. 56, no. 9, pp. 3593–3601,
Control: Theory and Practice, 2nd ed. Boca Raton, FL, USA: CRC Press, Sep. 2009.
2003. [27] S.-U. Lee and P. H. Chang, “Control of a heavy-duty robotic excavator
[2] T. Yoshikawa, Foundations of Robotics: Analysis and Control. Cambridge, using time delay control with integral sliding surface,” Control Eng. Pract.,
MA, USA: MIT Press, 1990. vol. 10, pp. 697–711, 2002.
[3] J. J. Craig, Introduction to Robotics: Mechanics and Control, 3rd ed. [28] J. Lee, P. H. Chang, and R. S. Jamisola, “Relative impedance control
Upper Saddle River, NJ, USA: Prentice-Hall, 2005. for dual-arm robots performing asymmetric bimanual tasks,” IEEE Trans.
[4] F. L. Lewis, C. T. Abdallah, and D. M. Dawson, Control of Robot Manip- Ind. Electron., vol. 61, no. 7, pp. 3786–3796, Jul. 2014.
ulators. Oxford, U. K.: Macmillan, 1993.
2152 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 22, NO. 5, OCTOBER 2017

[29] G. R. Cho, P. H. Chang, S. H. Park, and M. Jin, “Robust tracking under Sang Hoon Kang (M’12) received the B.S. (cum
nonlinear friction using time-delay control with internal model,” IEEE laude), M.S., and Ph.D. degrees in mechanical
Trans. Control Syst. Technol., vol. 17, no. 6, pp. 1406–1414, Nov. 2009. engineering from the Korea Advanced Institute
[30] Y. Jin, P. H. Chang, M. Jin, and D. G. Gweon, “Stability guaranteed of Science and Technology, Daejeon, South Ko-
time-delay control of manipulators using nonlinear damping and terminal rea, in 2000, 2002, and 2009, respectively.
sliding mode,” IEEE Trans. Ind. Electron., vol. 60, no. 8, pp. 3304–3317, He was a Research Associate at the Sensory
Aug. 2013. Motor Performance Program of the Rehabilita-
[31] J. Y. Lee, M. Jin, and P. H. Chang, “Variable PID gain tuning method using tion Institute of Chicago, Chicago, IL, USA, and
backstepping control with time-delay estimation and nonlinear damping,” a Postdoctoral Research Fellow in the Depart-
IEEE Trans. Ind. Electron., vol. 61, no. 12, pp. 6975–6985, Dec. 2014. ment of Physical Medicine and Rehabilitation,
[32] M. Jin, J. Lee, and K. K. Ahn, “Continuous nonsingular terminal sliding- Northwestern University, Chicago, from 2010 to
mode control of shape memory alloy actuators using time delay esti- 2015; was an Instructor in the Department of Biomedical Engineering,
mation,” IEEE/ASME Trans. Mechatronics, vol. 20, no. 2, pp. 899–909, Northwestern University, Evanston, IL, in 2012 and 2013; and was an
Apr. 2015. Adjunct Assistant Professor in the Department of Physical Medicine and
[33] J. Lee, M. Jin, and K. K. Ahn, “Precise tracking control of shape memory Rehabilitation, Northwestern University, from 2015 to 2017. He is cur-
alloy actuator systems using hyperbolic tangential sliding mode control rently an Assistant Professor in the Department of System Design and
with time delay estimation,” Mechatronics, vol. 23, pp. 310–317, 2013. Control Engineering, Ulsan National Institute of Science and Technology,
[34] V. Utkin and S. Jingxin, “Integral sliding mode in systems operating under Ulsan, South Korea. His current research interests include rehabilitation
uncertainty conditions,” in Proc. 35th Int. Conf. IEEE Dec. Control, Kobe, robotics and biomechanics of human movement, with an emphasis on
Japan, 1996, pp. 4591–4596. rehabilitation medicine.
[35] F. Castanos and L. Fridman, “Robust design criteria for integral sliding
surfaces,” in Proc. Int. Conf. IEEE Dec. Control, Seville, Spain, 2005,
pp. 1976–1981.
[36] F. Harashima, H. Hashimoto, and S. Kondo, “MOSFET converter-fed po-
sition servo system with sliding mode control,” IEEE Trans. Ind. Electron.,
vol. IE-32, no. 3, pp. 238–244, Aug. 1985.
[37] V. Utkin, J. Guldner, and J. Shi, Sliding Mode Control in Electro-
mechanical Systems. London, U. K.: Taylor & Francis, 1999, vol. 34. Pyung Hun Chang (S’86–M’89) received the
[38] C. C. Chen, S. S. D. Xu, and Y. W. Liang, “Study of nonlinear integral M.S. and Ph.D. degrees in mechanical engineer-
sliding mode fault-tolerant control,” IEEE/ASME Trans. Mechatronics, ing from the Massachusetts Institute of Technol-
vol. 21, no. 2, pp. 1160–1168, Apr. 2016. ogy, Cambridge, MA, USA, in 1984 and 1987,
[39] M. Fliess and C. Join, “Model-free control and intelligent PID controllers: respectively.
Towards a possible trivialization of nonlinear control?,” in Proc. IFAC He was a Professor at the Korea Advanced In-
Symp. Syst. Ident., Saint-Malo, France, 2009, pp. 1531–1550. stitute of Science and Technology (KAIST), Dae-
[40] M. Fliess and C. Join, “Model-free control,” Int. J. Control, vol. 86, jeon, South Korea, for 24 years. He is currently a
pp. 2228–2252, 2013. Professor with the Department of Robotics Engi-
[41] J. Luh, M. Walker, and R. Paul, “Resolved-acceleration control of mechan- neering, Daegu Gyeongbuk Institute of Science
ical manipulators,” IEEE Trans. Autom. Control, vol. 25, no. 3, pp. 468– and Technology, Daegu, South Korea. He has
474, Jul. 1980. authored more than 80 papers in international journals, 22 papers in
[42] J.-J. E. Slotine, “The robust control of robot manipulators,” Int. J. Robot. domestic journals, and 90 papers in conference proceedings. His cur-
Res., vol. 4, pp. 49–64, 1985. rent research interests include rehabilitation robotics, robust control of
[43] Y. Hong, Y. Xu, and J. Huang, “Finite-time control for robot manipulators,” nonlinear systems including kinematically redundant manipulators, and
Syst. Control Lett., vol. 46, pp. 243–253, 2002. impedance control of dual-arm manipulators.
[44] S. Yu, X. Yu, B. Shirinzadeh, and Z. Man, “Continuous finite-time control
for robotic manipulators with terminal sliding mode,” Automatica, vol. 41,
pp. 1957–1964, 2005.
[45] K. Youcef-Toumi and S. Y. Huang, “Analysis of a time delay controller
based on convolutions,” in Proc. Amer. Conf. Control, San Francisco, CA,
USA, 1993, pp. 2582–2586.

Jinoh Lee (S’09–M’12) received the B.S.


(summa cum laude) degree in mechanical engi-
Maolin Jin (S’06–M’08) received the B.S. de- neering from Hanyang University, Seoul, South
gree in material science and mechanical engi- Korea, in 2003, and the M.Sc. and Ph.D. de-
neering from the Yanbian University of Science grees in mechanical engineering from the Korea
and Technology, Jilin, China, in 1999, and the Advanced Institute of Science and Technology,
M.S. and Ph.D. degrees in mechanical engineer- Daejeon, South Korea, in 2012.
ing from the Korea Advanced Institute of Science He was a Postdoctoral Researcher in the De-
and Technology (KAIST), Daejeon, South Korea, partment of Advanced Robotics (ADVR), Istituto
in 2004 and 2008, respectively. Italiano di Tecnologia (IIT), Genoa, Italy, from
In 2008, he was a Postdoctoral Researcher 2012 to 2017. He is currently a Research Sci-
at the Mechanical Engineering Research Insti- entist with the Department of ADVR, IIT. His research interests include
tute, KAIST. He was a Senior Researcher at the robotics and control engineering, which includes whole-body manipu-
Research Institute of Industrial Science and Technology, Pohang, South lation of humanoids, robust control of highly nonlinear systems, and
Korea. He is currently a Principal Researcher with the Korea Institute compliant robotic system control for safe human–robot interaction.
of Robot and Convergence (KIRO), Pohang. His research interests in- Dr. Lee has been participating as a Technical Committee Member of
clude robust control of nonlinear plants, time-delay control, robot motion the International Federation of Automatic Control, TC4.3 Robotics, since
control, electro-hydraulic actuators, winding machines, disaster robotics, 2014. He is also a member of the IEEE Robotics and Automation, Con-
and factory automation. trol Systems, and Industrial Electronics Societies.

You might also like