490

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002

Pressure Observer-Controller Design for Pneumatic

Cylinder Actuators
Shunmugham R. Pandian, Member, IEEE, Fumiaki Takemura, Yasuhiro Hayakawa, Member, IEEE, and
Sadao Kawamura, Member, IEEE

AbstractRecent research has shown that robust and precise

high-speed control of pneumatic actuators is practicable, by
application of advanced control techniques such as model-based
adaptive and sliding-mode control. However, the resulting need for
full state-based feedback and feedforward controlin particular,
the measurement of air pressuresincreases both the cost and
complexity of the overall system. In this paper, we consider the
problem of design of observers to estimate the chamber pressure
variables in a cylinder actuator. Since the cylinder pressures are
not simultaneously observable because of the nature of cylinder
dynamics, we first propose a continuous gain observer in which
the pressure on one side of the cylinder is measured and the
pressure on the other side is estimated. Next, we propose a
sliding-mode observer where numerically estimated acceleration
is used in order to observe both pressures, with ultimate bounded
stability. A sliding-mode controller is proposed, whose sensitivity
to errors under the sliding-mode observer is studied. The proposed
observers are simple, effective and easy to implement. Results of
experimental implementation illustrate the practical effectiveness
of the new observers.
Index TermsActuators, model-based sliding-mode control,
pneumatic cylinders, pressure observers.

I. INTRODUCTION

NEUMATICALLY actuated systems have the advantages

of excellent compliance, high payload-to-weight and payload-to-volume ratios, high speed and large force capabilities
and a simple operational mechanism. As a result, compact,
light-weight manipulators can be developed cost-effectively for
a wide range of applications in domestic (e.g., rehabilitation and
prosthetics [1]), dexterous (e.g., multifingered hands [2] and
legged manipulators [3]) and heavy-duty industrial (e.g., materials handling [4]) environments. However, precise, high-speed
control of pneumatic systems has so far been difficult due to
the high-order, time-variant actuator dynamics, nonlinearities
due to compressibility of air, external disturbances such as
static and Coulomb friction and payload and pressure supply
variations.
Despite their many advantages over electrically actuated
robots which are widely used in robotics and mechatronics,
widespread use of pneumatic actuators and manipulators is
Manuscript received October 13, 2000; revised January 23, 2002. Recommended by Technical Editor C.-H. Menq.
S. R. Pandian is with the Engineering Science Program, University of
Michigan-Flint, Flint, MI 48502-1950 (e-mail: [email protected]).
F. Takemura and S. Kawamura are with the Department of Robotics, Ritsumeikan University, Kusatsu 525-8577, Japan.
Y. Hayakawa is with the Department of Control Engineering, Nara National
College of Technology, Nara 639-1058, Japan.
Digital Object Identifier 10.1109/TMECH.2002.805624

mainly inhibited by the above difficulties in controlling them.

However, with the advent of high-performance electropneumatic servovalves and low-cost, high-speed, real-time computer
control, it has now become possible to implement advanced
control algorithms to achieve precise, robust motion control
with pneumatic systems.
Recently, we have proposed practical design methods for
nonadaptive and adaptive sliding-mode controllers for pneumatic cylinders highlighting these aspects [5], [6]. We have also
applied this approach to control of vane-type air motors which
provide rotary actuation [7], [8]. Precise and robust control
performance has been obtained by use of third-order actuator
dynamics in model-based sliding-mode controller design. In
particular, we have shown that in contrast to the second-order
control laws using only position and velocity feedback that
have been mainly proposed in the literature, the use of pressure difference between the two cylinder chambers (between
successive chambers enclosed by vanes, in the case of motors)
in feedback and the use of individual chamber pressures in
feedforward control is essential for better performance.
A major limitation of this approach is that pressure sensors
are presently very expensive and a typical high-precision pressure sensor can be costlier than the actuator itself. Also, the use
of pressure sensors adds to the cost, size, weight, energy consumption and complexity of the overall system and reduces its
reliability.
In recent years, researchers have extensively studied the
application of observer theory to the design of velocity and
force observers for electrically actuated robot manipulators
and pressure and disturbance observers for hydraulic actuators.
The problem of observer design for pneumatic actuators and
manipulators has so far not received much attention in the
literature. Bourdat et al. [9] have considered the design of a
linear observer for estimation of velocity signal. Their results
are inconclusive concerning the superiority of the observer
over numerical differentiation to obtain velocity estimates.
Noritsugu and Takaiwa [10] have used a disturbance observer
to estimate friction and interaction forces. Their design is based
on linearized dynamics and requires both velocity and pressure
sensors.
In this work, we consider the design of observers for cylindrical actuators in order to estimate the pressure variables using
available measured signals, namely position and velocity. It is
shown that as the cylinder dynamics is not completely observable, both the chamber pressures are not simultaneously observable. This necessitates the use of one pressure sensor, or one
regulated pressure supply, or other alternative arrangements.

1083-4435/02$17.00 2002 IEEE

PANDIAN et al.: PRESSURE OBSERVER-CONTROLLER DESIGN FOR PNEUMATIC CYLINDER ACTUATORS

491

The valve dynamics are much faster than the mechanical and
pressure dynamics. Therefore, neglecting the valve dynamics as
well as the cylinder friction, we can write the equations of the
actuator dynamics in state-space form as [5]

Fig. 1. Model of pneumatic cylinders.

As an example of this approach, we develop a continuous

gain, Luenberger-type, observer for estimation of one chamber
pressure, with the pressure in the other chamber being measured
with a pressure sensor.
When the control valve dynamics are neglected, the dynamics of the cylinder can be modeled as a third-order system.
Therefore, we can numerically estimate the cylinder acceleration from the position and velocity values without significant
delay, using high-speed computation such as that facilitated by
personal computers. By neglecting the cylinder friction effect,
we can then estimate the chamber pressure difference using
the equation for piston motion. Finally, we can use a pressure
observer to estimate one chamber pressure and obtain the other
chamber pressure as the difference between the calculated
chamber pressure difference and the estimated pressure.
The partial observability, however, of the chamber pressures
implies that in the case of a sliding observer the error dynamics
of the estimated pressure can be guaranteed of ultimate boundedness only. Therefore, for purposes of control we apply the
sliding-mode control method and verify that the properties of
robustness of sliding-mode control can be made use of to ensure that the control performance is robust to estimation errors
when used in conjunction with the sliding-mode observer.
We illustrate the effectiveness of the proposed observers for
pressure estimation and for sliding-mode control in the case of
both position and trajectory control by experimental implementation on an industrial cylindrical actuator.
In the present methods, the velocity signal is obtained by numerical differentiation and filtering of the position signal. It is
also possible to develop a combined velocity-pressure observer.
This approach is also advantageous compared to the traditional
use of tachogenerators, which are very expensive and overcomes
the problem of noise in numerical differentiation. The design of
the combined observer is similar to that of the pressure observers
and is omitted here due to lack of space [11].

II. DYNAMICS OF CYLINDRICAL ACTUATORS

A simplified model of the double-acting type pneumatic
cylindrical actuator considered in this study is shown in Fig. 1.
are the chamber air pressures, ,
are the ram
Here, ,
are the valve cross-section areas.
is the
areas and ,
mass of the piston and external payload and is the coefficient
of viscous friction and damping. is the position of the piston
endpoint and is the stroke of the piston.

(1)

are nonlinear
where is the ratio of specific heats and ,
1, 2) are positive valve parameters and
functions. and (
(
1, 2) are voltage inputs to the valves.
III. ESTIMATION OF CYLINDER PRESSURES
In the following analysis, we assume that the constant paramand
are known exactly. The nonlinear,
eters, , , ,
and
are functions of chamber
time-varying parameters
pressures, but can be assumed to be constants under conditions
of sonic air flow through the valves. For simplicity of notation,
we hereafter represent velocity by instead of .
Since (1) represents a nonlinear, time-varying system with
state-dependent input-coefficient matrix, a quantitative analysis
of its observability conditions is difficult [12]. Using the rank
criterion for observability of linear, time-invariant systems as a
heuristic test, we can infer that for the system of (1) we cannot
and
[11]. A physical intersimultaneously observe both
pretation of this situation is that since changes in chamber presand
have equal and opposite effects on acceleration
sures
of the cylinder, we cannot estimate the two pressures simultaneously when the velocity is zero. However, from similar conor
siderations we can show that at least one of the variables
is observable, if the other is known.
For purposes of feedback control using the sliding-mode approach, it is sufficient to use the pressure difference
in addition to use of both position and velocity variables. On the other hand, feedforward compensation using both
and
is necessary for precise and robust position and trajectory control.
Since we are interested in bidirectional motion of the cylinder
piston by varying the difference pressure across chambers, it is
possible to treat one of the pressures, say , as constant in the
case of position control and as having a given trajectory in the
case of trajectory control.
at a constant level,
For example, we may seek to maintain
or along a specified trajectory, by open-loop control of valve
input . However, an experimental implementation of this approach shows that due to the compressibility of air the values
are widely different from the regulation levels during the
of

492

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002

transient phase [11]. This means that it is necessary to either use

a pressure regulator for controlling air supply to chamber 2, or
to use a pressure sensor for one of the pressure variables, ,
say and develop a pressure observer to estimate .

where is the gain vector of the reduced-order observer.

, we can
Defining the observation error as
obtain the error dynamics by subtracting (8) from (6) as
(9)

IV. CONTINUOUS GAIN OBSERVER DESIGN

In this method, we treat
as a disturbance that can be
measured precisely using a pressure sensor, so that the system
dynamics is of third-order, as given by

(2)

In (2), the parameters are either known constants or timevarying as known functions of the measured states and .
Therefore, we can use a reduced-order nonlinear observer to es[13].
timate the unmeasured state
In vector-matrix form, we can rewrite (2) as
(3)
,
and is a known, constant
where
parameter vector.
It is clear from the above that the matching condition for invariance of the state vector with reference to the disturbance is
not satisfied.
We assume that errors in the measurement of are bounded,
such that
so that there exists a positive constant
(4)
where indicates the measured variable and [0, ] is the time
interval of interest.
as the measured state vector and
Denoting
as the unmeasured state variable, we partition the
system dynamics as [14]
(5)
(6)
. The partitioned matrices and vectors can readily
where
be inferred from (2) and their state dependence is omitted for
simplicity of presentation.
Defining the effective output as
(7)
the estimate

of the unmeasured state is specified as

(8)

.
since
From (9), we note that since the input coefficient is a function only of the (precisely) known state variable the observer
design is simplified. This is in contrast to the general nonlinear
case where the observability properties depend on the input [12].
) in (3) and hence the pair (
) is controlThe pair (
lable. Since (9) is of the first order, the local asymptotic stability
can be established from the simple
of the observation error
.
stability condition
and
Since the piston motion is restricted as
,(
the chamber pressures are restricted as
1, 2), where
and
are the atmospheric and supply presand are
sures, respectively, the parametric variations in
also bounded. Therefore, in (9) we can choose the observer gain
and the exvector as a function of the disturbance bound
pected bounded variations in the state-dependent parameters of
and .
If the pressure is not known precisely, then the observation
does not tend to zero, as the measurement error in
error
appears as a disturbance in (9). Moreover, increasing observer
gains also amplifies this disturbance, so a proof of bounded stability of the observer cannot be established. In practice, though,
pressure measurement errors are sufficiently small so using high
observer gains results in satisfactory performance.
V. SLIDING-MODE PRESSURE OBSERVER DESIGN
The pressure observer proposed in the previous section has
the drawback that it requires one pressure sensor. Many authors
have made use of piston acceleration obtained from sensors or
numerical differentiation as a feedback variable in the control of
pneumatic actuators. Others have used the equivalence between
pressure difference and acceleration to employ full-state feedback. In this section, we use numerically obtained acceleration
variable as an aid to overcome the problem of partial unobservability of chamber pressures.
Sliding-mode observers have many advantages over Luenberger observers and Kalman filters, such as robustness to model
parameter inaccuracies and measurement noise, effective estimation with low gains and so on. As is well-known, a finite-gain
sliding-mode controller is equivalent to a high-gain continuous
feedback controller. Therefore, based on the duality of controllability and observability, we can expect improved performance
using the new approach as compared to the performance of the
continuous gain observer [15], [16].
For the case when the cylinder friction is negligible and the
and
are known precisely, we can obtain acparameters
celeration values by numerical differentiation of velocity and
using the equation of mechanical dynamics
then estimate

PANDIAN et al.: PRESSURE OBSERVER-CONTROLLER DESIGN FOR PNEUMATIC CYLINDER ACTUATORS

in the form
(10)
where and represent estimation/measurement noise in the
acceleration and velocity variables, respectively. Here, the velocity variable itself can be obtained using a tachogenerator,
by numerical differentiation of position, or by use of a combined
velocity-pressure observer.
is again treated as a disturbance and the
In this case,
system dynamics is of third order, as given by (2). From the
and the numerically estimated value of
observed value of
, we can calculate
for use in feedforward control as
(11)
is the observed value of
and
is given by (10).
where
This implies, of course, that errors in the estimation of
and
will result in error in estimation of . However, by
improving the response of the observer for and by calculating
precisely the effects of these errors can be minimized.
is not available for meaSince in (2) the pressure variable
surement, the variable structure switching is based only on the
measured variables, namely, and . A full-order observer is
employed, where the estimates of and are used to improve
the observer performance.
The observer dynamics is derived from (2) in the form [15]

493

The condition for reachability of the th sliding surface

0 by the error trajectories starting from anywhere in the
state space is given by
(15)
From (15), we have the conditions

which can be used to choose the gains and . Note, that since
are unknown in the second of the above equations, has to
be set sufficiently large.
During sliding motion, the conditions
(16)
are satisfied, so that from (13) and (14), we get

The dynamics of the observed pressure error in the sliding phase

and its stability can be determined using the equivalent control
approach [15].
During sliding, from the second equation of (16)

where is an equivalent switching vector, replacing the discontinuous switching vector

(12)
and
are the position and velocity
where
observation errors, respectively.
Subtracting (12) from (2), we can obtain the observation error
dynamics as

and we have used the fact that

Substituting

we obtain the equivalent dynamics on the reduced order-sliding

manifold as

From the third equation of the above, we obtain

(13)
are chosen so as to place the eigenThe continuous gains
values of the 3 3 system matrix in (13) at desired locations in
the left-half plane.
as the state vector of the obSpecifying
server error dynamics, we choose the sliding surface as
(14)

(17)
We see from (17), that the observer pressure error can be sta, but this also has the effect of
bilized by setting a large
which acts as a
enlarging the effect of the estimation error
disturbance.
However, we can also see from (13) that a choice of suffigains means that in the reaching phase toward
ciently large

494

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002

the sliding plane the state variables , , and , are rendered

small. Once in the sliding phase, any marginal increase in the
has an effect of pulling the variables and
magnitude of
away from the sliding plane (
0,
0). At this stage, the
switching control satisfying the condition (15) pushes the states
back toward the sliding surface.
From the above arguments, we can conclude that the observed
either remains bounded according to (17) in
pressure error
the case of small , or remains bounded at levels corresponding
to switching around the sliding plane in the case of large .
is ultimately bounded and the magnitude of its
As a result,
ultimate bound is determined by the error in estimation of
and the gains ,
and , as well as the ratio
.
may be made
It may be noted that the continuous gains
zero and the reachability conditions following (15) satisfied by
and . This simplifies the observer design conincreasing
siderably, though increasing and also has the effect of increasing the chattering.
The observer can be further simplified by eliminating the first
in (12). The bounded stability of pressure
equation and
observation error can be established as before.

From (18), we have

where

We specify the control algorithm in the form

(21)
(22)
where

.
if
if

are switching feedback gains and , are relay gains.

are feedforward inputs obtained from

(23)
and

VI. SLIDING-MODE CONTROLLER

The partial observability of cylinder pressure dynamics implies that pressure observation errors do not converge to zero,
but are ultimately bounded. In this case, we consider the design
of the sliding-mode controller to ensure that the control performance is robust to observation errors. For simplicity, we consider the case of position control [5].
as the state vector, we choose
Defining
a linear, time-invariant sliding surface as
(18)

Accordingly, we have

As discussed in [5], we can obtain the switching gains as

. Here,
is the desired endpoint for
where
and
are the position
piston displacement.
and velocity errors, respectively.
, when the system is in sliding
Defining
0, so that
mode,

and

This implies that during the sliding phase the system dynamics can be expressed as

are chosen of the same sign as

where for robustness,
.
The effect of observation errors on control performance can
be compensated for by choosing relay gains as [17]

(19)
) to ensure that the poles of the
Therefore, by choosing (
second-order system are sufficiently to the left of the imaginary
axis in the complex plane, we can ensure that the errors and
are ultimately bounded. The size of the error bounds can be
by employing the
determined as a function of the bound on
Lyapunov function approach.
We next choose the switching gains to guarantee that the
sliding surface is reached from anywhere in the state space.
The condition for reachability of the sliding surface is
(20)

(24)

(25)

(26)
(27)
is given by (17).
where
Note that the inequalities (26)(27) can be satisfied over
the time interval of interest by choosing the magnitudes of
relay gains sufficiently large. This is always possible, since
the cylinder position and chamber pressures are limited in
range, implying that the pressure estimation errors are bounded
too. Setting relay gains too large, however, may increase the

PANDIAN et al.: PRESSURE OBSERVER-CONTROLLER DESIGN FOR PNEUMATIC CYLINDER ACTUATORS

Fig. 2.

Schematic of the experimental system.

495

Fig. 3. Estimation of P .

chattering of the control and state variables in which case it is

necessary to employ dual mode control by applying high gain
control in the vicinity of the sliding surface [17].
VII. EXPERIMENTAL RESULTS
The effectiveness of the proposed observer designs is illustrated by results of implementation on an industrial doubleacting piston-type cylindrical actuator (SMC CM2L32-500). A
schematic of the experimental system is shown in Fig. 2.
The external payload is moved by the piston through a linear
guide and the air flow to the cylinder chambers is regulated by
two-way three-port magnetic-type proportional valves (SMC
VEF312). The piston displacement is measured using an incremental encoder and the chamber pressures are measured with
pressure sensors through 16-channel, 12-bit A/D converters.
The velocity signals were obtained by differentiating and
digitally filtering the position measurements.
The observer algorithm is implemented on a 200-MHz Pentium processor-based PC. The observer and controller programs
are written in C and implemented with a sampling period of 0.55
ms. Experiments were performed with the cylinder in the horizontal state. The maximum air supply pressure is 0.6 MPa and
the piston stroke is 0.5 m. The ram cross-sectional areas were
m
m .
and areas
In the case of the cylindrical actuator, the mass
are known accurately and the parameters can be calculated
fairly precisely. However, it is difficult to get an exact estimate
of the viscous friction and damping parameter . To circumvent
this problem, we tuned the model parameter value of so as
to provide the best estimation performance for a given set of
observer gains under changes in payload mass . The resulting
parameter value is expected to be close to the actual value of .
were calculated under assumptions of sonic
Parameters
flow, using estimates of valve constants and obtained from
valve frequency responses. Fig. 3 shows the result of pressure
estimation using the continuous-gain observer for the case of
is estimated, with
a step change in the piston endpoint.
being measured using the pressure sensor. The figure shows a
with the accomparison of the estimated chamber pressure
as measured with pressure sensor. The initial
tual pressure

Fig. 4.

Estimation of P : Large friction.

estimation error is zero, as both pressures start from the atmospheric pressure level. It can be seen that the observer tracks the
actual pressure value quite closely.
The proportional valves used in the experiment employed
dither-based friction compensation, yet a finite valve dead time
exists in practice due to valve stiction. The observer model does
not include this time delay in the effect of valve input on pressure variation. Therefore, we see that a major source of estimation error is the dead time of the valves, as a small time delay
results in a large estimation error.
Another comparison of the measured and estimated pressures
is shown in Fig. 4. In this case, we have deliberately reduced the
lubrication of the linear guide over which the cylinder moves.
Further, the bias voltages of the valves are larger than those used
in the case of Fig. 3.
In this case, we see that there are large fluctuations in pressure variation due to the high valve bias. However, the increased
cylinder friction effectively filters out any potential effect of the
pressure fluctuations on cylinder motion (see Fig. 5). Since the
cylinder model used in the observer design does not incorporate
this friction effect, the estimated pressure shows only limited
fluctuations in its variation. Also, unlike in the case of Fig. 3,
there is a finite steady-state observation error due to cylinder
friction.

496

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002

Fig. 7. Estimation of P : Trajectory control.

Fig. 5. Cylinder motion under low/high friction.

Fig. 8.

Fig. 6.

Estimation of P : Trajectory control, Large friction.

Position control response.

To verify the effectiveness of the pressure observer for

control, the sliding-mode controller proposed in the previous
section was implemented for positioning of the cylinder. The
problem was to move the piston endpoint from an initial position of 0.1 m to a final position of 0.4 m. Both feedback control
and feedforward compensation are employed, using measured
position, velocity obtained by numerical differentiation and the
measured/observed pressures.
It should be noted here that the level of performance obtained
using the pressure observer depends partly on the pressure dif) feedback gain used in the controller.
ference (
gains are used the response under the observer-conIf low
troller combination will be quite close to that obtained with presand . However, this controller fairly
sure sensors for both
resembles a PD controller which would perform poorly when
payload is changed. In the latter case, it is necessary to choose
gains. The resulting position control responses are
higher
compared in Fig. 6. The deterioration in response in the case
when the observer is used is mainly the result of estimation errors due to cylinder and valve friction.
Figs. 7 and 8, show the performance of the pressure observer
in the case of tracking a cycloidal reference trajectory. Again,
the observer responses are fairly close to the actual levels.

Fig. 9. Trajectory control response.

Unlike in the case of position control, the feedback and feedforward control are initially insufficient to overcome the effects
of static friction in the cylinder and valves. The cylinder friction also contributes to large pressure estimation errors initially
as well as in the final stages when the velocity approaches zero.
In the case of Fig. 8, the observer is not able to track the high frequency fluctuations in actual pressure which were deliberately
induced in a manner similar to that for Fig. 4.
Finally, Fig. 9 shows the performance of the observer-based
controller in the case of tracking a cycloidal reference trajec-

PANDIAN et al.: PRESSURE OBSERVER-CONTROLLER DESIGN FOR PNEUMATIC CYLINDER ACTUATORS

Fig. 10.

Estimation of P : Sliding observer.

Fig. 11.

Estimation of

1P : Sliding observer.

tory. The response is close to the reference trajectory and the

actual trajectory achieved with measured pressures. The observer-based response is slightly oscillatory and results in a significant steady-state error.
Figs. 1012 illustrate the results of the sliding-mode pressure
observer. The results are qualitatively similar to those obtained
using the reduced-order observer. Here, the feedback gains are
made high so that the resulting chattering of inputs introduces
a fluctuation in pressure. The cylinder friction, however,
smoothes out any potential effect of this pressure variation on
piston movement. As the observer model does not incorporate
cylinder friction, the observer is not able to reproduce this
fluctuation in pressure (Fig. 10).
A comparison of the measured and estimated pressure differences is shown in Fig. 11. The estimation is seen to be fairly accurate. Here, we see that the main source of error is the cylinder
friction. At the start of the motion, there is a small time delay
due to static friction in the cylinder. While the pressure difference starts to build up during this delay period, the estimated
pressure difference remains zero since the velocity and accelerand
ation are zero [see (10)]. As the errors in estimation of
are out of phase, we see from (11) that the error in estima-

Fig. 12.

Estimation of P : Sliding observer.

Fig. 13.

Comparison of estimation errors: Sliding observer.

497

tion of pressure
is larger than both the errors
and
.
deviates signifAs a result, in Fig. 12 the estimated pressure
icantly from the actual pressure .
Fig. 13 shows the time trajectories of the estimation errors
in chamber pressures and pressure difference. The contribution
and
, respectively, to the
of the phase lead and lag of
as seen in Fig. 12 is clear.
relatively large magnitude of
In Figs. 11 and 13, we can see the presence of noise in the
pressure measurements and consequently its effect on the estimated pressure. The position control responses with use of
observer/sensors are compared in Fig. 14. Both feedback and
feedforward control have been employed. The deterioration in
response in the case when the observer is used may again be
attributed to the estimation errors due to cylinder and valve friction.
Fig. 15 shows the performance of the sliding-mode observerbased controller in the case of tracking a cycloidal reference trajectory. The observer-based tracking response is close to that
obtained with pressures sensors. The observer-based response is
slightly oscillatory and results in a noticeable steady-state error.
The similarity of response to that of the continuous-gain observer using a single pressure sensor is clear.

498

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002

Fig. 14.

Position control: Sliding observer.

Fig. 15.

Trajectory control: Sliding observer.

VIII. DISCUSSIONS
In this paper, we have shown that the limitation of partial
observability of chamber pressures can be overcome in practice by use of numerically estimated piston acceleration. A
sliding-mode controller with robust performance in the face of
bounded pressure observer errors can be designed by setting
both the switching surface parameters and the relay gains sufficiently high. Since we have employed a simple sliding-mode
control algorithm with a linear, time-invariant sliding surface
and nonadaptive switching/relay gains here, problems arising
from actuator saturation and high-frequency chattering can
be avoided by developing a more sophisticated sliding-mode
controller with an adaptive or nonlinear switching surface and
adaptive switching/relay gains.
The difficulties traditionally associated with the use of
numerically estimated acceleration variable have largely
been eclipsed in recent years by the availability of low-cost,
high-speed computing power. While we have used a simple
second-order Butterworth filter in our experiments for digital
filtering of numerically estimated velocity and acceleration,
it is also possible to use more advanced techniques such as
predictive filtering to achieve low-noise, low-delay numerical
differentiation (e.g., [18]). In this study, as the sampling period
is very small we have found that nominal changes in filter

cutoff frequencies do not contribute much to estimation delay,

as compared to the friction effect.
The experimental results presented in this paper demonstrate
that the main obstacles to the precise estimation of pressure in
pneumatic actuators are the friction forces in the actuators as
well as the valves. The friction forces also impose a limit on the
level of precision possible using advanced control methods for
pneumatic actuators [19]. Therefore, research on modeling and
identification of friction forces in actuators and valves is necessary to improve the performance of observers and controllers
for pneumatic systems.
Variations in pressure supply (set by valve bias voltages) affect control performance as well as energy efficiency of the actuator. As the results presented here indicate, pressure estimation too is very sensitive to pressure supply variations. Therefore, it is essential to incorporate this factor into future models
used in observer design. Though we have assumed sonic flow
conditions in this study to simplify the theoretical treatment, in
experiments we have found that use of parameter values of
based on estimated pressures yields satisfactory results, even
when this assumption is violated.
The work so far has assumed that the parameters, especially
payload mass , are known precisely. In cases where the payloads are variable and robust performance is required, design of
adaptive observers is necessary. Further, by combining the proposed pressure observer with models for estimation of cylinder
friction, it is possible to implement hybrid position/force control algorithms without the need for high-cost force sensors.
Due to their advantages of compliance and simplicity, pneumatic actuators are increasingly used for multijoint mechanisms
in rehabilitation and prosthetics. The use of pressure observers
proposed here can result in considerable reduction in the cost,
weight, energy consumption and complexity of such systems,
particularly since high levels of observer precision are not essential in these applications. The pressure observer proposed here
has also been implemented successfully in the case of rotary actuators of the first two links of a pneumatic SCARA robot [20].
IX. CONCLUSION
In this paper, we have proposed design methods for the
estimation of pressure variables in the case of pneumatic
cylinder actuators. The proposed methods are simple and easy
to implement and the accuracy is mainly limited by the friction
in cylinder and control valves. Several problems for future
research have been identified. Results of experimental implementation on an industrial-type pneumatic cylinder illustrate
the effectiveness of the proposed observers.
REFERENCES
[1] Y. Hayakawa, M. Tamai, S. R. Pandian, T. Isaka, and S. Kawamura, Development of an active orthosis for standing up motion by using hexahedron rubber actuator, in Proc. Japan-France/Asia-Europe Congr.
Mechatronics, Kitakyushu, Japan, 1998, pp. 6770.
[2] H. Murrenhoff, Y.-S. Hong, and A. Czinki, Component design and
control strategies for dexterous robot hands with servopneumatic finger
joints, in Proc. Int. Conf. Advanced Mechatronics, Okayama, Japan,
1998, pp. 275280.
[3] P. Gorce and M. Guihard, On dynamic control of pneumatic bipeds,
J. Robot. Syst., vol. 15, pp. 421433, 1998.

PANDIAN et al.: PRESSURE OBSERVER-CONTROLLER DESIGN FOR PNEUMATIC CYLINDER ACTUATORS

[4] J. E. Bobrow and B. W. McDonell, Modeling, identification and control of a pneumatically actuated, force controllable robot, IEEE Trans.
Robot. Automat., vol. 14, pp. 732742, Oct. 1998.
[5] S. R. Pandian, Y. Hayakawa, Y. Kanazawa, Y. Kamoyama, and S. Kawamura, Practical design of a sliding-mode controller for pneumatic actuators, ASME J., Dyn. Syst., Meas. Control, vol. 119, pp. 666674, 1997.
[6] S. R. Pandian, F. Takemura, Y. Hayakawa, and S. Kawamura, Practical
design of adaptive sliding-mode control of pneumatic actuators, Fluid
Power Syst.: J. Japanese Hydraul. Pneumat. Soc., vol. 31, pp. 107114,
2000.
[7] S. R. Pandian, Y. Hayakawa, K. Ieda, Y. Kamoyama, and S. Kawamura,
Modeling and control of a pneumatic rotary actuator, in Proc. Int.
Workshop Power Transmission and Motion Control, PTMC , Bath, U.K.,
1998, pp. 363377.
[8] S. R. Pandian, F. Takemura, Y. Hayakawa, and S. Kawamura, Control
performance of an air motor: Can air motors replace electric motors?,
in Proc. IEEE Int. Conf. Robotics and Automation, Detroit, MI, 1999,
pp. 518524.
[9] S. Bourdat, E. Richard, and S. Scavarda, A modified linear reduced
observer for a pneumatic servodrive, in Proc. 4th Bath Int. Fluid
Power Workshop: Fluid Power Systems Modeling and Control, Bath,
U.K., 1991, pp. 357371.
[10] T. Noritsugu and M. Takaiwa, Motion control of parallel link manipulator using disturbance observer, in Proc. Japan/U.S. Symp. Flexible
Automation, Boston, MA, 1996, pp. 173180.
[11] F. Takemura, S. R. Pandian, Y. Hayakawa, and S. Kawamura, Observer
design for control of pneumatic cylinder actuators, in Proc. Bath Int.
Workshop Power Transmission and Motion Control PTMC, Bath, U.K.,
1999, pp. 223236.
[12] H. Nijmeijer and T. I. Fossen, New Directions in Nonlinear Observer
Design, H. Nijmeijer and T. I. Fossen, Eds. London, U.K.: SpringerVerlag, 1999.
[13] B. L. Walcott, M. J. Corless, and S. H. Zak, Comparative study of
nonlinear state observation techniques, Int. J. Control, vol. 45, pp.
21092132, 1987.
[14] P. R. Belanger, Control Engineering: A Modern Approach. Fort Worth,
TX: Saunders, 1995.
[15] J.-J. E. Slotine, J. K. Hedrick, and E. A. Misawa, On sliding observers
for nonlinear systems, ASME J., Dyn. Syst., Meas. Control, vol. 109,
pp. 245252, 1987.
[16] FI. Haskari, U. Ozguner, and V. Utkin, On sliding-mode observers via
equivalent control approach, Int. J. Control, vol. 71, pp. 10511067,
1998.
[17] S. R. Pandian and M. Hanmandlu, A model-based sliding mode-controller for robot manipulators, Int. J. Robot. Automat., vol. 10, pp.
2934, 1995.
[18] O. Vainio, Adaptive derivative estimation for delay-constrained acceleration measurement, Trans. IEEE Trans. Ind. Electron., vol. 46, pp.
933935, Oct. 1999.
[19] J. Pu, C. B. Wong, and P. R. Moore, Methods of overcoming friction
effects in the control of pneumatic servo systems, in Proc. ASME Int.
Mechanical Engineering Congr., San Francisco, CA, 1995, pp. 139144.
[20] H. Mizutani, F. Takemura, T. Ogata, M. Inoue, S. R. Pandian, Y.
Hayakawa, and S. Kawamura, Practical design of sliding-mode control
of pneumatic robots: Verification of pressure observer and model-based
control (in Japanese), in Proc. Annu. Meet. Robotics Soc. Japan,
Kusatsu, Japan, 2000, pp. 171172.

499

Shunmugham R. Pandian (M95) received the

Bachelors degree in electronics and communication
engineering from Madurai-Kamaraj University,
Madurai, India and the Masters and Ph.D. degrees
in electrical engineering from the Indian Institutes of
Technology, in Kanpur and New Delhi, respectively.
From 1992 to 2000, he was with the College of
Science and Engineering, Ritsumeikan University,
Kusatsu, Japan as a Research Associate and a
full-time Lecturer. Currently, he is an Assistant
Professor in the Engineering Science Program at
the University of Michigan-Flint. His fields of interest include robotics and
mechatronics, control, pneumatics, low-cost renewable energy systems and
rehabilitation engineering.

Fumiaki Takemura received the B.S. degree in engineering from the Kyushu Institute of Technology,
Fukuoka, Japan, in 1996, the M.S. degree from
Ryukyu University, Okinawa, Japan, in 1998, and the
Ph.D. degree from Ritsumeikan University, Kusatsu,
Japan, in 2001. His Ph.D. dissertation was on the
control of pneumatic motors and the development of
a hybrid pneumatic/electric motor.
During 2001, he did research on sensors at Matsushita Electric Industrial Company, Ltd. Currently,
he is a part-time Lecturer at Ritsumeikan University.
He is currently working on the development of a wearable system using pneumatics.

Yasuhiro Hayakawa (M95) was a Research

Associate from 1985 to 1990, an Assistant Professor
from 1990 to 1993, and is currently an Associate
Professor in the Electrical and Control Engineering
Department, Nara National College of Technology,
Nara, Japan.
Dr. Hayakawa is a member of the Robotics
Society of Japan (RSJ), Japan Society of Mechanical
Engineers (JSME), and the Society of Instrument
and Control Engineers (SICE). His main fields of
interest are pneumatic drive systems, rehabilitation
engineering, and energy-efficient systems.

Sadao Kawamura (M89) received the B.S. degree

in biophysical engineering and the M.S. and Ph.D.
degrees in mechanical engineering all from Osaka
University, Osaka, Japan, in 1981, 1983, and 1986,
respectively.
From 1986 to 1987, he was an Assistant Professor
at Osaka University. From 1987 to 1994, he was
an Associate Professor at Ritsumeikan University,
Kyoto, Japan. Since 1995, he has been a Professor
in the Department of Mechanical Engineering,
and since 1996, a Professor in the Department
of Robotics, at Ritsumeikan University. From 1998 to 2000, he was a
Vice-Director of the Research Organization of Science and Engineering at
Ritsumeikan University. Currently, he is the Dean of Academic Affairs in
the Biwako-Kusatsu Campus at Ritsumeikan University. His current research
interests are in robotics and human motion science.
Dr. Kawamura received the Society of Instrument and Control Engineers
(SICE) Best Paper Award in 1987 and 2000, Institute of Systems, Control and
Information Engineers (ISCIE) Best Paper Award in 1987, Fluid Power Technology Promotion Foundation Best Paper Award in 1992, and the Good Design
Special Award in 2001. From 1999 to 2000, he was a trustee of the Robotics
Society of Japan (RSJ).