Simultaneous Force and Stiffness

Control of a Pneumatic Actuator

Xiangrong Shen
Michael Goldfarb
Department of Mechanical Engineering,
Vanderbilt University,
Nashville, TN 37235
This paper proposes a new approach to the design of a robot actuator with physically
variable stiffness. The proposed approach leverages the dynamic characteristics inherent
in a pneumatic actuator, which behaves in essence as a series elastic actuator. By replac-
ing the four-way servovalve used to control a typical pneumatic actuator with a pair of
three-way valves, the stiffness of the series elastic component can be modulated indepen-
dently of the actuator output force. Based on this notion, the authors propose a control
approach for the simultaneous control of actuator output force and stiffness. Since the
achievable output force and stiffness are coupled and configuration-dependent, the au-
thors also present a control law that provides either maximum or minimum actuator
output stiffness for a given displacement and desired force output. The general control
and maximum/minimum stiffness approaches are experimentally demonstrated and shown
to provide high fidelity control of force and stiffness, and additionally shown to provide a
factor of 6 dynamic range in stiffness. 关DOI: 10.1115/1.2745850兴

1 Introduction
The modulation of actuator output stiffness can serve several
purposes in robotic applications, many of these motivated by bio-
logical motor control strategies. For example, research suggests
that humans achieve stable and effective interaction with a wide
variety of environments by leveraging their ability to modulate
joint impedance independently of joint torque 共e.g. 关1–4兴兲, and a
significant body of research exists that highlights the role of vari-
able compliance in enhancing the energetic efficiency of mamma-
lian locomotion 共e.g. 关5–7兴兲. As proposed by Salisbury 关8兴 and
generalized by Hogan 关9兴, one means of modulating actuator out-
put stiffness is via feedback control. Such an approach can effec-
tively modulate actuator 共or manipulator兲 output stiffness, but as
with all feedback control systems, provides the desired character-
istics in a limited frequency range, and can jeopardize system
stability, especially in cases of noncollocated load sensing. Addi-
tionally, closed-loop strategies offer little with respect to energetic
efficiency, since feedback control is generally energetically non-
conservative. In order to provide variable compliance without
these limitations, several researchers have developed robot actua-
tors with physically variable stiffness, which incorporate some
open-loop mechanism to enable simultaneous control of actuator
force and stiffness. Specifically, in order to implement a biologi-
cally inspired strategy of interaction during robotic manipulation
without the limitations imposed by feedback control, Laurin-
Kovitz et al. 关10兴 proposed the design of a variable stiffness ac-
tuator that is loosely based on the configuration of the human
musculoskeletal system. Their approach incorporates two 共non-
backdrivable兲 motors for a single joint 共similar to the agonist/
antagonist musculoskeletal relationships in animals兲, each actuat-
ing a tendon through a nonlinear spring. As with a biological
motor control system, the joint torque is a function of the differ-
ence of the motor efforts, while the joint stiffness is a function of
the sum of the motor efforts, thus providing simultaneous control
of joint torque and stiffness. Koganezawa and Yamazaki 关11兴,
Koganezawa and Ban 关12兴, and English and Russell 关13,14兴 pro-
posed a variable stiffness actuator design of a similar structure. In
order to leverage the energetic benefits of variable compliance for
robotic legged locomotion, Hurst et al. 关15兴 presented a different
design for a variable stiffness actuator. Like the previously de-
Contributed by the Dynamic Systems, Measurement, and Control Division of
ASME for publication in the JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT, AND CON-
TROL. Manuscript received January 9, 2006; final manuscript received January 5,
2007. Review conducted by Huei Peng.
scribed approaches, their design incorporates two motors for each
共kinematic兲 degree-of-freedom, but unlike the agonist/antagonist
approaches, the design of Hurst et al. is essentially a series elastic
actuator, where the stiffness of the series elastic element is modu-
lated by the second motor 共i.e., the second motor adjusts a spring
pretension, which modulates its stiffness兲. Tonietti et al. 关16兴 pre-
sented the design of a variable stiffness actuator that is structurally
different but conceptually similar to the design of Hurst et al.
Finally, in order to enhance the intrinsic safety of human/robot
interaction, Bicchi et al. 关17兴 and Tonietti and Bicchi 关18兴 pro-
posed the use of an agonist/antagonist pair of McKibben artificial
muscles to provide simultaneous control of position and 共open-
loop兲 stiffness in a similar manner to a biological motor control
system. In essence they control the position via the difference of
actuator pressures, while they control the stiffness with the sum.
This paper presents an alternate configuration for an actuator
with a physically variable output stiffness that offers a more com-
pact package and requires less mechanical complexity than the
aforementioned motor-spring designs. Unlike the work presented
in 关17,18兴, the proposed approach requires only a single actuator
rather than two. Further, as subsequently shown, the proposed
approach enables a greater dynamic range and bandwidth in the
control of stiffness and force, relative to that demonstrated in 关18兴.
Like the work presented in 关17,18兴, the proposed approach lever-
ages the open-loop behavior of a pneumatic actuator, which inher-
ently provides a series elastic component via the compressible gas
dynamics. A typical pneumatic actuator is controlled via a single
four-way spool valve, and as such, the actuator output stiffness is
not controllable independently of the actuator force. By decou-
pling the single four-way spool valve into a pair of three-way
valves, however, the pressure in each cylinder chamber can be
independently controlled, and thus the actuator output stiffness of
the cylinder actuator can be controlled independently of the output
force. Like the agonist/antagonist systems previously described,
the actuator force is a function of the difference between the
chamber pressures, while the actuator output stiffness is a function
of their sum. Thus, the only additional hardware required is an
extra valve 共for each actuator兲. As with the other 共previously pro-
posed兲 actuators, the stiffness is physical in nature 共i.e., open-
loop兲, and therefore maintains its constitutive behavior throughout
the frequency spectrum, with no potential for nonpassive behav-
ior. It should be noted that Raibert in 关19兴 suggests an actuator
configuration similar to the one proposed herein, but does not
describe the simultaneous force and stiffness control of this con-
figuration. Rather, as utilized by Raibert, the actuator alternates

Journal of Dynamic Systems, Measurement, and Control JULY 2007, Vol. 129 / 425
Copyright © 2007 by ASME

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 between a force source 共when the valves are open兲 and a spring
共when the valves are closed兲. This paper presents an approach for
the simultaneous stiffness and force control of the pneumatic ac-
tuator, and presents experimental results that characterize the
force and stiffness tracking performance.

2 Modeling the Pneumatic Actuator

In order to implement simultaneous force and stiffness control,
a two-input, two-output dynamic model of the actuator is briefly
described. For the proposed variable stiffness actuator, the two
model outputs are actuator force and output stiffness. The two
inputs to the actuator are the respective valve commands. As in a Fig. 1 Pneumatic actuator controlled by pair of three-way
typical pneumatic servo system, the commands are assumed to be valves
valve areas, which are algebraically related to the valve spool
displacement. Note that the servovalve spool dynamics are ne-
glected, since these are typically significantly faster than the ac-
tuator dynamics. Finally, since mass flow rate through the valve is
an algebraic function of the valve area, model formulation is sim- Ḟ = ṖaAa − ṖbAb 共4兲
plified by assuming the model inputs are the respective mass flow Assuming air is an ideal gas undergoing an isothermal process, the
rates into 共positive兲 or out of 共negative兲 the respective sides of the rate of change of the pressure inside each chamber of the cylinder
pneumatic cylinder. Specifically, modeling the flow through the can be expressed as a function of mass flow rate as
valve as the flow of an ideal gas through a converging nozzle, the
mass flow is algebraically related to the valve area command by RT P共a,b兲
the following relation: Ṗ共a,b兲 = ṁ共a,b兲 − V̇共a,b兲 共5兲
V共a,b兲 V共a,b兲
ṁ共Pu, Pd兲 = Av⌿共Pu, Pd兲 共1兲 where P共a,b兲 is the absolute pressure inside each side of the cylin-
where der, ṁ共a,b兲 is the mass flow rate command 共where as previously

冑 冉 冊

冦 冣
described, a positive command indicates mass flowing into the
␥ 2 共␥+1兲/共␥−1兲 chamber, negative indicates mass flowing out兲, and V共a,b兲 is the
C f Pu
RT ␥ + 1 volume of each cylinder chamber. Thus, the dynamics from mass
Pd flow input to the force output is given by
if ⱕ Cr 共choked兲
⌿共Pu, Pd兲 = Pu RTAa RTAb P aA a P bA b

冑 2␥
RT共␥ − 1兲
冑 冉 冊 冉 冊
1−
Pd
Pu
共␥−1兲/␥
Pd
Pu
共1/␥兲
C f Pu
Ḟ =
Va
ṁa −
Vb
ṁb −
Va
V̇a +
Vb
The volume in each chamber is a geometric function of piston
V̇b 共6兲

otherwise 共unchoked兲 displacement x, given by

and Av is the valve area command, Pu and Pd are the upstream and
共2兲
V共a,b兲 = A共a,b兲 冉 冊
L
2
±x 共7兲
downstream pressures, respectively, ␥ is the ratio of specific heats,
where L is the length of the actuator and 兩x 兩 ⬍ L / 2 共i.e., the cham-
R is the universal gas constant, T is the gas temperature at the
ber volume is never zero兲. Thus, the dynamics from mass flow
orifice, C f is the discharge coefficient of the valve 共which ac-
input to force output can be written as a function of the measur-
counts for irreversible flow conditions兲, and Cr is the pressure
able states 共i.e., the chamber pressures and piston displacement兲 as
ratio that divides the flow regimes into unchoked 共subsonic兲 and
choked 共sonic兲 flow through the orifice. Thus, by commanding the RT RT PaAaẋ PbAbẋ
valve orifice areas, the servovalves are algebraically commanding Ḟ = ṁa − ṁb − − 共8兲
the mass flow rates into or out of each chamber of the cylinder, L L L L
+x −x +x −x
and as such, the mass flow rates through the respective valves are 2 2 2 2
assumed to be the two actuator inputs. Note that, since the two The output stiffness of the actuator is defined by
valves are three-way valves, a positive valve command connects
the pressure supply to the cylinder so that mass flows into the ⳵F
chamber 共defined as positive mass flow兲, while a negative com- K=− 共9兲
⳵x
mand connects the cylinder to exhaust, such that mass flows out of
the chamber 共defined as negative mass flow兲. The model is thus where the actuator force is given by 共3兲. Substituting 共3兲 into 共9兲
derived by describing the respective relationships between actua- yields
tor force and chamber mass flow, and between actuator output
stiffness and chamber mass flow. Based on the schematic of the ⳵ 共PaAa − PbAb − PatmAr兲 ⳵ Pa ⳵ Pb
K=− = − Aa + Ab 共10兲
pneumatic actuator shown in Fig. 1, the force generated by the ⳵x ⳵x ⳵x
actuator is given by
The chamber pressures can be described as a function of displace-
F = PaAa − PbAb − PatmAr 共3兲 ment by assuming air is an ideal gas, such that

where Pa and Pb are the absolute pressures inside each chamber m共a,b兲RT
of the actuator, Aa and Ab are the effective areas of each side of P共a,b兲 = 共11兲
V共a,b兲
the piston, and Ar is the cross-sectional area of the piston rod. In
order to recover the mass flow rate inputs, the force equation is Substituting 共7兲 into 共11兲 yields an expression for pressures as a
differentiated with respect to time function of piston displacement:

426 / Vol. 129, JULY 2007 Transactions of the ASME

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 m共a,b兲RT RT

冉 冊 冉 冊
P共a,b兲 = 共12兲 b21 = 2 共23兲
L L
A共a,b兲 ±x +x
2 2
Differentiating with respect to x gives RT

冉 冊
b22 = 2 共24兲
⳵ P共a,b兲 m共a,b兲RT L

冉 冊
= ⫿ 共13兲 −x
⳵x L 2 2
A共a,b兲 ±x
2
3 Sliding Mode Controller Design
Thus, the stiffness can be written as a function of the respective
masses of air in each cylinder chamber and the piston displace- Given the two-input, two-output model provided by 共16兲 and
ment as 共24兲, a standard multi-input-multi-output sliding mode approach
can be utilized for the simultaneous control of actuator force and

冢冉 冊 冉 冊冣
ma mb stiffness. Note that sliding mode control was chosen to accommo-
K = RT 2 + 2 共14兲 date the model uncertainty in the nonlinear dynamics 共e.g., the
L L
+x −x actual pressure dynamics will vary somewhat from the isothermal
2 2 condition assumed in the model derivation兲. The bound for the
In order to recover the mass flow rate command inputs, 共14兲 is uncertainty associated with the above model can be expressed by
differentiated with respect to time, which after simplification
兩f̂ i − f i兩 ⱕ Fi i = 1, 2 共25兲
based on substitution from Eq. 共12兲 yields the dynamics from
mass flow input to the stiffness output as a function of measured
states, b = 共I + ⌬兲b̂ 兩⌬ij兩 ⱕ Dij i = 1, 2 j = 1, 2 共26兲

RT RT 2PaAaẋ 2PbAbẋ where f̂ i is the estimated value of f i, b̂ is the estimated matrix of

冉 冊 冉 冊 冉 冊 冉 冊
K̇ = 2 ṁa + 2 ṁb − 2 + 2 b, I is the 2 ⫻ 2 identity matrix, and Dij is a positive constant.
L L L L In order to develop an MIMO sliding mode controller for the
+x −x +x −x
2 2 2 2 system, first define a sliding surface vector s as

Note that the force dynamics 共8兲 are influenced by the difference
共15兲
s= 冋册冋 册
s1
s2
=
F − Fd
K − Kd
共27兲
in valve commands 共i.e., mass flow rates兲, while the stiffness dy-
where Fd and Kd are the desired actuator force and stiffness, re-
namics 共15兲 are influenced by the sum of the mass flow rates.
spectively. The corresponding sliding conditions for the two states
The model as described by 共1兲–共15兲 can be expressed in control
are
canonical form by defining the input vector such that it consists of
the two mass flow rates, u = 关ṁa ṁb 兴T, and by defining the output 1d 2
vector such that it consists of the force and stiffness, respectively, s ⱕ − ␩i兩si兩 i = 1, 2 共28兲
2 dt i
x = 关F K 兴T. As such, the system dynamics can be expressed as
where ␩i is a strictly positive constant. In order to satisfy the
ẋ = f + bu 共16兲 sliding conditions in the presence of the model uncertainty, the
MIMO sliding mode control law 共as described in 关20兴兲 is given by
where

冋册
u = b−1关ẋd − f − ␬ sgn共s兲兴 共29兲
f1
f= 共17兲 where ␬ sgn共s兲 is the robustness vector 关k1 sgn共s1兲 k2 sgn共s2兲兴T,
f2 and ẋd is the time derivative of the desired input vector, which is
and defined as xd = 关Fd Kd 兴T . k1 and k2 are chosen such that

冋 册
2

兺 D k = F + 兺 D 兩ẋ
b11 b12
b= 共18兲 共1 − Dii兲ki + ij j i ij di − f̂ j兩 + ␩i i = 1, 2
b21 b22 j⫽i j=1

where 共30兲

PaAaẋ PbAbẋ The existence of a solution to these equations is guaranteed by the

f1 = − − 共19兲 Frobenius-Perron theorem, as described in 关20兴.
L L

再
+x −x Finally, the commands to the control valves are calculated by
2 2
ṁa/⌿共Ps, Pa兲 if ṁa ⱖ 0
Av,a = 共31兲
2PaAaẋ 2PbAbẋ ṁa/⌿共Pa, Patm兲 otherwise

冉 冊 冉 冊 再
f2 = − 2 + 2 共20兲
L L ṁb/⌿共Ps, Pb兲 if ṁb ⱖ 0
+x −x
2 2 Av,b = 共32兲
ṁb/⌿共Pb, Patm兲 otherwise
RT where ⌿共Pu , Pd兲 is defined by 共2兲.
b11 = 共21兲
L
+x 4 Force Control With Maximum/Minimum Output
2 Stiffness
RT Many cases exist for the simultaneous control of force and stiff-
b12 = − 共22兲 ness for which one might desire either a maximum or a minimum
L actuator output stiffness. For example, during a kinematically con-
−x
2 strained manipulation task 共such as opening a drawer兲, a manipu-

Journal of Dynamic Systems, Measurement, and Control JULY 2007, Vol. 129 / 427

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 冦
lator should ideally maintain a low output stiffness, which mini- Fd + PatmAa PatmAb
mizes the effective gain of the disturbance transfer function from +
L L if Fd ⱖ 0
position error normal to the kinematic constraint to the resulting +x −x
force of interaction. Alternatively, for a manipulation task such as 2 2
Kmin = 共37兲
carving with a knife, the manipulator should ideally maintain a PatmAa − Fd + PatmAb
high output stiffness, which will minimize the effective gain from +
L L otherwise
environmental disturbance force to a resulting error in desired +x −x
trajectory. As described by Eqs. 共16兲–共24兲, the actuator output 2 2
stiffness and force are coupled, and thus the achievable stiffness at As such, the cases of minimum and maximum actuator stiffness
any given time depends on the actuator force and the actuator could be implemented by utilizing either Eq. 共36兲 or 共37兲, respec-
state. This coupling is expressed more explicitly by substituting tively, to determine the desired stiffness command to the previ-
Eq. 共12兲 into Eq. 共14兲, so that the stiffness can be expressed in ously described MIMO controller for a given desired force. Such
terms of pressure and displacement as follows: an approach, however, operates outside of the control loop, and
thus does not account for disturbances or tracking error 共i.e., does
P aA a P bA b not take into consideration the actual force, or more accurately,
K= + 共33兲 the actual cylinder pressures兲.
L L
+x −x In order to implement a more robust approach to obtaining a
2 2 maximum/minimum actuator stiffness 共i.e., one that is sensitive to
disturbances and tracking error兲, the authors developed a control
For a non-negative desired force Fd, the set of Eqs. 共3兲 and 共33兲
approach that operates inside the control loop on the control com-
can be solved explicitly for K and Pa, such that the expression for
mands 共i.e., on the mass flow rates兲, rather than the aforemen-
stiffness becomes
tioned command generator described by Eqs. 共36兲 and 共37兲.
Rather than specify a given stiffness at any given time, the pro-
Fd + PbAb + PatmAr PbAb posed approach leverages the presence of saturation in stiffness
K= + 共34兲
L L 共i.e., the presence of Eqs. 共36兲 and 共37兲兲 and the fact that stiffness
+x −x is a strictly positive quantity. Given these, the stiffness can be
2 2
made and maintained at its maximum 共in general fluctuating兲
Normally Pb is a positive value between atmospheric and supply value by maximizing its time rate of change 共i.e., driving it as
pressure, and thus the minimum stiffness can be written as a func- quickly as possible to its maximum value兲. Similarly, the stiffness
tion of desired force and displacement as can be made and maintained at its minimum 共in general fluctuat-
ing兲 value by minimizing its time rate of change 共i.e., driving it as
Fd + PatmAa PatmAb
Kmin = + if Fd ⱖ 0 共35兲
L L
+x −x
2 2

since Aa = Ab + Ar. Further analysis of other cases gives the follow-

ing expressions for the maximum and minimum stiffness:

冦
PsAa − Fd + PsAa − PatmAr
+
L L if Fd ⱖ 共Ps − Patm兲Ar
+x −x
2 2
Kmax =
Fd + PsAb + PatmAr PsAb
+
L L otherwise
+x −x
2 2
共36兲

Fig. 2 Experimental setup for the second set of experiments,

in which force and stiffness control were performed in the pres-
ence of motion control. The experimental setup for the first set
of experiments was essentially identical, except the inertial Fig. 3 Simultaneous force and stiffness tracking, each of
load was clamped at x = 0. 4.0 Hz sinusoid, with piston rod constrained at x = 0

428 / Vol. 129, JULY 2007 Transactions of the ASME

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Table 1 Model and controller parameters for experimental implementation

Parameter Description Value Unit

Ps Supply pressure 653 kPa

Aa Piston area side a 572 mm2
Ab Piston area side b 540 mm2
Ar Rod area 32 mm2
Av max Maximum valve area 12.57 mm2
Cf Discharge coefficient 0.8
Cr Pressure ratio 0.5
␥ Ratio of specific heats 1.4
R Universal gas constant 0.287 kJ/ kg K
T Gas temperature at the orifice 293 K
k1 Force error robustness gain 6 ⫻ 103 kg m / s3
k2 Stiffness error robustness gain 1.5⫻ 105 kg/ s3
Kp Outer loop proportional gain 1000 N/m
Kd Outer loop derivative gain 185 N s/m

冦
quickly as possible to its minimum value兲. As described by Eq. ṁa,max
共15兲, the time rate of change of stiffness is linearly influenced by
the 共input兲 mass flow rates, and as such the maximum or mini- − f 1 + Ḟd − k1 sgn共s1兲 − b11ṁa,max
if ⱕ ṁb,max
mum time derivative of stiffness can be obtained via the linear b12
optimization of ṁa =
− f 1 + Ḟd − k1 sgn共s1兲 − b12ṁb,max
J = b21ṁa + b22ṁb 共38兲 b11
where b21 and b22 are defined by 共23兲 and 共24兲. For stable force otherwise
tracking, the mass flow rates should satisfy the sliding condition 共45兲
共28兲 for i = 1,
and

冦
1d 2
s ⱕ − ␩1兩s1兩 共39兲 − f 1 + Ḟd − k1 sgn共s1兲 − b11ṁa,max
2 dt 1
b12
This condition can be satisfied by the following equation between
the inputs: − f 1 + Ḟd − k1 sgn共s1兲 − b11ṁa,max
ṁb = if ⱕ ṁb,max
b12
b11ṁa + b12ṁb = − f 1 + Ḟd − k1 sgn共s1兲 共40兲 ṁb,max
where b11 and b12 are given by 共21兲 and 共22兲, f 1 is given by 共19兲, otherwise
Ḟd is the time derivative of the desired actuator force, and k1 is the 共46兲
robustness gain satisfying 共39兲. Moreover, the inputs ṁa and ṁb
are bounded by an upper limit for which the corresponding valve The control inputs for minimizing stiffness are given by

冦
is fully open to the air supply and a lower limit for which the ṁa,min
corresponding valve is fully open to the exhaust. These constraints
can be expressed as − f 1 + Ḟd − k1 sgn共s1兲 − b11ṁa,min
if ⱖ ṁb,min
b12
ṁa,min ⱕ ṁa ⱕ ṁa,max 共41兲 ṁa = 共47兲
− f 1 + Ḟd − k1 sgn共s1兲 − b12ṁb,min
ṁb,min ⱕ ṁb ⱕ ṁb,max 共42兲 b11
where ṁ共a,b兲共min,max兲 are functions of the state as given by 共1兲, otherwise
evaluated at the maximum valve area. Specifically, ṁ共a,b兲共min,max兲 and

冦
are given by
− f 1 + Ḟd − k1 sgn共s1兲 − b11ṁa,min
ṁ共a,b兲max = Av,max⌿共Ps, P共a,b兲兲 共43兲 b12
and − f 1 + Ḟd − k1 sgn共s1兲 − b11ṁa,min
ṁb = if ⱖ ṁb,min 共48兲
b12
ṁ共a,b兲min = − Av,max⌿共P共a,b兲, Patm兲 共44兲
ṁb,min
where Ps is the supply pressure and Patm is the exhaust pressure. otherwise
Note that the negative sign for ṁ共a,b兲 min pertains to the direction
The corresponding commands to the control valves are calculated
of mass flow, as previously discussed. Thus, the whole problem
by 共31兲 and 共32兲.
can be defined as maximizing or minimizing the objective func-
tion J as defined by 共38兲 under the constraints 共40兲–共42兲.
The optimization problem defined above can be solved with a 5 Experimental Results
linear programming approach. The control inputs for maximizing The force/stiffness controller was implemented on an experi-
stiffness are given by mental setup to demonstrate and validate the approach. The ex-

Journal of Dynamic Systems, Measurement, and Control JULY 2007, Vol. 129 / 429

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 4 Force tracking of 0.25 Hz sinusoid and simultaneous
stiffness tracking of 4.0 Hz sinusoid with piston rod con-
strained at x = 0. The pressure trajectories „in chambers a and
b, respectively… that generate the force and stiffness tracking
are shown below the force and stiffness plots.
perimental setup, which is shown schematically in Fig. 1 and in a
1
photograph in Fig. 2, consists of a 2.7 cm 共1 16 in.兲 inner diam-
eter, 10 cm 共4 in.兲 stroke pneumatic cylinder 共Numatics 1062D04-
04A兲 controlled by a pair of proportional servovalves 共FESTO
MPYE-5-M5-010-B兲, each configured as a three-way valve as
shown in Fig. 1, and each supplied with air at an absolute pressure
of 653 kPa 共95 psia兲. The setup is instrumented with pressure
sensors 共FESTO SDE-16-10V/20mA兲, which measure the pres-
sure in each cylinder chamber. Two sets of experiments were per-
formed. In the first set, the piston rod was clamped to the table
such that the piston displacement was constrained to remain at x
= 0, and the desired force and stiffness were commanded directly
430 / Vol. 129, JULY 2007
Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
Fig. 5 Force tracking of 4.0 Hz sinusoid and simultaneous
stiffness tracking of 0.25 Hz sinusoid with piston rod con-
strained at x = 0. The pressure trajectories „in chambers a and
b, respectively… that generate the force and stiffness tracking
are shown below the force and stiffness plots.
to the MIMO controller. In the second set of experiments 共the
setup for which is depicted in Fig. 2兲, rather than command the
force directly, the force command was derived indirectly from a
motion control outer loop, while the stiffness was commanded
directly. Although this approach is arguably a less direct assess-
ment of force control relative to clamping the piston and provid-
ing a desired force trajectory, it demonstrates force control in the
presence of motion, and thus involves all terms in the force dy-
namics described in 共8兲. In these experiments, the output of the
pneumatic cylinder was connected to a linear slide, upon which a
10 kg mass was mounted, while the displacement of the slide 共and
actuator兲 was measured with a linear potentiometer 共Midori LP-
Transactions of the ASME
 Fig. 6 Force and stiffness tracking with force commands generated by the position control loop. The plots
show actual and desired position tracking of a 0.5 Hz sinusoidal motion and corresponding error, force
tracking of the force command resulting from the position control loop and corresponding error, and si-
multaneous stiffness tracking for commanded sinusoidal stiffness variation of 0.5 Hz, with corresponding
stiffness error.

100F兲, as depicted in Fig. 2. The model and controller parameters
used for the force/stiffness controller are listed in Table 1. The PD
control gains used in the position control loop in the motion con-
trol set of experiments are also listed in the table. Note that the
model parameters 共i.e., the first ten entries in the table兲 were either
measured, calculated, or found via standard references, whereas
the controller gains 共i.e., the last four entries兲 were obtained by
experimental tuning to balance tracking performance and sensor
noise 共i.e., higher gains provide better performance and better
robustness to model uncertainty, but introduce more sensor noise
into the output vector兲. Finally, it should be noted that actuator
output stiffness, which cannot be measured directly, was com-
puted based on 共33兲 and the combination of measured chamber
pressures and piston displacement.
Representative results of the first set of experiments 共desired
force and stiffness commanded directly and the piston displace-
ment clamped兲 are shown in Figs. 3–5. Figure 3 shows simulta-
neous force and stiffness tracking, where the force command is a
4.0 Hz sinusoid with 50 N amplitude, and the stiffness command
is also a 4.0 Hz sinusoid, varying between 7.0 and 9.0 N / mm.
Note that, despite unmatched initial conditions, the trajectories

Journal of Dynamic Systems, Measurement, and Control JULY 2007, Vol. 129 / 431

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 7 Force and stiffness tracking with force commands generated by the position control loop. The plots
show actual and desired position tracking of a 2.0 Hz sinusoidal motion and corresponding error, force
tracking of the force command resulting from the position control loop and corresponding error, and si-
multaneous stiffness tracking for commanded sinusoidal stiffness variation of 4.0 Hz, with corresponding
stiffness error.

converge fairly quickly to the desired trajectories. Figures 4 and 5
demonstrate the ability of the system to independently track force
and stiffness. Specifically, Fig. 4 shows 0.25 Hz force tracking
and simultaneous 4.0 Hz stiffness tracking 共with the same ampli-
tude variations in Fig. 3兲, while Fig. 5 shows the reverse. In ad-
dition to the force and stiffness trajectories, the pressure variations
in each cylinder chamber are also shown in Figs. 4 and 5. As is
clear in the pressure plots, the force is a function of the pressure
difference, while the stiffness is a function of the pressure sum.
Representative results of the second set of experiments 共desired
force commanded inside of a PD motion control loop, desired
stiffness commanded directly兲 are shown in Figs. 6 and 7. Figure
6 shows the force and stiffness tracking when the force commands
are generated by a PD motion controller tracking a 20 mm ampli-
tude, 0.5 Hz sinusoidal motion trajectory, and when the stiffness
commands are generated directly as a 0.5 Hz sinusoidal variation
in stiffness between 7.0 and 9.0 N / mm. Specifically, Fig. 6 shows
the desired and actual motion, along with the motion tracking
error; the desired and actual force, along with the force tracking
error; and the desired and actual stiffness, along with the stiffness

432 / Vol. 129, JULY 2007 Transactions of the ASME

Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 8 Force tracking with commands generated by the posi-
tion control loop, while maximizing the output stiffness of the
actuator. The plots show actual and desired position tracking
of a 0.5 Hz sinusoidal motion, force tracking corresponding to
the force command resulting from the position control loop,
and the corresponding stiffness of the actuator.
tracking error. Note that the force command is closely representa-
tive of the position tracking error. Both the nonsinusoidal shape of
the position error 共and force command兲 and the fact that the maxi-
mum errors occur at motion reversals indicate the clear presence
of Coulomb friction in the system, presumably in the piston seals.
Note that the maximum force tracking error is approximately 5%,
while the maximum stiffness tracking error is approximately 2%
共relative to the amplitude variation兲 with the major difference be-
ing the significantly higher frequency content of the force trajec-
tory, due to the aforementioned Coulomb friction.
Figure 7 shows the force and stiffness tracking when the force
commands are generated by the same PD motion controller track-
Journal of Dynamic Systems, Measurement, and Control
Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
Fig. 9 Force tracking with commands generated by the posi-
tion control loop, while minimizing the output stiffness of the
actuator. The plots show actual and desired position tracking
of a 0.5 Hz sinusoidal motion, force tracking corresponding to
the force command resulting from the position control loop,
and the corresponding stiffness of the actuator.
ing a 2.0 Hz sinusoidal motion trajectory, and when the stiffness
commands are generated directly as a 4.0 Hz sinusoidal variation
in stiffness 共with the same amplitude variations represented in Fig.
6兲. Like Fig. 6, Fig. 7 shows the desired and actual motion, along
with the motion tracking error; the desired and actual force, along
with the force tracking error; and the desired and actual stiffness,
along with the stiffness tracking error. Note that the position error
and corresponding force command are more sinusoidal in this
case, since the dynamics become more inertially dominated 共and
hence proportionally less influenced by friction兲 due to the higher
JULY 2007, Vol. 129 / 433
 accelerations. Note that both the force and stiffness maximum
tracking errors are approximately 10%, relative to the amplitude
of variation.
Figures 8 and 9 show the force tracking corresponding to a
0.5 Hz sinusoidal command in motion while maximizing stiffness
and minimizing stiffness, respectively. Note that the maximum
and minimum stiffnesses are not constant, but rather are functions
of the cylinder configuration as expressed in 共14兲. Note also the
“frequency doubling” effect in the stiffness variation, due to the
squared terms in 共14兲. The variation in stiffness between the two
conditions 共maximum relative to minimum兲 is approximately a
factor of 6, and as such provides a significant dynamic range in
achievable stiffness. The dynamic range of the stiffness could be
increased further by using a higher supply pressure 共relative to the
80 psig used in these experiments兲.
6 Conclusion
This paper proposes the use of a pneumatic actuator as a vari-
able stiffness actuator. These actuators are well suited to such a
task, since they are inherently series elastic actuators, and through
the use of two three-way servovalves rather than a single four-way
valve, the output stiffness can be controlled simultaneously with
the actuation force. The net result is a robot actuator capable of
modulated output stiffness with a minimal amount of associated
hardware. This paper describes an approach to simultaneously
control actuator force and stiffness, and presented experimental
results that demonstrate the effectiveness of the proposed ap-
proach. Among the performance attributes are a stiffness tracking
bandwidth equivalent to the force tracking bandwidth, and an ap-
proximate factor of 6 dynamic range in stiffness 共which could be
increased with an increased supply pressure兲.
References
关1兴 Hogan, N., 1985, “The Mechanics of Multi-Joint Posture and Movement Con-
trol,” Biol. Cybern., 52, pp. 315–331.
关2兴 Hogan, N., 1984, “Adaptive Control of Mechanical Impedance by Coactiva-
tion of Antagonist Muscles,” IEEE Trans. Autom. Control, 29共8兲, pp. 681–
690.
关3兴 Hogan, N., 1980, “Tuning Muscle Stiffness can Simplify Control of Natural
Movement,” in Advances in Bioengineering, ASME Winter Annual Meeting,
pp. 279–282.
434 / Vol. 129, JULY 2007
Downloaded 14 Mar 2008 to 129.59.78.62. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
关4兴 Mussa-Ivaldi, F. A., Hogan, N., and Bizzi, E., 1985, “Neural, Mechanical, and
Geometric Factors Subserving Arm Posture in Humans,” J. Neurosci., 5, pp.
2732–2743.
关5兴 Alexander, R., 1990, “Three Uses for Springs in Legged Locomotion,” Int. J.
Robot. Res., 9共2兲, pp. 53–61.
关6兴 Cavagna, G. A., Heglund, N. C., and Taylor, C. R., 1977, “Mechanical Work in
Terrestrial Locomotion: Two Basic Mechanisms for Minimizing Energy Ex-
penditure,” Am. J. Physiol., 223, pp. 243–261.
关7兴 Taylor, C. R., and Heglund, N. C., 1982, “Energetics and Mechanics of Ter-
restrial Locomotion,” Annu. Rev. Physiol., 44, pp. 97–107.
关8兴 Salisbury, J. K., 1980, “Active Stiffness Control of a Manipulator in Cartesian
Coordinates,” in Proceedings of the IEEE Conference on Decision and Con-
trol, pp. 383–388.
关9兴 Hogan, N., 1985, “Impedance Control: An Approach to Manipulation,” ASME
J. Dyn. Syst., Meas., Control, 107共1兲, pp. 1–24.
关10兴 Laurin-Kovitz, K. F., Colgate, J. E., and Carnes, S. D. R., 1991, “Design of
Components for Programmable Passive Impedance,” in Proceedings of the
IEEE International Conference on Robotics and Automation, Vol. 2, pp. 1476–
1481.
关11兴 Koganezawa, K., and Yamazaki, M., 1999, “Mechanical Stiffness Control of
Tendon-Driven Joints,” in Proceedings of the IEEE/RSJ International Confer-
ence on Intelligent Robots and Systems, pp. 818–825.
关12兴 Koganezawa, K., and Ban, S., 2002, “Stiffness Control of Antagonistically
Driven Redundant DOF Manipulator,” in Proceedings of the IEEE/RSJ Inter-
national Conference on Intelligent Robots and Systems, pp. 2280–2285.
关13兴 English, C. E., and Russell, D., 1999, “Mechanics and Stiffness Limitations of
a Variable Stiffness Actuator for Use in Prosthetic Limbs,” Mech. Mach.
Theory, 34共1兲, pp. 7–25.
关14兴 English, C. E., and Russell, D., 1999, “Implementation of Variable Joint Stiff-
ness Through Antagonistic Actuation Using Rolamite Springs,” Mech. Mach.
Theory, 34共1兲, pp. 27–40.
关15兴 Hurst, J. W., Chestnutt, J. E., and Rizzi, A. A., 2004, “An Actuator with
Physically Variable Stiffness for Highly Dynamic Legged Locomotion,” in
Proceedings of the IEEE International Conference on Robotics and Automa-
tion, pp. 4662–4667.
关16兴 Tonietti, G., Schiavi, R., and Bicchi, A., 2005, “Design and Control of a
Variable Stiffness Actuator for Safe and Fast Physical Human/Robot Interac-
tion,” in Proceedings of the IEEE International Conference on Robotics and
Automation, pp. 528–533.
关17兴 Bicchi, A., Rizzini, S. L., and Tonietti, G., 2001, “Compliant Design for In-
trinsic Safety: General Issues and Preliminary Design,” in Proceedings of the
IEEE/RSJ International Conference on Intelligent Robots and Systems, pp.
1864–1869.
关18兴 Tonietti, G., and Bicchi, A., 2002, “Adaptive Simultaneous Position and Stiff-
ness Control for a Soft Robot Arm,” in Proceedings of the IEEE/RSJ Interna-
tional Conference on Intelligent Robots and Systems, pp. 1992–1997.
关19兴 Raibert, M. H., 1986, Legged Robots that Balance, MIT Press, Cambridge,
MA, pp. 33–34.
关20兴 Slotine, J. E., and Li, W., 1991, Applied Nonlinear Control, Prentice-Hall, NJ.
Transactions of the ASME