Designing Robust Force Control of

Despite System and Environmental Uncertainties

By Navid Niksefat
and Nariman Sepehri

T
his article presents the de-
sign of a robust force con-
troller for a hydraulic
actuator interacting with an
uncertain environment via
quantitative feedback the-
ory (QFT). After the derivation of a realistic
nonlinear differential equation model, a
linearized plant transfer function is devel-
oped. The effects of nonlinearities are ac-
counted for by describing the linearized
model parameters as structured uncer-
tainty. The impact of environmental vari-
ability as well as variations in hydraulic
component parameters are also included
as uncertainty in the model. The QFT de-
sign procedure is carried out to design a ro-
bust controller that satisfies performance
specifications for tracking and disturbance ©ABB Control Valves.
rejection. The designed controller enjoys
the simplicity of fixed-gain controllers, is
easy to implement, and at the same time is
robust to the variation of hydraulic func-
tions as well as environmental stiffness.
The controller is implemented on an industrial hydraulic ac- Introduction
tuator equipped with a low-cost proportional valve. The ex- Many industrial applications, such as manufacturing auto-
perimental results show that robust stability against system mation and material handling, involve interaction with the
uncertainties and under varying conditions is achieved and environment. In these applications, force control is re-
the performance goals are satisfied. quired. Hydraulic actuators are advantageous for such ap-

Niksefat and Sepehri (nariman@cc.umanitoba.ca) are with the Department of Mechanical Engineering, The University of Manitoba, Win-
nipeg, Manitoba, R3T 5V6 Canada.

0272-1708/01/$10.00©2001IEEE
66 IEEE Control Systems Magazine April 2001
plications because of their high force-to-weight ratio and
fast response time. Additionally, they are able to maintain
their loading capacity indefinitely, which would usually
cause excessive heat generation in electrical components
[1]. Unlike in electric actuators, however, force control in
hydraulic actuators is a difficult problem [2], [3].
In a hydraulic actuator, the control signal activates the
spool valve that controls the flow of hydraulic fluid into and
out of the actuator. This flow in turn causes a pressure dif-
ferential buildup that is proportional to the actuator force.
Even if the spool valve dynamics are ignored, the control
signal fundamentally controls the derivative of the actuator
force and not the force itself. Furthermore, hydraulic sys-
tems are highly nonlinear and subject to parameter uncer-
tainty; parameters change with time as a result of variations
in operating conditions and component degradation. For ex-
ample, the supply pressure is subject to variation that may
be generated by the operation of other actuators in a
multiuser environment [4]. The flow and pressure coeffi-
cients, characterizing fluid flow into and out of the valve, are
functions of load and supply pressure and can vary under
different operating conditions [5]. Also, the effective bulk
modulus in hydraulic systems can significantly change un-
der various load conditions, oil temperature, and air con-
tent in the oil [6]. Design of a controller in the face of such a
range of parameter variations and disturbances is challeng-
ing. This article presents the application of QFT to the de-
sign of a robust force controller for hydraulic actuators.
In the literature, several force control strategies have
been proposed for hydraulic actuators. Conrad and Jensen
[2] used combinations of velocity feedforward, output feed-
back, and a Luenberger observer with state estimate feed-
back for force control of a double-rod hydraulic actuator.
The simulation and experimental results for a constant
setpoint force showed superior performance of the pro-
posed method over conventional (P or PI) force feedback
controllers. However, the variations of load and supply
pressure were not considered in their study.
Chen et al. [7] designed a sliding-mode controller for a
single-rod hydraulic actuator interacting with a spring as an
environment. Using position, velocity, acceleration, force,
and pressure feedback, the variable-structure controller
proved to be capable in both static and dynamic force con-
trol tasks. The effect of servo-amplifier gain variation was
also examined; however, the effect of variations in environ-
mental stiffness was not studied.
Sun et al. [8] employed a sliding-mode controller with a
perturbation observer for a single-rod electrohydraulic sys-
tem. The effect of cylinder position and velocity on the pres-
sure dynamics was considered as perturbation to the
control model, which was estimated by an observer. The ex-
perimental results verified improved steady-state and tran-
sient performance as compared with traditional
proportional-integral-differential (PID) controllers.
Adaptive control strategies have also been considered
for hydraulic force control. Liu and Alleyne [9] developed a
April 2001 IEEE Control Systems Magazine
switching control scheme using a Lyapunov-based adaptive
law to reduce parametric uncertainty. The implementation
of the controller, which is based on the measurements of po-
sition, velocity, acceleration, pressure, and spool displace-
ment, showed good performance for high-frequency force/
pressure tracking.
Wu et al. [10] applied a generalized predictive control al-
gorithm to a hydraulic force control system. The controller
was experimentally evaluated for various environmental
stiffnesses and setpoints. The method, however, relies
heavily on online parameter estimation and consequently is
computationally expensive.
Laval et al. [11] used an H ∞ approach to robustly control
the force exerted by a double-acting symmetric hydraulic
cylinder with a servovalve. The importance of uncertainties
in the environment, measurement, and nonlinearities on the
performance of hydraulic force control systems was high-
lighted. Limited test results, demonstrating the achieve-
ment of a stability/performance trade-off utilizing an H ∞
approach, were presented.
Despite the existence of a great number of force control
concepts, methods, and algorithms, there is still a large gap
between theory and industrial practice. The reasons have
been ascribed to the poor industrial control architecture,
which does not allow the implementation of sophisticated
algorithms [12]. In this article, we employ the QFT tech-
nique to design an explicit force controller for an industrial
hydraulic actuator. The goal is to arrive at a fixed-gain con-
troller that: 1) is of low order and easy to implement, 2) is ro-
bust against uncertainties in both environmental stiffness
and actuator functions, and 3) does not require exact knowl-
edge of the system’s parameters.
QFT is a robust controller design methodology aimed at
plants with parametric and unstructured uncertainties. The
concept was first introduced by Horowitz in the early sixties
and was later refined by him and others into a technique
[13]-[15]. QFT emphasizes the fact that feedback is only nec-
essary because of uncertainty and that the amount of feed-
back should therefore be directly related to the extent of
plant uncertainty and unknown external disturbances. Mini-
mizing the cost of feedback, as measured by the amount of
controller bandwidth, is the main objective of QFT [16].
Therefore, the plant uncertainty and the closed-loop toler-
ances are formulated quantitatively so that the cost of feed-
back can be assessed at each stage of the design process
[17]. (See the sidebar “An Overview of QFT” on p. 68.) The
method has been applied to a wide range of engineering
problems, including flight control [18], robot position con-
trol [19], and manufacturing systems [20]. Regarding the ap-
plication of QFT to hydraulic systems, the technique was
used to control an electrohydrostatic actuator as part of a
flight control system by Pachter et al. [21]. The controller’s
efficacy was validated by simulation results. Thompson and
Kremer [22] developed a QFT controller for a variable-dis-
placement pump system based on a linearized model with
67
 An Overview of QFT

Q uantitative feedback theory (QFT) is a frequency domain method

intended for practical control system design given robust perfor-
mance specifications. Its goal is to design a low-bandwidth controller
that satisfies performance specifications, despite system uncertainties
and disturbances. A low-bandwidth controller is a key issue in any
practical design to avoid problems with noise amplification, resonance,
and unmodeled dynamics. A brief outline of the classical QFT problem
is presented here. A full treatment is given in [13] and [15]. Short over-
views can also be found in [28] and [29].
α 0 , yields the nominal plant P0 ( s ) = P ( s , α 0 ) . Many problems of
practical interest can be expressed by the above plant model with
parametric uncertainty.
QFT Design Problem
The QFT design problem is to synthesize a pair of strictly proper,
rational, stable functions G( s ) and F ( s ) such that the following
specifications are satisfied while the bandwidth of the controller
is kept as low as possible.

QFT Configuration Robust Stability: The closed-loop system

A two degree-of-freedom system is typically assumed for the QFT L( s ,α )

technique when only output Y and command input R can be T ( s ,α ) =
1 + L( s ,α )
measured independently. Plant P contains system uncertainties
and is exposed to unknown disturbance D. G is the controller, must be stable ∀ α ∈ Ω. L( s , α ) = P ( s , α ) G( s ) is the open-loop
which can help reduce the variation of the plant output due to transfer function.
uncertainties and disturbances, while F is used merely as a filter to
Robust Tracking: A tracking specification for the reference
tailor the response to meet the control system’ s specifications.
input is given as
The uncertain plant can be described as
m
Tl (ω ) ≤ |F (iω )T (iω ,α )| ≤ Tu (ω ) ω ∈ [0 , ∞ )
∑ p (α )s
i
i

P( s ,α ) = i=0
n
where the frequency response bounds, Tl (ω ) and Tu (ω ) , are
∑ q (α )s
i=0
i
i

specified by the designer.

where α ∈ Ω ⊂ R p is an uncertain parameter vector. Ω is a Disturbance Attenuation: The requirement for disturbance
compact set of parameter variations and may be given as rejection at plant output is expressed as

Ω = {α :α i ∈ [α i ,α i ] ,i = 1,… , p} max TD ( jω ,α ) ≤ M D (ω )
α ∈Ω

where each uncertain parameter α i ( i = 1,... , p ) varies indepen-

dently within an interval. Choosing a reference parameter vector, where

parametric uncertainty. The closed-loop frequency domain Electrohydraulic Actuator Modeling

specifications were satisfied, and the corresponding step
responses were found to be reasonable. The successful ap- Presentation of
plication of this method to a wide range of engineering prob- Nonlinear Dynamic Equations
lems provides evidence of its potential effectiveness for A schematic diagram of a hydraulic actuator in contact
with the environment is shown in Fig. 1. In this study, the
hydraulic systems. One objective of this study is to evaluate
well-known model of manipulator-sensor-environment
and demonstrate the benefits of QFT when applied to hy-
[23], [24] is coupled with the nonlinear hydraulic actua-
draulic force control systems.
tor dynamics. The sensor with stiffness ks and damping ds
The article is organized as follows. In the next section, we
connects the actuator piston, represented by the mass
derive a mathematical model of a hydraulic actuator in con-
ma , to the environment. The environment is represented
tact with the environment. This model is then used to con- by mass me , damping de , and stiffness ke . The dynamic
struct a parametrically uncertain linear transfer function to equations are
describe the system. The uncertainty reflects variation in
operating-point-dependent parameters, changes in the en- ma x

= fa − ds ( x
− x
e ) − dx
− ks ( x − x e ) (1)
vironment, and changes in hydraulic actuator’s functions.
Then, we use the QFT method to design a single fixed-gain
robust force controller for the above uncertain plant. Fol- me x

e = ds ( x
− x
e ) − de x
e + ks ( x − x e ) − ke x e (2)
lowing that, we demonstrate the controller’s performance
on an experimental hydraulic actuator test rig. Finally, we
present conclusions. f = ks ( x − x e ) (3)

68 IEEE Control Systems Magazine April 2001

 1 frequency to keep the controller bandwidth small. During this
TD ( s ,α ) =
1 + L( s ,α )
is the transfer function from the disturbance to the output and
MD (ω ) is the magnitude of disturbance rejection.
QFT Design Procedure
The design procedure involves the following steps.
1) Generating the plant templates: The region in the Nichols
chart occupied by the complex values of P ( jω , α ) at each
frequency is called the plant template (see the figure at right).
This essentially characterizes the plant uncertainty by capturing
the gain and phase variations of the plant at a given frequency.
The templates are used to generate bounds on the Nichols chart
for the controller design.
stage, the designer should effect a trade-off between conflicting
specifications, controller complexity, and the cost of the
feedback in the bandwidth. Once a satisfactory L0 ( s ) is arrived
at, the controller can be extracted from L0 ( s ) by dividing it by the
nominal plant transfer function P0 ( s ).
4) Design of prefilter: Design of a proper L0 ( s ) only guaran-
tees that the variation in the closed-loop transfer function,
T ( s , α ) , is less than or equal to that allowed. Therefore, a
prefilter is required to bring the response within the upper
and lower tolerances, Tl (ω ) and Tu (ω ) . The prefilter magni-
tude bound at each frequency can be calculated from the ro-
bust tracking specification.

2) Generating performance bounds: The design specifications

60 L0(jω1)
outlined above should be translated into certain bounds on the ω1
50
nominal open-loop transfer function, L0 ( s ) = P0 ( s ) G( s ) , to Plant Template at ω = ωi
40
reveal the trade-offs between performance specifications and

Open-Loop Gain (dB)

30 ω2
robustness at each frequency. These bounds are derived either
20
by moving the plant template between the closed-loop
10 QFT Bounds ω3
magnitude contours on the Nichols chart or by a computer
search. Examples of bounds are shown in the figure. These 0
−10 ω4
bounds are used in the next step as a guide for loop shaping the
nominal open-loop transfer function. −20 Nominal Loop, L0(jω)
−30 ω4
3) Loop shaping: Once the QFT bounds are determined, the
−40
nominal loop transfer function, L0 ( s ), should be designed, by −350 −300 −250 −200 −150 −100 −50 0
Open-Loop Phase (deg)
adding proper poles and zeroes, to yield a stable nominal closed
loop, while at the same time satisfying all bounds. A typical plot
of L0 ( s ) is given in the figure. An optimum L0 ( s ) is one that
satisfies the bounds and decreases as rapidly as possible with Typical bounds and nominal loop function in the Nichols chart.

where f is the sensed force and x, x e , and d are the actua- 2 2

tor’s displacement, the environmental displacement, and q i = cdwx sp ( pi − pe ) q o = cdwx sp ( ps − po )
ρ ρ (5b)
the equivalent viscous damping coefficient of the actua-
tor, respectively. The force originated from the hydraulic
actuator, fa , is where q i and q o represent fluid flows into and out of the
valve, respectively. cd is the orifice coefficient of dis-
fa = pi Ai − po Ao (4) charge, ρ is the mass density of the fluid, ps is the pump
pressure, and pe is the return (exit) pressure. w is the area
where Ai and Ao are the piston effective areas and pi and po gradient that relates the spool displacement( x sp ) to the or-
are the input and output line pressures, respectively. The ifice area. Continuity equations for oil flow through the cyl-
governing nonlinear equations describing the fluid flow dis- inder, neglecting the leakage flow across the actuator’s
tribution in the valve are written, in their simplest forms, as piston, are
[25]
x sp ≥ 0 (extension) dx 1 dpi
q i = Ai + Vi
dt β dt (6a)
2 2
q i = cdwx sp ( ps − pi ) q o = cdwx sp ( po − pe )
ρ ρ (5a)
dx 1 dpo
q o = Ao − Vo .
x sp < 0 (retraction) dt β dt (6b)

April 2001 IEEE Control Systems Magazine 69

Here β is the effective bulk modulus of the hydraulic fluid, as uncertainty. In this section, the derivation procedure will
and Vi and Vo are the volumes of fluid trapped at the sides of be followed step-by-step to highlight the assumptions that
the actuator. They can be expressed as functions of actua- are made during the derivation and simplification.
tor displacement The linearized flow relationships from (5) are

q i = K si x sp − K pi pi
Hydraulic actuators are (9a)

advantageous because of their

q o = K so x sp + K po po
high force-to-weight ratio and (9b)

fast response time. where K si (K so ) and K pi (K po ) represent flow and pres-

sure sensitivity gains of the valve, respectively. Super-
scripts i and o stand for input and output, respectively.
Vi ( x ) = Vi + xAi (7a) They are given as follows:

for extension ( x sp ≥ 0 ) for retraction ( x sp < 0 )

Vo ( x ) = Vo − xAo (7b) 2 2
K si = cdw ( ps − pi ) K si = cdw ( pi − pe )
ρ ρ
where Vi and Vo are the initial volumes trapped in the blind 2 2
and rod sides of the actuator. The relationship of the spool K so = cdw ( po − pe ) K so = cdw ( ps − po )
ρ ρ
displacement, x sp , and the input voltage, u, to the propor-
cdwx sp − cdwx sp
tional valve can be expressed as a first-order system that is K pi = K pi =
accurate enough to model low-cost proportional valves [26] 2ρ( ps − pi ) 2ρ( pi − pe )
cdwx sp − cdwx sp
K po = K po = .
τ dx sp 1 2ρ( po − pe ) 2ρ( ps − po )
u= + x sp
ksp dt ksp (8)

As can be seen, K si (K so ) and K pi (K po ) are load- and pres-

where τ and ksp are gains describing the valve dynamics.
sure-dependent variables. Hence, in this study, they are
Equations (1)-(8) express the relationship between the con-
considered uncertain but bounded parameters. Assuming
tact force, f, and the input control voltage, u.
small piston displacements within the vicinity of the
midstroke, the following approximation is made:
Derivation of Linear Transfer Function
The nonlinear equations derived above are now linearized
Vi ( x ) Vo ( x ) 1  Vi + Vo 
about an operating point to obtain a linear transfer function. ≈ ≈   = C.
The objective is to utilize this model within the QFT ap- β β β 2  (10)
proach. As will be seen later, the variation of the operating-
point-dependent parameters will be included in the model Thus, (6) can be written in the Laplace domain as

x xe Qi = Ai sX + CsPi (11a)
ks ke
po
pi Ai ma me
Ao Qo = Ao sX − CsPo . (11b)
ds de
qi qo d
Substituting (9) into (11) and rearranging for line
pressures, we have

xsp
Ai s K si
Pi = − X+ X sp
Cs + K p
i
Cs + K pi (12a)

Pe Pe
Ps
Ao s K so
Figure 1. Schematic model of the hydraulic actuator interacting with the Po = X − X sp .
environment. Cs + K po Cs + K po (12b)

70 IEEE Control Systems Magazine April 2001

Incorporating (12) into (4) we have Table 1. Operating values and parameter ranges
pertaining to the linear transfer function.
 KoA K si Ai   Ai2 s Ao2 s  Parameter Nominal Value Range
Fa =  s o o +  X sp −  + o 
X.
Cs + K p Cs + K p  Cs + K p Cs + K p  (13)
i i
ke 75 (kN/m) 50-100
3
Equations (1)-(3) are now combined to form the following Ks 0.375 (m /pa.s) 0.25-0.5
transfer functions: −12 2 −12
Kp 2.5(10) (m /s) 0-5(10)
−11 −11 −11
Fa ( s ) 1 ( ma s + ds )( me s + ( de + ds )s + ks + ke )
3
2 2 C 1.5(10) (m /pa) 1(10) -3(10)
= 
F ( s ) ks  me s 2 + de s + ke d 700 (N/m/s) 600-800

( ds s + ks )( me s 2 + de s + ke )  ma 20 (kg) 19.9-20.1
+ 
me s 2 + de s + ke  Ai 0.00203 (m )
2
0.00193-0.00213
(14)
2
Ao 0.00152 (m ) 0.00144-0.00160

X ( s ) 1 me s 2 + ( de + ds )s + ks + ke ksp 0.0012 (m/V) 0.0011-0.0013

= .
F ( s ) ks me s 2 + de s + ke (15)
τ 35 (ms) 30-40

Substituting Fa ( s ) and X ( s ) from (14) and (15) into (13) and

rearranging it, the transfer function describing the relation Including the dynamics of the valve from (8), the transfer
between contact force and spool displacement is derived as function between measured contact force, F ( s ), and control
(1 6 ) , s h o w n b e l o w , w h e r e ϕ( s ) = me s 2 + de s + ke , voltage, U ( s ), is written as
ψ( s ) = ma s 2 + ds, and ζ( s ) = ds s + ks .
To simplify the above transfer function, K si (K so ) and K pi F (s) ksp  K s ke ( Ai + Ao ) 
o =  .
( K p ) are replaced by K s and K p . Hence transfer function U ( s ) ( τs + 1) ( K p + Cs )( ma s + ds + ke ) + ( Ai s + Ao s ) 
2 2 2

(16) is reduced to (17), shown below.

Any variations or approximation in K si (K so ) and K pi (K po )
are now included as uncertainty in K s and K p , respectively.
The above relation can be further simplified if the stiffness
of the force sensor and the piston rod are high compared
with the environmental stiffness and the hydraulic compli-
ance. Hence their dynamics are not excited during the con-
tact and can be lumped together as a rigid body. Based on
this assumption, (17) is rewritten as
F (s) K s ( Ai + Ao )( me s 2 + de s + ke )
= .
X sp ( s ) ( K p + Cs )( ma s 2 + ( d + de )s + ke ) + ( Ai2 s + Ao2 s )
(18)
Further, we assume the dynamics of the environment is
dominated by a pure stiffness, ke . This type of environment
has enjoyed popularity among many researchers [27]. The
system transfer function is finally written in its simplest
form as
F (s) K s ke ( Ai + Ao )
= .
X sp ( s ) ( K p + Cs )( ma s 2 + ds + ke ) + ( Ai2 s + Ao2 s) (19)
(20)
Equation (20) is now considered as a parametrically uncer-
tain system. For example, the uncertainty ranges in K s and
K p reflect variations in the operating point, supply pres-
sure, and, in part, the orifice area gradient. The variations
in the environmental stiffness and damping of the system
are included in parameters ke and d, respectively. Further-
more, the uncertainty in C represents the changes in the
fluid bulk modules and the volumes of fluid trapped at the
sides of the actuator. The uncertainty in the valve charac-
teristic is captured in variations of τ and ksp . All these pa-
rameters are known to have an effect on the system’s
stability.
Table 1 lists the nominal values of all parameters in (20)
and their corresponding uncertainty ranges. These values,
which are representative of the hydraulic actuator under in-
vestigation (see the “Experimental Setup” section), were ob-
tained from available references (including manufacturer
specifications), direct measurements of some geometric pa-
rameters, and indirect identification based on experimental
observation of the responses [26]. These uncertain parame-

F (s)
=
(
ks ( me s 2 + de s + ke ) K si Ai ( K po + Cs ) + K so Ao ( K pi + Cs ) )
X sp ( s ) ( K pi + Cs )( K po + Cs )( ψ ( s )ϕ( s ) + ζ( s )ϕ( s ) + ψ ( s )ζ( s )) + ( Ai2 s( K po + Cs ) + Ao2 s( K pi + Cs ))(ϕ( s ) + ζ( s )) (16)

F (s) ks K s ( Ai + Ao )( me s + de s + ke )
2

= . (17)
X sp ( s ) ( K p + Cs )(( ψ ( s )ϕ( s ) + ζ( s )ϕ( s ) + ψ ( s )ζ( s )) + ( Ai2 s + Ao2 s )(ϕ( s ) + ζ( s ))

April 2001 IEEE Control Systems Magazine 71

 Controller Synthesis
Prefilter Controller Uncertain Plant D (s)
The goal of this section is to design a robust force
controller for the hydraulic actuator that is rep-
R(s)
F (s)
+
G(s) P(s,α) + Y(s)
resented by the uncertain transfer function (21).
− +
A typical two-degree-of-freedom feedback sys-
tem configuration in QFT is shown in Fig. 2. A
strictly proper controller, G( s ), and a strictly
proper prefilter, F ( s ), are to be designed such
that the following conditions are satisfied (read-
Figure 2. Two-degree-of-freedom QFT control system. ers are referred to [28] for a description of the
QFT design procedure):
ters are grouped into a vector, denoted as α. Then the sys- i) Closed-loop robust stability. The associated QFT ro-
tem’s open-loop transfer function can be defined as bustness constraint in terms of the nominal loop
transfer function, L0 ( s ) = P0 ( s )G( s ), is given by [28]:
F (s)
P( s ,α ) = .
U(s) (21)  
 L0 (iω ) 
The plant evaluated at its nominal operating point is re-  p (iω )  ≤ M = 14
. ∀ω ∈ [0 , ∞ )
 0
+ L0 (iω )
ferred to here as the nominal plant and is shown as  p(iω ,α )  (23)

P0 ( s ) = P( s ,α 0 ) (22) which implies an approximately 3 dB gain margin for

the closed-loop system.
where α o is the vector containing all nominal values of the
plant transfer function.

ω=0.01
120
20 100
Upper Bound ω=0.05
0 80 ω=0.1
Magnitude (dB)

−20 60
ω=0.5
Magnitude (dB)

Tu (s)
−40 40 ω=1

−60 Lower Bound 20 ω=5

ω=70 ω=10
−80 Tl (s) 0 ω=100
ω=50
−100 −20
−350 −300 −250 −200 −150 −100 −50 0
−120 Phase (deg)
100 101 102 (a)
Frequency (rad/s)
(a) ω=0.01
120 0.01
1.4
100
1.2 Upper Bound 0.05 0.05
80 0.1
Magnitude (dB)
Force (Normalized)

1
60 0.5
0.8 0.5
Lower Bound 40 1 1
0.6 5
20 10 5
0.4 10
0 50
0.2 100 70
−20
−350 −300 −250 −200 −150 −100 −50 0
0 Phase (deg)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s) (b)
(b)
Figure 4. QFT bounds on Nichols chart with: (a) uncompensated
Figure 3. Frequency and time domain tracking bounds. plant and (b) nominal loop.

72 IEEE Control Systems Magazine April 2001

 ii) Robust reference input. To meet tracking performance transfer function, L0 , should be determined such that it lies
requirements, the controller should satisfy the follow- on or above the corresponding bounds at every frequency.
ing inequality: This can be done using a procedure that is called loop shap-
ing in QFT. Finally, the controller can be extracted from L0 by
dividing by the nominal plant transfer function, P0 .
L0 (iω ) Design frequencies are selected as ω = {0.01,0.05,0.1,
Tl (iω ) ≤ F (iω ) ≤ Tu (iω ) ∀ω ∈ [0, ∞ )
p0 (iω ) 0.5,1,5,10,50,70,100} rad/s. The bounds generated by con-
+ L0 (iω )
p(iω ,α ) (24)
straints (23), (24), and (26) are shown in Fig. 4(a) for the un-
compensated plant. The nominal open-loop transfer func-
where the upper and lower tracking bounds are de-
fined as 20
Upper Bound δ
0
 s 
 + 1
 2.8  −20
Tu ( s ) =

Magnitude (dB)
s s s 
 + 1  + 1  + 1 −40
4 7 8  (25a) δall
Lower Bound
−60

−80
1
Tl ( s ) = .
s 9.6 s 
2
 s  s −100
 + 1  + 1  + + 1
 4.8   80   50 50  (25b) −120
100 101 50 102
Frequency (rad/s)
These bounds are built from the time domain figures of (a)
merit for step responses such as peak overshoot, peak time,
and settling time [14]. In this work, the desired lower track- 20
ing bound, Tl ( s ), is built to have an overdamped response Upper Bound
0
with ≈1 s settling time. For this purpose, a model with a real
pole at s = −4.8 and a pair of complex poles are chosen. The −20
Magnitude (dB)

real pole must be more dominant than the complex poles −40
[14]. Moreover, a high-frequency pole at s = −80 is inserted Lower Bound
in Tl ( s ), which does not affect the desired performance −60
specification but widens the range between Tu ( s ) and Tl ( s ) −80
in the high-frequency band. The figures of merit for upper
bound, Tl ( s ), are a 5% peak overshoot and ≈1 s settling time. −100
Tu ( s ) is selected with three real poles and a zero. The zero is −120
closer to the origin than the poles to have an underdamped 100 101 102
Frequency (rad/s)
response. The frequency and time domain plots of these (b)
bounds are shown in Fig. 3(a) and (b), respectively.
1200
iii) Closed-loop disturbance attenuation (sensitivity reduc- Upper Bound
tion). For disturbance rejection at the plant output, an 1000
upper tolerance is imposed on the sensitivity function.
Here we consider only a constant upper bound to limit 800 Lower Bound
Force (N)

the peak value of disturbance amplification as follows:

600

400
 p0 (iω ) 
 p(iω ,α )  200
S (iω ,α ) max =  ≤ M D (ω ) = 1.2 ∀ω ∈ [0 ,10].
p (iω )
 0 + L0 (iω )
 p(iω ,α ) max 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(26) Time (s)
(c)
The above design specifications impose constraints on Figure 5. System responses over range of parametric uncertainty
the allowable loop gain,|L0|. These constraints can be shown (simulation): (a) closed-loop frequency responses without prefilter,
on a Nichols chart as a curve boundary at each design fre- (b) closed-loop frequency responses with prefilter, and (c) time
quency. Once the bounds are derived, the nominal open-loop responses.

April 2001 IEEE Control Systems Magazine 73

tion, L0 , must be shaped to satisfy all these bounds at each
frequency. For low frequencies, L0 should lie on or above the
bounds to satisfy the constraints. However, for higher fre-
quencies in the range, e.g., ω = 50, 70, and 100 rad/s, the con-
straints generate closed boundaries which L0 should not
enter. For the industrial hydraulic actuator under investiga-
tion, the valve dead band produces a steady-state error in the
system response in the absence of an integrator. For a zero
Maintaining a low-order controller
was an important factor in the design
process.
steady-state error, L0 must contain an integrator. Thus, a pos-
sible loop can be obtained by cascading an integrator. Fur-
ther, a gain of 0.0065 is required to bring the open-loop
transfer function within the specification bounds. Two ze-
roes should be added to the structure of the controller to sat-
isfy the bounds and to force the nominal loop, L0 , to the
right-hand side of the closed boundaries. After several itera-
tions the controller zeroes were chosen at s = −4 and s = −40.
At this stage, L0 closely follows the bounds up to ω = 70 rad/s.
3 3 4 5
Analog and Digital u 
Interface Board
6 F (s) =
1 
2  5.5   9 
Figure 6. Experimental test station. 1: proportional value, 2:
pump with pressure regulator, 3: pressure transducer, 4:
incremental encoder, 5: force sensor, 6: environment.
Figure 7. Close-up photograph of the environment.
74 IEEE Control Systems Magazine
The next step is to add two high- frequency poles to allow for
a quick descend of the nominal loop to decrease the cost of
feedback. Fig. 4(b) shows the final loop shaping of the sys-
tem. The controller that satisfies the specifications is
s  s 
0.0065  + 1  + 1
 4   40 
G( s ) = .
 s  s 
s + 1  + 1
100   300  (27)
To satisfy the tracking specification, a pre-
filter, F ( s ), is required to place the closed-
loop frequency response in the specified
range, betweenTu ( s ) andTl ( s ). Fig 5(a) shows
the frequency response of the closed-loop
system without a prefilter where the designed controller
(27) only guarantees that the variation in the system’s mag-
nitude will be less than or equal to that allowed. The pre-
filter is synthesized by calculating its magnitude bound at
each frequency. For example, in Fig. 5(a), consider the fre-
quency response of the control system at ω = 50 rad/s. The
maximum frequency variation, δ, is about 17 dB = 2 − ( −15 ),
which is less than the amount allowed for this frequency,
i.e., δ all = 25 dB =|Tu (i 50 )|−|Tl (i 50 )| = −30 − ( −55 ). The ac-
ceptable magnitude of the prefilter at this frequency is then
found to be −55 − ( −15 ) ≤ |F (i50 )|≤ −30 − (2). By applying
the above procedure for the design frequency range, a suit-
able prefilter is found to be
 s 
+ 1
150 
.
 s s 
+ 1  + 1
(28)
The closed-loop Bode plots of the plant with prefilter are
given in Fig. 5(b). Note that all the tracking specifications
are met in the frequency domain. The corresponding
closed-loop simulated time responses to a step input of 1000
N are shown in Fig. 5(c). As can be seen, for all cases pertain-
ing to extreme parts of the operating envelope, the specifi-
cations are satisfied.
Experiments
Experimental Setup
The test station consists of a hydraulic unit, a 486/66-based
PC equipped with a Metrabyte M5312 quadrature incre-
mental encoder card, and a DAS-16 analog-to-digital (A/D)
conversion card (Fig. 6). The pump provides constant op-
erational supply pressure up to ≈1000 psi. The hydraulic
valve is a low-cost closed-center four-way proportional
valve. The positioning of the valve spool is based on the
pulse-width modulation principle. A spring is used to rep-
April 2001
resent the environment (Fig. 7); replacing the spring Conclusions
changes the stiffness of the environment. A force trans- This article has described the application of the QFT tech-
ducer with a capacity of 1000 lb (4.5 kN) is inserted be- nique to the development of a force controller for hydraulic
tween the environment and the piston rod. The 12-bit A/D actuators. A linear parametrically uncertain fourth-order
conversion allows a force resolution of 10 N when the full model was developed to represent the relation between the
range is used. Three pressure transmitters read pump, control signal and the force acting on the environment. Plant
supply line, and return line pressures with ±1% accuracy.
An incremental encoder with sensing resolution of 0.06
mm reads the displacement of the cylinder piston. The 600
control signal, generated by the control algorithm, is con-
verted to an analog signal by the A/D card and then trans- 500
Tu
mitted to the hydraulic valve amplifier. The low-cost valve
operates within the range ±1.8 V and has a dead band of 400
Tl

Force (N)
±10% (±0.15 V) within which the actuator does not move.
No attempt was made to eliminate or bypass these 300
nonideal effects due to the intended application.
200
Reference Force = 500 N
Results Ke = 50 kN/m
100
The controller described by transfer function (27) was Ke = 100 kN/m
descretized and implemented on the experimental test
0
stand. The sampling frequency for the controller was 200 Hz. 0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s)
Several experiments were performed to study the effects of (a)
variations of environmental stiffness, supply pressure, and
0.5
force setpoint. First, the variation of environmental stiffness
was studied by using different springs. Two different springs 0.45
Ke = 50 kN/m
with stiffnesses of 50 and 100 kN/m were used for this pur- 0.4
Ke = 100 kN/m
pose. The results are shown in Fig. 8. As can be seen, the 0.35
Control Signal (V)

controller is capable of handling the changes in environ- 0.3

mental stiffness and the steady-state errors are small. With
0.25
reference to Fig. 8(a), responses exhibit initial delays,
0.2
mainly due to the dead band (±10%) and dry friction in the
low-cost valve used in this experiment. Consequently, the 0.15
responses do not precisely fit within the design specifica- 0.1
tion bounds shown with dotted lines. Indeed, the effect of 0.05
valve dead band was not incorporated into the QFT design 0
procedure and was only handled by cascading an integrator 0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s)
in the controller. The actuator displacement and the control (b)
signal for such a trial are shown in Fig. 8(b) and (c), respec-
tively. The control signal is smooth, but it contains high-fre- 0.012
quency oscillations that originated from the noise in the
force sensor. Fig. 9 compares the test results for two differ- 0.01
ent reference forces (500 N and 1000 N) and with similar en-
Displacement (m)

0.008
vironmental stiffness (50 kN/m). As shown in this figure,
despite changing the loading condition, the system’s rise
0.006
time did not change considerably. The ability of the control-
ler to cope with pump pressure variations was also tested.
0.004
Typical results are shown in Fig. 10, where the pump pres-
sure was varied 100%. As can be seen, the control effort is re- Ke = 50 kN/m
0.002
duced for the higher pump pressure. Finally, a test was Ke = 100 kN/m
arranged to investigate the repeatability of the controller
0
and the effect of long-term system operation on the re- 0 0.5 1 1.5 2 2.5 3 3.5 4
sponse. A 1000-N step response was performed once and Time (s)
(c)
then again after the machine had been in continuous oper-
ation for more than one hour. As shown in Fig. 11, there is Figure 8. Step responses with different environmental stiffnesses
no noticeable difference between the two responses. (experiment): (a) force, (b) control signal, and (c) displacement.

April 2001 IEEE Control Systems Magazine 75

uncertainty was quantified within the linear model to com-
pensate for time-varying dynamics and to allow for variations
in the environmental stiffness, hydraulic fluid flow and pres-
sure gains, actuator compliance, and many other parameters
specific to the hydraulic actuator components. A robust con-
troller was then designed that, along with a prefilter, main-
tains a satisfactory force control performance against the
environment despite a wide range of uncertainty. Main-
1200
1000
800
Force (N)
taining a low-order controller was an important factor in the
design process to ensure easy implementation and demon-
strate its suitability for industrial applications.
The QFT controller designed here was implemented suc-
cessfully on an industrial hydraulic actuator equipped with
a low-cost proportional valve. The transparency of the de-
sign process facilitated the quick adjustment of the control-
ler for experimental implementation. Several tests were
performed to show the force-regulating ability of the devel-
oped controller. In particular, the experimental results dem-
onstrated the robustness of the QFT controller for up to
100% variations in environmental stiffness, supply pres-
sure, and reference force. Good performance was also ob-
tained despite significant variations in actuator dynamics.

The single-loop fixed-gain controller was easy to implement,

600
required very little computational effort, needed no online
tuning or gain scheduling, and resulted in good perfor-
400
mance in both transient and steady-state periods. Knowl-
200 edge of the lower and upper bounds of uncertain parameters
Ke = 50 kN/m
was the only requirement for the controller design.
0 In summary, the results of this work, which represent a
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) contribution to the published work on the application of
QFT, provide insight into the potential and effectiveness of
Figure 9. Step responses for different reference forces the technique for the design of robust force controllers for
(experiment). hydraulic actuators. In particular, we have shown that it is
feasible to employ a single fixed-gain force controller for a
1200

1000 1200

800 Stiffness = 75 kN/m 1000

Force (N)

600 800
Force (N)

400 Pump Pressure = 500 psi 600 Stiffness = 75 kN/m

Pump Pressure = 1000 psi Pump Pressure = 750 psi
200 400

200
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) 0
(a) 0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s)
0.6 (a)
0.6
0.5 Pump Pressure = 500 psi
Pump Pressure = 1000 psi 0.5
Control Signal (v)

0.4
Control Signal (v)

0.4
0.3
0.3
0.2
0.2
0.1
0.1
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s) 0
0 0.5 1 1.5 2 2.5 3 3.5 4
(b) Time (s)
(b)
Figure 10. Step force responses for different supply pressures
(experiment). Figure 11. Repeatability test (experiment).

76 IEEE Control Systems Magazine April 2001

hydraulic actuator, and that control performance can be in-
sensitive to structured plant variation. It also became ap-
parent that the uncertain linear fourth-order model
adequately characterizes the major features of the class of
hydraulic systems under investigation.
In future work, we plan to use the QFT method to design
automatic force controllers for heavy-duty hydraulic equip-
ment such as excavators. These machines strongly interact
with environments, and accurate models for the interac-
tions are difficult to obtain. Currently, we are investigating
the feasibility of assuming the environmental force as a
combination of disturbance and unstructured uncertainty
to be handled by the QFT controller.
References
[1] A. Alleyne, “Nonlinear force control of an electrohydraulic actuator,” in
Proc. Japan/USA Symp. Flexible Automation, Boston, MA, 1996, pp. 193-200.
[2] F. Conrad and C.J.D. Jensen, “Design of hydraulic force control systems
with state estimate feedback,” in Proc. IFAC 10th Triennial Congr., Munich,
Germany, 1987, pp. 307-312.
[3] N. Niksefat and N. Sepehri, “Robust force controller design for an
electrohydraulic actuator based on nonlinear model,” in Proc. IEEE Conf. Ro-
botics and Automation, Detroit, MI, 1999, pp. 200-206.
[4] R.F. Pannett, P.K. Chawdhry, and C.R. Burrows, “Alternative robust con-
trol strategies for disturbance rejection in fluid power systems,” in Proc. ACC,
San Diego, CA, 1999, pp. 739-743.
[5] J. Watton, “On linearized coefficients for an underlapped servo-valve cou-
pled to a single-rod cylinder,” ASME J. Dynam. Syst., Measure., Contr., vol. 112,
no. 4, pp. 794-796, 1990.
[6] J. Yu, Z. Chen, and Y. Lu, “The variation of oil effective bulk modulus with
pressure in hydraulic systems,” ASME J. Dynam. Syst., Measure., Contr., vol.
116, no. 1, pp. 146-150, 1994.
[7] Y.N. Chen, C.B. Lee, and C.H. Tseng, “A variable-structure controller de-
sign for an electrohydraulic force control servo system,” J. Chinese Soc. Mech.
Eng., vol. 11, no. 6, pp. 520-526, 1990.
[8] H. Sun and G.T.C. Chiu, “Nonlinear observer based force control of
electrohydraulic actuators,” in Proc. ACC, San Diego, CA, 1999, pp. 764-768.
[9] R. Liu and A. Alleyne, “Nonlinear force/pressure tracking of an
electrohydraulic actuator,” in Proc. IFAC 14th Triennial Congr., Beijing, China,
1999, pp. 469-474.
[10] G. Wu, N. Sepehri, and K. Ziaei, “Design of a hydraulic force control sys-
tem using a generalized predictive control algorithm,” Proc. Inst. Elec.
Eng.-Contr. Theory Applicat., vol. 145, no. 5, pp. 428-436, 1998.
[11] L. Laval, N.K. M’Sirdi, and J.C Cadiou, “H ∞ force control of a hydraulic
servo-actuator with environmental uncertainties,” in Proc. IEEE Conf. Robot-
ics and Automation, Minneapolis, MN, 1996, pp. 1566-1571.
[12] G. Ferretti, G. Magnani, and P. Rocco, “On the stability of integral force
control in case of contact with stiff surface,” ASME J. Dynam. Syst., Measure.,
Contr., vol. 117, no. 4, pp. 547-553, 1995.
[13] I.M. Horowitz, Quantitative Feedback Design Theory (QFT). Boulder, CO:
QFT Publications, 1992.
[14] J.J. D’Azzo and C.H. Houpis, Linear Control System Analysis and Design.
New York: McGraw-Hill, 1988.
[15] O. Yaniv, Quantitative Feedback Design of Linear and Nonlinear Control
Systems. Norwell, MA: Kluwer, 1999.
[16] S. Jayasuriya, O. Yaniv, O.D.I. Nwokah, and Y. Chait, “Benchmark prob-
lem solution by quantitative feedback theory,” J. Guidance, Contr., Dynam.,
vol. 15, no. 5, pp. 1087-1093, 1992.
[17] S.G. Breslin and M.J. Grimble, “Longitudinal control of an advanced com-
bat aircraft using quantitative feedback theory,” in Proc. ACC, Albuquerque,
NM,1997, pp. 113-117.
April 2001 IEEE Control Systems Magazine
[18] S.N. Sheldon and S.J. Rasmussen, “Development and first successful
flight test of a QFT flight control system,” in Proc. IEEE National Aerospace and
Electronics Conf., NAECON, Dayton, OH, 1994, pp. 629-636.
[19] M.B. Leahy, D.E. Bossert, and P.V. Whalen, “Robust model-based control:
An experimental case study,” in Proc. IEEE Conf. on Robotics and Automation,
Sacramento, CA, 1990, pp. 1982-1987.
[20] S.J. Rober, Y.C. Shin, and O.D.I. Nwokah, “A digital robust controller for
cutting force control in the end milling process,” ASME J. Dynam. Syst., Mea-
sure., Contr., vol. 119, no. 2, pp. 146-152, 1997.
[21] M. Pachter, C.H. Houpis, and K. Kang, “Modeling and control of an
electrohydrostatic actuator,” Int. J. Robust Nonlinear Contr., vol. 7, no. 6, pp.
591-608, 1997.
[22] D.F. Thompson and G.G. Kremer, “Quantitative feedback design for a
variable-displacement hydraulic vane pump,” in Proc. ACC, Albuquerque,
NM, 1997, pp. 1061-1064.
[23] R. Volpe and P. Khosla, “Theoretical analysis and experimental verifica-
tion of a manipulators/sensor/environment model for force control,” in Proc.
IEEE Conf. Systems, Man, Cybernetics, Los Angles, CA, 1990, pp. 784-790.
[24] S.D. Eppinger and W.P Seering, “On dynamic models of robot force con-
trol,” Proc. IEEE Conf. Robotics and Automation, San Francisco, CA, 1986, pp.
29-34.
[25] H.E. Merritt, Hydraulic Control Systems. New York: Wiley, 1967.
[26] K. Ziaei, “Development of a nonlinear adaptive control scheme for hy-
draulic actuator,” M.Sc. thesis, Dept. Mech. Eng., Univ. of Manitoba, Winni-
peg, Canada, 1998.
[27] H. Seraji and R. Colbaugh, “Force tracking in impedance control,” Int. J.
Robot. Res., vol. 16, no. 1, pp. 97-117, 1997.
[28] S. Jayasuriya, “Frequency domain design for robust performance under
parametric, unstructured, or mixed uncertainties,” ASME J. Dynam. Syst.,
Measure., Contr., vol. 115, no. 2, pp. 439-451, 1993.
[29] I.M. Horowitz, “Survey of quantitative feedback theory (QFT),” Int. J.
Contr., vol. 53, no. 2, pp. 255-291, 1991.
Navid Niksefat received the B.Sc. degree from Sharif Uni-
versity of Technology, Iran, and the M.Sc. degree from
Isfahan University of Technology, Iran, in 1994 and 1997, re-
spectively, both in mechanical engineering. Since January
1998, he has been with the Experimental Robotics and
Teleoperation Laboratory, University of Manitoba, Canada,
where he has held teaching and research assistantship posi-
tions. His current research interests include robust control
and control of discontinuous and nonlinear systems, with
applications to robotics and hydraulic systems. He re-
ceived several awards and a fellowship from the University
of Manitoba during 1999-2001.
Nariman Sepehri joined the Department of Mechanical and
Industrial Engineering at the University of Manitoba in 1991,
where he is currently a Professor. He received an M.Sc. in
computer-aided process planning in 1986 and a Ph.D. in ro-
botics and controls in 1990, both from the University of Brit-
ish Columbia. His areas of interest are mathematical
modeling and real-time simulation, robust control design,
and fluid power. An area of research interest is the automa-
tion of heavy-duty industrial (forest and mining) machinery
using robotics and advanced control technology. He is a
member of the American Society of Mechanical Engineers
and the IEEE.
77