732 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 14, NO.

5, OCTOBER 1998

Modeling, Identification, and Control of a

Pneumatically Actuated, Force
Controllable Robot
James E. Bobrow and Brian W. McDonell

Abstract—This research focuses on modeling and control of a Several controllers have been used in the past for pneumatic
light-weight and inexpensive pneumatic robot that can be used systems. Most have been fixed gain linear controllers based
for position tracking and for end-effector force control. Unlike on a nominal transfer function model obtained by linearizing
many previous controllers, our approach more fully accounts for
the nonlinear dynamic properties of pneumatic systems such as the air flow dynamics about the cylinder midstroke position
servovalve flow characteristics and the thermodynamic properties [1]–[3]. Later work by Liu and Bobrow [4] has used a
of air compressed in a cylinder. We show with theory and linearized state space model to develop an optimal regulator
experiments that pneumatic actuators can rival the performance for a fixed operating point. In order to deal with the un-
of more common electric actuators. Our pneumatic robot is
certainties and the highly nonlinear behavior of pneumatic
controlled by extending existing manipulator control algorithms
to handle the nonlinear flow and compressibility of air. The systems, a number of approaches have also been developed
control approach uses the triangular form of the coupled rigid that incorporate some form of learning. Pu and Weston [5]
body and air flow dynamics to establish path tracking. In addition describe an algorithm which is trained to provide feed-forward
to the trajectory tracking control law, a hybrid position/force signals to optimize various point to point motions. McDonell
control algorithm is developed. The experimental results indicate
that the tip forces on the robot can be controlled without the need and Bobrow [6] use a real time identification scheme to
for an expensive force/torque sensor usually required by electric identify a locally linear time-varying model for the system
motor driven systems. about arbitrary reference trajectories.
Index Terms— Force control, hierarchical control, modeling, In the pursuit of better performance in pneumatic systems
nonlinear control, pneumatic control systems. for less cost, research has been done on different hardware
configurations. One component that is costly and difficult to
model for pneumatic controllers is the servovalve. A possible
I. INTRODUCTION substitute is to replace the variable flow servovalve with a

W HERE originally nearly all controlled motion was

done using electric motors, as computers have become
more powerful, other forms of actuation have become fea-
linear or rotary solenoid valve. Using a solenoid valve, a
pulse-width modulated control law that has a fixed-gain inner
pressure loop and an outer position loop was developed in
sible for providing motion. In addition, as the complexity Lai et al. [7]. Other controllers using solenoid valves based
of robotic tasks has increased, position control alone has on sliding mode control theory have been developed by Tang
not been sufficient. Useful interaction with the environment and Walker [8] and by Paul et al. [9]. A reduced order sliding
requires producing controlled forces in addition to moving the surface was used for the second approach to eliminate the need
robot with a prescribed motion. Pneumatic actuation, originally for pressure feedback. A novel rotary type air flow control
discounted as not being suitable for anything other than motion valve and its control system has been developed by Kunt and
between two hard stops, is now becoming a reasonable sub- Singh [10]. In this work, a pulse-width modulated form of
stitute for electric actuation in some applications. Pneumatic control is applied to a linear time-varying system model with
actuators are of interest for robotic applications because of good results.
their large power output at a relatively low cost. They are More recently, nonlinear control laws have been developed
also clean, easy to work with, and lightweight. In addition, using the full nonlinear dynamics of pneumatic systems using
compressed air is readily available at nearly every industrial feedback linearization techniques [11], [12], or Lyapunov
facility. Unfortunately, position stabilization of a pneumatic stability arguments [13]. When compared to control laws based
actuator is difficult if a high bandwidth closed-loop system on linearization, the nonlinear control laws obtained with these
is desired. approaches dramatically extend the range of stability. The
Manuscript received September 4, 1997; revised July 13, 1998. This work basic form of the control law obtained in the above references
was supported by the Rockwell International Graduate Fellowship Program, by is similar to that obtained in our research. That is, they all
NSF Grant MSS-9006667, and by Parker-Hannifin, Parker-Bertea Aerospace
Division. This paper was recommended for publication by Associate Editor require a model of the air flow dynamics, and they all use air
K. Tanie and Editor A. De Luca upon evaluation of the reviewers’ comments. pressure feedback in a level of control that drives the pressure
J. E. Bobrow is with the Department of Mechanical and Aerospace dynamics to a desired output. As suggested by Kawamura et
Engineering, University of California, Irvine, CA 92697 USA.
B. W. McDonell is with ATL Products, Irvine, CA 92614 USA. al. [13], we develop a natural separation between the rigid
Publisher Item Identifier S 1042-296X(98)07435-7. body dynamics and the actuator dynamics in a hierarchical
1042–296X/98$10.00  1998 IEEE
BOBROW AND MCDONELL: MODELING, IDENTIFICATION, AND CONTROL 733

control approach. We focus on obtaining an accurate air flow

model that can be inverted to implement any of these nonlinear
control approaches.
Although previous applications of pneumatic actuators to
robotics have primarily used linearized controllers, their low
cost, light weight, and ability for output force control makes
them attractive alternatives to electric motors. For instance, an
important early use of pneumatics was for tendon actuation
of the Utah/MIT dexterous hand [14]. Later work on accurate
force control of these tendons was conducted by Henri and
Hollerbach [15] using a linearized model of air flow through a
jet-pipe. Ben-Dov and Salcudean [16] also developed a force
controllable pneumatic system with low friction cylinders that
can be used for robotic teleoperation applications. Their con-
troller is based on a linearized model of flow through a flapper Fig. 1. Schematic diagram of a pneumatic actuator.
valve. Another application of pneumatics is a three degree of
freedom wrist and linear controller presented in Pfruendschuh
et al. [17]. This system uses the compressibility of air to obtain each cable of approximately 700 lb (3100 N). The radius of
natural compliance on a parallel tripod platform. Chirikjian each pulley was chosen to give the robot a maximum payload
[18] has also exploited the low cost feature of pneumatic capacity of 20 lb (89 N) in the fully outstretched configuration.
systems in the development of “binary manipulators” which The total weight of the robot, including the valves and base is
use a large number of pneumatic actuators in a state of full approximately 50 lb (220 N). An interesting property of this
extension or full contraction. pneumatic robot is that it can hold an object in the outstretched
Our work has the advantages of the above mentioned position indefinitely with no heating or power loss. This is
servovalve actuated pneumatic systems and as mentioned because once the correct mass of air has entered the cylinders
earlier, is not limited by the choice of operating point as to balance the load, no further air flow or valve control signal
are the controllers based on linearization. In the following, is needed.
we first derive a nonlinear dynamic model of the actuator The robot was controlled using a 486 PC. The feedback
dynamics and then demonstrate with experimental evidence sensors for each joint included a piezoresistive pressure trans-
that common assumptions made for the valve flow dynamics ducer on each side of the piston (two sensors needed on each
are not accurate. Next we combine the actuator dynamics with joint) and a 432 000 pulse/revolution position encoder. The
the nonlinear rigid body dynamics of a 3R robot to obtain a control signal was a voltage obtained from the PC through a
model of the coupled system. We then develop a model based 12 bit D/A converter. This voltage was converted to a valve
controller that provides exponentially stable path tracking and current using an analog voltage to current loop. The servovalve
output force control for the coupled system using the full was a standard jet-pipe type valve [19] with a maximum flow
nonlinear dynamics. Finally experiments are presented that rate of 10 cfm (283 liter/min) at a pressure drop of 100 psi
demonstrate the strengths of our approach and its limitations (689 kPa) across the valve. Although the supply pressure was
due to realistic considerations such as friction and valve regulated to 100 psig, a pressure transducer was also used
flow saturation. in the feedback control loop that measured the instantaneous
supply pressure near the inlet to the servovalves.
II. ROBOT DESCRIPTION
The experimental 3R robot used in this research is shown
in five different positions in Fig. 5. Fig. 1 shows the basic III. PNEUMATIC ACTUATOR MODEL
configuration of each of the three actuators used. Each actuator A dynamic model for the behavior of air in the cylinder
consists of a double acting pneumatic cylinder driving, via is developed by following Shearer [1] and Liu and Bobrow
cables, a rotary joint. A control current is sent to a servovalve [4]. In their approach, by using conservation of energy, the
to control the spool position that creates an air flow into or relationship between mass flow rate of air and the change
out of the two cylinder chambers. of pressure in chamber A (or chamber B) can be found.
Joint 1 of the robot is attached to the robot base and creates The internal energy of the mass flowing into the system is
a rotation out of plane for the shoulder-elbow configuration (these variables are defined in the Appendix) the
of joints 2 and 3 which are both 18 in (45.7 cm) in length. rate at which work is done on the moving piston is
A unique feature of the robot is that the structural portion of the rate of heat transfer through the wall is and the rate
each link is also the pneumatic cylinder that provides rotary of change of the internal energy of the air in the cylinder
actuation for the subsequent link. That is, the cable in Fig. 1 is Here is the mass flow rate of
drives the right-hand pulley, and this pulley is attached to the air to chamber A, and are the constant volume and
next cylinder in the kinematic chain. Each cylinder has a bore constant pressure specific heats of air, is the universal gas
of 3 in (7.6 cm) and is operated at a supply pressure of 100 constant, is the volume of chamber A, is the air supply
psig (689 kPa) which gives a maximum linear force output on temperature, and is the cylinder pressure. Summing these
734 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 14, NO. 5, OCTOBER 1998

terms in an energy balance yields Substituting (6) and (7) into (8) and separating into the
terms affected by the servovalve spool position and the terms
(1) which are functions only of the position and velocity of the
joint we obtain
For this actuator shown in Fig. 1, the cylinder volumes are
related to the pulley position by
(2) (9)
(3) or
where is the cylinder cross-sectional area, is the stroke, (10)
is the radius of the pulley that rotates the pulley, and
are the fixed volumes at the ends of the stroke, and is where
the angle of rotation of the pulley with 0 when the piston
is at the left end of the cylinder.
The rate of heat transfer is governed by the temperature (11)
difference between the inside and outside of the cylinder and
the coefficients of thermal conductivity of the stainless steel
cylinder and the aluminum end-caps. For the experimental (12)
system the thermal time constant for this heat transfer process
is on the order of several minutes, which was assumed to be and we have assumed that and are functions of and
much slower than the response time of the robot. We therefore the cylinder pressures.
assume that 0. Experimental studies related to the validity To more easily represent multiple degree of freedom pneu-
of this assumption have been reported in [20]. With 0, matic systems, these equations will now be considered -
expand (1), and use the fact that to obtain vectors with each element corresponding to each independent
joint and actuator, i.e.,
(4)
.. .. ..
. . . (13)
(5)

(6) IV. VALVE FLOW CHARACTERIZATION

The control law developed in the following section requires
Similarly for the side of the cylinder knowledge of all of the parameters and states of the air flow
dynamics in (9). Most of the parameters used in (9) are easily
(7) measured such as the cylinder area and the pulley radius. The
most difficult quantities to characterize are the mass flow rates
The mass flow rates, and are nonlinear functions of through the valve. For the purpose of modeling, it is often
the servovalve spool position and the supply and cylinder assumed that the mass flow of air through the servovalve
pressures. In addition, the servovalve has dynamics relating closely matches that of airflow through an orifice mounted
the control current to spool position which have a 3 dB in a pipe where the orifice area varies with control current.
amplitude ratio and a 90 phase lag at 40 hertz as described We will call this the “theoretical” flow rate in what follows.
in the manufacturers literature [19]. To simplify the model, The equations governing mass flow through an orifice with no
the spool valve position is assumed to be directly proportional losses are given in [21] (14), shown at the bottom of the page,
to the control current. The dynamics of the servovalve can be where is the effective valve orifice area, is the upstream
included in the formulation if one extends the levels of control pressure, is the downstream pressure, is the upstream
hierarchy described in Section V-B to three. temperature, and 1.4 for air.
The torque acting on the joint is related to the difference For the pneumatic system there are two situations which
in pressure between the two sides of the cylinder, so that are encountered:
or
1) a constant pressure supply flowing into the cylinder;
(8) 2) the cylinder exhausting into atmospheric pressure.

(unchoked)
(14)
(choked)
BOBROW AND MCDONELL: MODELING, IDENTIFICATION, AND CONTROL 735

Fig. 2. Curve fit for cp Ts m

_ a as a function of pa with a valve control Fig. 3. New function curve fit using the same data as Fig. 2.
current of 6 mA.

tally measured data to the theoretically predicted flow-rates,

In both the filling and exhausting situations it is possible to least squares surface fitting was done with different surfaces as
reach conditions of choked flow. We tested the validity of basis functions. Two surfaces are needed—one that represents
(14) using the following experimental procedure. First note cylinder filling from the supply and the other that represents
that (6) shows that for a fixed cylinder volume and constant cylinder exhausting to atmosphere. Our control law requires
is directly proportional to We measured the internal that these functions of cylinder pressure and servovalve current
energy of mass flowing into the system, rather than be able to be back-solved (i.e., one to one and invertable) for
since is needed for the control law. This was also the control that will produce a given at any given cylinder
done to eliminate the need for an upstream temperature sensor pressure.
on the assumption that in terms of absolute temperatures, the After many experiments and trials, the basis functions which
supply temperature will be changing very little. With yielded a good curve fit for the change in internal energy as a
chamber A was moved to its maximum volume configuration function of cylinder pressure were quadratic in
and exposed to atmospheric pressure. The chamber was then
closed and a constant control was sent to the servovalve. For filling (15)
this constant the pressure and the supply pressure exhausting (16)
were recorded at 200 Hz as the cylinder filled with
where and are coefficients found from a
air. Given the measured we then used a central
least squares fit of the experimental data. A different set
difference numerical derivative to obtain the corresponding
of coefficients was identified for each side of the cylinder
In this manner, is also known from (6). Note
since different valve orifices are used to produce flow in each
that although the supply pressure is usually assumed to be
chamber. Note that these equations use the upstream pressure.
constant, we found it better to measure the actual upstream
If this pressure is constant, then it need not be measured at each
pressure near the valve. The upstream pressure did fluctuate
sample. However, as mentioned previously, we found that the
during period of high flow transients.
upstream pressure dropped during periods of high flow, and
Given the above data, the only unknown in (14) is the
that it is helpful to measure this pressure at each sampling
effective orifice area We chose the unknown constant
instant. Fig. 3 shows the curve fit using the new function and
that produced the least mean-squared error from our
the same data that was used in Fig. 2.
measured data and the theoretical curve given by (14). A
The control law in the next section sets a desired value
typical resulting best fit solution for a given valve current
for at each sampling instant. The above flow equations
6 mA is shown in Fig. 2. The solid line is a plot of
must then be back-solved at each for the that produces
versus downstream pressure from (14) with this best
the desired This can be done since the entire state is
choice of and with the upstream pressure regulated at
measured, and the only unknown in (10) given is In
100 psig, or 790 kPa absolute. The ’s denote experimentally
order to solve for we first determine the desired and then
measured data points. The discrepancy between the curves is
insert and in (11) with the appropriate
probably due to the assumption used in the derivation of (14)
functions. If the desired is positive, the proper value of
that there are no losses in the flow through the orifice. Nearly
will be one that makes chamber A fill with air and chamber
all previous results on pneumatic control incorrectly assume
B exhaust air. Therefore will be replaced with (15)
(14) is true, or worse, assume a linear relation between and
for chamber A and with (16) for chamber B. If the
Note that as varies, a surface of flow curves of this form
desired is negative, will be replaced with (17)
are defined.
and with (15). Following these replacements, we
We measured a least squares surface fit by sweeping the
now have a quadratic equation which can be solved for
current though its range of possible operating values and
performing the above curve fitting procedure at 40 discrete
values. This produced 40 effective areas for which V. CONTROL LAW FOR THE PENUMATIC SYSTEM
varied approximately linearly with the valve current. Because As mentioned in the introduction, there are several choices
of the relatively poor match shown in Fig. 2 of the experimen- possible for the control law for the pneumatic system. In
736 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 14, NO. 5, OCTOBER 1998

Fig. 4. The left-hand plots are 2 and 5 Hz sin wave tracking using the proposed control law. The tracking error is shown below them. The right-hand
plots are results using the same inputs and a well-tuned proportional control law.

[22], three distinct approaches were analytically developed, with its derivative Substituting (17)
and tested experimentally. Once suitable feedback gains were into (10) and using the definition for we get
chosen, similar performance was achieved for all the ap-
(18)
proaches. The following control law is based on the stability of
hierarchical systems and it has the advantage that it is simple (19)
to implement. The control law is also similar to integrator Thus since the torque error, will converge to zero
backstepping [23], where the control law for one subsystem is exponentially.
the reference signal for another. We show that any control law The control law (17) has a feedforward term a
for standard torque controlled robots can easily be extended to term that accounts for motion of the cylinder and an
pneumatic systems as long as the control law is continuously expected proportional term. Although the proportional term
differentiable. appears to be a linear control, the inversion of the
flow equation (11) needed to obtain the servovalve current
A. Torque Control Subsystem is highly nonlinear. Fig. 4 shows some typical experimental
The first step in the control design is to control the actuator results for force tracking. In this case, the robot was held
torques. This controller will be the first subsystem in the stationary. The left-hand plots show the response obtained
hierarchical system. Joint torque control is necessary for any with our control law for joint 1. In the top left plot, the desired
robotic output force control operation. It will also bridge the torque is a 10 Nm amplitude, 1 Hz sin wave. As shown in
gap between the standard torque input controllers and the the plot below it, the torque tracking error is very small in this
pneumatic robot. Consider the equation relating the valve case. The next lower plot is the response to a 5 Hz sin wave
current and the actuator torque (10), Given any input. At 5 Hz, some deterioration in tracking performance
desired torque, and its time derivative, let was measured as shown on the bottom left hand plot. The peak
error was about 1.5 N or 15% of the input wave amplitude.
(17) The degradation was probably due to the fact that we have
ignored servovalve dynamics and valve current limitations.
Where is a gain matrix with The right hand plots show the system response using only
Now, using the procedure described in the last section, solve a standard proportional control law, i.e., with
(11) for the servovalve currents, which will produce this no nonlinear inversion of needed to get The gain
This controller can easily be shown to exponentially converge was chosen to be the largest value that produced a stable
to the desired torque as follows. Define the torque error as response in the operating range of the cylinder. The lower
BOBROW AND MCDONELL: MODELING, IDENTIFICATION, AND CONTROL 737

right hand plot shows that the standard proportional controller C. Inverse Dynamics Hierarchical Control
fails completely due to the large phase lag at 5 Hz. One of the most popular control laws for manipulators
Given an ideal system, our torque controller drives the is inverse dynamics, or computed torque control (see, e.g.,
torques to any differentiable desired torque One Spong and Vidyasagar [26]). In order to use inverse dynamics
should then be able to choose the desired torque to be the same control for the pneumatic system, assume the robot equations
as that needed to control a standard torque controlled robot. of motion are
The following analysis demonstrates how exponential stability
of the entire system can be established if one defines to (27)
be an inverse dynamics control law, or to be the control law
of Slotine and Li [24]. and let the desired torque be
(28)
B. Hierarchical Control
where Then with use the equations of
First we review a result from Vidyasagar [25] on the sta-
motion, (27) and (28) to obtain
bility of systems in hierarchical or triangular form. Theorem
“Hierarchical System Stability”: Consider a system in the
form (29)

Canceling terms, and using the fact that the inertia matrix is
invertible we get
..
. (30)
(20)
To fit the form of (20), this can be rewritten as
where each represents a vector. Suppose for each
the following conditions are satisfied:

(21) (31)

by defining
(22)
(32)
and there exists constants and such that
and recalling Note that because is
(23)
a function of and time, i.e., which is in the
form required for the theorem.
where Clearly is and its derivatives with respect to any
state variable are bounded. Also 0.
.. Since is the linear system
. (24)
and can easily be chosen to make an expo-
nentially stable equilibrium of the system. This subsystem,
Under these conditions, is a exponentially stable satisfies all the requirements necessary
equilibrium of the system (20) if and only if is a for a subsystem of a hierarchical system. Therefore by the
exponentially stable equilibrium of the systems hierarchical system stability theorem, the entire system is
exponentially stable.
(25)

for each In this theorem, is any norm, but D. Slotine Hierarchical Control
is the most convenient to use. The notation means The Slotine and Li [24] algorithm can also be used to control
an open ball of radius the robot. For the nonadaptive version, a procedure similar to
For our pneumatic robot, the hierarchy consists of the the above shows that the closed loop system dynamics are
torque control subsystem dynamics, which will be and
the controlled robot dynamics, which will be These two (33)
subsystems will be shown to satisfy the criteria for stabilization
where
of a hierarchical system. If we define then (19)
We can then rewrite the dynamics as (34) and (35), shown
becomes
at the bottom of the next page. The function is continuously
(26) differentiable and its derivatives with respect to all state
variables are bounded. The criteria of and
Because of its simple linear structure, it is clear that it satisfies being an exponentially stable equilibrium of the system
all the criteria for a subsystem in a hierarchical system. are also satisfied, therefore the complete
738 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 14, NO. 5, OCTOBER 1998

(a) (b) (c)

(d) (e)
Fig. 5. Five positions used for testing three degree of freedom algorithms.

system of and is exponentially stable. The adaptive where is the Jacobean matrix defined by and
version of the Slotine and Li algorithm requires the definition is a desired Cartesian acceleration determined from an outer
of a third level of the hierarchy to handle the adaptation position loop. After substituting this in to the dynamics of the
dynamics. robot in (27) with the additional term to represent
contact with the environment, the closed loop dynamics with
E. Hybrid Position/Force Control 0 become The task space motion has been globally
linearized and decoupled allowing the position and force
As robotic tasks become more advanced, there will be a controllers to be designed independently. For our experiment
greater need for robots which don’t simply move without we will assume position control in the and Cartesian
contact with the environment. One of the advantages of a directions and force control in the direction. Let
pneumatic system is the ability to accurately produce and
monitor joint torques based on the cylinder pressures. For this
reason, an ideal application for a pneumatic robot is force
control. Typically in force control applications a force/torque (37)
sensor will be incorporated to close the force control loop.
Unfortunately, these sensors are not only costly but fragile.
If the sensor could be eliminated while still maintaining good where is the error in the Cartesian workspace coordinates.
force control, a savings in cost would be achieved with an This will control the robot motion in the plane. The force
increase in reliability. control normal to the plane will be provided through the
The control law used in the following experiments is a open loop force term with While many values
modified version of a hybrid position/force controller given of were tested, in the following experiments was set to
in Lewis et al. [27]. The desired joint torque is defined as a 50 N downward force in one test and to 220 N in another
test. The stability of this control law when combined with the
torque control subsystem is similar to the proof of the inverse
(36) dynamics controller of Section V-C.

(34)

(35)
BOBROW AND MCDONELL: MODELING, IDENTIFICATION, AND CONTROL 739

Fig. 6. Trajectory used for testing three degree of freedom algorithms.

Fig. 8. Error while tracking the above path with the hierarchical
torque-Slotine controller.

Fig. 7. Tracking the trajectory of Fig. 6 with the hierarchical torque-Slotine

controller.

VI. EXPERIMENTS
Numerous experiments were performed on UCI’s three de-
gree of freedom pneumatic robot to test the control algorithms.
A trajectory used to test the tracking control law is shown in
Fig. 5. The joint space motion that blends the five positions
(a)–(e) shown in Fig. 5 is shown in Fig. 6. The positions are
Fig. 9. Force/torque sensor attached to the robot.
labeled in the plot with vertical lines. The individual joint
trajectories are also labeled 1, 2, and 3 for the base, shoulder,
reduction used in an electric robot—our robot is a direct-
and elbow, respectively.
drive device and should be compared to these. One of the
The hierarchical inverse dynamics control law and the hier-
few experimental studies reported that attempts to quantify
archical Slotine control law were tested under many different tracking error is given in [28]. For the two robots considered
conditions including different sampling rates. A representative in that study and using their best control laws, the root-mean-
experimental result is shown in Fig. 7. This plot has the desired squared errors varied from about 1–3 , which is very close to
trajectory of Fig. 6 shown as a dotted line, and the actual the errors observed for our pneumatic robot.
trajectory shown as a solid line. Fig. 8 shows the tracking
error in degrees.
The tracking error should theoretically converge to zero, A. Tip Force Control
however due to modeling errors such as friction, air flow The tip force control law described in Section V-E was
limitations, and the fact that we have ignored servovalve tested on the three degree of freedom robot with a force/torque
dynamics, this was not the case. One might wonder how our sensor attached to the tip. Mounted to the sensor was a small
tracking errors compare to those observed for most electric fixture for holding a short Teflon rod. The attached sensor is
motor driven robots? This is not easy to answer since the shown in Fig. 9. This sensor was used only to monitor the
errors are trajectory dependent and there is no standard for forces exerted by the robot on the table, its output was not
comparison between robots. Another consideration is the gear sent to the controller to be used in a feedback loop.
740 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 14, NO. 5, OCTOBER 1998

Fig. 10. Position tracking of circle while applying a 50 Newton normal force. Fig. 13. Position tracking of circle while applying a 220 Newton normal
force.

Fig. 11. Z component of tip force from force/torque sensor.

Fig. 14. Z component of tip force from force/torque sensor.

Fig. 12. Z component of tip force from measured joint torques. Fig. 15. Z component of tip force from measured joint torques.

Figs. 10–15 show the results of two tests of tracking a robot configuration and the air supply pressure. Shown are the
circle in the plane while applying a 50 Newton (11 lb) desired and actual Cartesian trajectories, as well as the actual
and a 220 Newton (50 lb) downward force. The 220 Newton measured tip force from the force-torque sensor (not filtered)
force is just under the maximum possible force given the and the predicted tip force as computed from the joint torques
BOBROW AND MCDONELL: MODELING, IDENTIFICATION, AND CONTROL 741

and the dynamic model of the robot. The periodic nature of Mass flow rate of air into chambers A and
both plots is due to the fact that the tip of the robot traverses the B.
circle about three times during the test. The fact that the errors Air supply temperature.
repeat suggest that the errors could be removed by careful Absolute pressures in chambers A and B.
calibration. Note that in both output force plots even though Supply and exhaust pressures.
the force controller is completely open loop, the force is still Volume of chambers A and B.
relatively close to the desired value. Volumes not swept by piston in chambers
A and B.
Constant volume and pressure specific
VII. CONCLUSION heats of air.
The theory and experiments developed in this research For air this ratio is 1.4
demonstrate that the key to a high performance pneumatic Rate of heat transfer to the cylinder.
system is a good model. Through system modeling and ex- Universal gas constant.
perimental results, it was found that previous assumptions Joint torque.
regarding air flow through a servovalve were not accurate. Servovalve input current.
The experimental data for the air flow through the servovalve Valve flow constants for fill and exhaust.
did not agree with the predicted flow assumed by previous
researchers. This is not surprising due to the complex, three REFERENCES
dimensional flows present inside the servovalve. Using the [1] J. L. Shearer, “Study of pneumatic processes in the continuous control
experimental data, a new model was developed which much of motion with compressed air—i, ii,” Trans. ASME, pp. 233–242, Feb.
more closely matched the actual performance of the valve. Ex- 1956.
[2] C. R. Burrows and C. R. Webb, “Use of root loci in design of pneumatic
perimental tests of the model were performed which revealed servo motors,” Control, pp. 423–427, Aug. 1966.
that the model did perform well, and could be used to achieve [3] D. R. Vaughan, “Hot-gas actuators: Some limits on the response speed,”
ASME J. Basic Eng., pp. 113–119, Mar. 1965.
excellent torque tracking at low frequencies ( 5 Hz) with [4] S. Liu and J. E. Burrows, “An analysis of a pneumatic servo sustem and
some degradation in torque tracking due to flow saturation its application to a computer-controlled robot,” ASME J. Dynam. Syst.,
occurring at higher frequencies. Meas., Contr., vol. 115, no. 3, pp. 427–433, Sept. 1988.
[5] J. Pu and R. H. Weston, “A new generation of pneumatic servos for
Once the valve flow characteristics were determined, a industrial roots,” Robotica, vol. 7, pp. 17–23, Jan. 1989.
model based control approach was developed that permits the [6] B. W. McDonell and J. E. Bobrow, “Adaptive tracking control of an
use of most existing control laws for standard torque motor air powered robot actuator,” ASME J. Dynam. Syst., Meas., Contr., vol.
115, no. , pp. 427–433, Sept. 1993.
controlled robots. The pneumatic controller is cast into a hier- [7] J. Y. Lai, C. H. Meng, and R. Singh, “Accurate position control of a
archical system of torque controller and rigid-body dynamics pneumatic actuator,” ASME J. Dynam. Syst., Meas., Contr., pp. 734–739,
controller which can be treated as two separate subsystems Dec. 1990.
[8] J. Tang and G. Walker, “Variable structure control of a pneumatic
for stability analysis. The control law tested exhibited good actuator,” ASME J. Dynam. Syst., Meas., Contr., vol. 117, no. 1, pp.
trajectory tracking characteristics for the multiple degree of 88–92, Mar. 1995.
[9] A. K. Paul, J. K. Mishra, and M. G. Radke, “Reduced order sliding mode
freedom robot used in the experiments. control for pneumatic actuator,” IEEE Trans. Contr. Syst. Technol., vol.
One of the most promising capabilities of the pneumatically 2, pp. 271–276, Sept. 1994.
actuated robot system is its ability to produce tip forces [10] C. Kunt and R. Singh, “A linear time varying model for on-off valve
controlled pneumatic actuators,” ASME J. Dynam. Syst., Meas., Contr.,
from the measured joint torques. Using only open loop force pp. 740–747, Dec. 1990.
control in conjunction with a controller for tracking motion [11] A. Bouhal, E. Richard, and S. Scavarda, “An experimental comparative
tangential to a surface, the robot was able to produce a tip study of linear and nonlinear adaptive pressure regulation,” in Proc. 6th
Bath Int. Fluid Power Workshop Model. Simulat., 1993, pp. 225–238.
force without the use of a force/torque sensor to close the loop. [12] T. Kimura, S. Hara, T. Fujita, and T. Kagawa, “Control for pneumatic
This was performed on a rigid surface without a compliant end actuator systems using feedback linearization with disturbance rejec-
effector. In contrast to some previous experimental research, tion,” in Proc. 1995 Amer. Contr. Conf., Seattle, WA, June 1995, pp.
825–829.
no instabilities in tip force control were encountered. Hence, [13] S. Kawmura, K. Miyata, H. Hanafusa, and K. Isisa, “Pi type hierarchical
the natural compliance of the compressed air is advanta- control scheme for pneuamtic robots,” in Proc. 1989 IEEE Int. Conf.
Robot. Automat., Scottsdale, AZ, May 1989, pp. 1953–1858.
geous for force control applications. The experimental results [14] S. C. Jacobsen, J. E. Wood, D. E. Knutti, and K. B. Biggers, “The
achieved in this research demonstrate the remarkable potential utah/mit dextrous hand: Work in progress,” Int. J. Robot. Res., vol. 3,
of pneumatically actuated robot systems. no. 4, pp. 21–50, 1984.
[15] P. D. Henri and J. M. Hollerbach, “An analytical and experimental
investigation of a jet pipe controlled electropneumatic actuator,” in
Proc. IEEE Int. Conf. Robot. Automat., San Diego, CA, May 1994, pp.
APPENDIX 300–306.
LIST OF SYMBOLS RELATED TO THE ACUATATOR [16] D. Ben-Dov and S. E. Salcudean, “A force-controlled pneumatic actu-
ator,” IEEE Trans. Robot. Automat., vol. 11, pp. 906–911, Dec. 1995.
Angular position of the pulley. [17] G. H. Pfruendschuh, V. Kumar, and T. G. Sugar, “Design and control
of a 3 dof in-parallel actuated maniplator,” in Proc. Int. Conf. Robot.
Radius of the cable pulley. Automat., Sacramento, CA, 1991, pp. 1659–1664.
Length of cylinder stroke. [18] G. S. Chirikjian and I. Ebert-Uphoff, “Efficient workspace generation
Cylinder cross sectional area. for binary manipulators with many actuators,” J. Robot. Syst., vol. 12,
no. 6, pp. 383–400, June 1995.
Effective valve orifice area. [19] R. Atchley, Servovalve Model 211APN, Typical Performance Specifica-
A, B Reference to sides of cylinder. tions. Los Angeles, CA: Atchley Controls.
742 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 14, NO. 5, OCTOBER 1998

[20] Y. Kawakami, J. Akao, S. Kawai, and T. Machiyama, “Some consider- Brian W. McDonell received the B.S. and Ph.D. de-
ations on the dynamic characteristics of pneumatic cylinders,” J. Fluid grees in mechanical engineering from the University
Contr., vol. 19, no. 2, pp. 22–36, Sept. 1988. of California, Irvine, in 1989 and 1996, respectively.
[21] R. L. Daugherty, J. B. Franzini, and E. J. Finnemore, Fluid Mechanics His research interests are in the area of computer
with Engineering Applications. New York: McGraw-Hill, 1985. controlled mechanical systems. He has designed
[22] B. W. McDonell, Modeling, Identification, and Control of a Pneumati- novel controllers for electrically, pneumatically, and
cally Actuated Robotic Manipulator, Ph.D. dissertation, Dept. Mech. & hydraulically actuated systems as well as for internal
Aerosp. Eng., Univ. California, Irvine, 1996. combustion engines. He is currently working at ATL
[23] P. V. Kokotovic, “The joy of feedback: Nonlinear and adaptive,” IEEE Products, Irvine, CA, a manufacturer of automated
Contr. Syst. Mag., pp. 7–17, June 1992. computer data storage products.
[24] J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood
Cliffs, NJ: Prentice-Hall, 1991.
[25] M. Vidyasagar, Nonlinear Systems Analysis, 2nd ed. Englewood Cliffs,
NJ: Prentice-Hall, 1993.
[26] M. W. Spong and M. Vidyasagar, Robot Dynamics and Control. New
York: Wiley, 1989.
[27] F. L. Lewis, C. T. Abdallah, and D. M. Dawson, Control of Robot
Manipulators. New York: MacMillan, 1993.
[28] L. L. Whitcomb, A. A. Rizzi, and D. E. Koditschek, “Comparative
experiments with a new adaptive controller for robot arms,” IEEE Trans.
Robot. Automat., vol. 9, pp. 59–70, Feb. 1993.

James E. Bobrow received the M.S. and Ph.D.

degrees in engineering from the Mechanics and
Structures Department, University of California, Los
Angeles, in 1979 and 1983, respectively.
He is a Professor of Mechanical and Aerospace
Engineering at the University of California, Irvine
(UCI). After graduate school, he was a Senior
Programmer Analyst with McDonnell Douglas Au-
tomation Company, where he developed CAM soft-
ware for the Unigraphics system. In June 1984, he
joined UCI as an Assistant Professor, where he con-
ducted research in robotics and applied control systems. From 1991 to 1992,
he was a Visiting Associate Professor in the Computer Science Department,
Stanford University, Stanford, CA, where he investigated applications of
numerical optimization algorithms to learning systems. He has published over
70 research papers in robotics, applied control systems, numerical optimization
techniques, and computer-aided design.