Accepted Manuscript

Title: Model-based feedforward and cascade control of

hydraulic McKibben muscles

Author: Michael Meller Boris Kogan Matthew Bryant

Ephrahim Garcia

PII: S0924-4247(17)31199-8
DOI: https://doi.org/doi:10.1016/j.sna.2018.03.036
Reference: SNA 10702

To appear in: Sensors and Actuators A

Received date: 30-6-2017

Revised date: 1-3-2018
Accepted date: 23-3-2018

Please cite this article as: Michael Meller, Boris Kogan, Matthew Bryant,
Ephrahim Garcia, Model-based feedforward and cascade control of hydraulic
McKibben muscles, <![CDATA[Sensors & Actuators: A. Physical]]> (2018),
https://doi.org/10.1016/j.sna.2018.03.036

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 Model-based feedforward and cascade control of
hydraulic McKibben muscles

t
Michael Mellera,∗, Boris Kogana , Matthew Bryantb , Ephrahim Garciaa,1

ip
a Laboratory for Intelligent Machine Systems, Sibley School of Mechanical and Aerospace
Engineering, Cornell University, Ithaca, NY 14853-7501, USA

cr
b Intelligent Systems and Structures Research Lab, Department of Mechanical and Aerospace

Engineering, North Carolina State University, Raleigh, NC 27695-7910, USA

us
Abstract

an
McKibben artificial muscle actuators are predominantly pneumatically pow-
ered. Recently, hydraulic operation has gained interest due to its higher rigidity
and efficiency. While there has been extensive control system development for
M
pneumatic artificial muscles, little has been conducted hydraulically. This pa-
per investigates three different controllers developed for a loaded robotic arm
actuated with oil-hydraulic McKibben muscles. The goal was to achieve good
d

angular position tracking over a range of frequencies up to 1 Hz. The first

te

scheme, serving as the baseline, is a proportional-integral controller. The sec-

ond architecture adds a nonlinear model-based feedforward term to the baseline
controller; the feedforward includes the expected flowrate demands based on
p

the actuator kinematics as well as the valve flow gain. The last scheme adds
ce

an inner pressure feedback loop to the second architecture. All controllers were
evaluated with frequency and step response experiments. The results show that
a simple proportional-integral controller has significant phase lag and attenu-
Ac

ation at the higher frequencies tested; including the feedforward term almost
completely eliminates these. The cascaded loop improves rise and settling times.
Keywords: McKibben muscle, fluidic artificial muscle, hydraulic artificial
muscle, fluid power control, hydraulic control, robotics

∗ Corresponding author
Email address: mam627@cornell.edu (Michael Meller)
1 Deceased.

Preprint submitted to Sensors and Actuators A: Physical February 28, 2018

Page 1 of 35
 1. Introduction

t
ip
McKibben muscles are most commonly used in anthropomorphic robots [1,
2, 3, 4] and assistive devices [5, 6, 7]. Their high force and power densities, low

cr
fabrication cost, skeletal muscle-like performance, and scalability make them
5 appealing for use in such systems [8]. In their original form, McKibben muscles
are made up of an elastomeric tube that is inserted into a helically woven mesh

us
[9]. The mesh is typically braided with a bias angle (with respect to the central
axis) less than the neutral angle of 54.7◦ , thereby allowing it to expand radially

an
and contract axially [8]; while it is not nearly as common, extensile operation
10 is also possible if the bias angle of the weave is greater than the neutral angle
[10]. One end of the actuator is plugged while the other end connects to the
M
fluid power supply. Pressurization of the McKibben muscle via the fluid port
causes contraction and/or tension depending on the load. Since these actuators
are single-acting, either an antagonistic pair or a restoring force is required to
d

15 achieve bi-directional motion.

te

Compressed air, rather than pressurized liquid, is by far the most popu-
lar means to power McKibben muscles, hence the moniker pneumatic artificial
p

muscle (PAM) [8]. This fluid medium is often chosen to take advantage of its
inherent compliance, making these systems more suitable for human-robot in-
ce

20 teraction, in addition to the benefits of cleanliness, availability of components

at the human scale, and ability to use the atmosphere as the return reservoir
[11, 7, 12, 13]. However, the compliance of pneumatic systems makes accu-
Ac

rate position control more challenging, and results in lower energetic efficiencies
[14, 13]. When considering the amount of energy required to bring the working
25 fluid from atmospheric pressure to a system pressure of 5.17 bar (75 psi), as
well as the losses associated with the actuator’s transduction from fluid energy
to mechanical work, it was shown that over one full actuation cycle a typical
McKibben muscle is 61% efficient when operated hydraulically, while only 27%
efficient when run pneumatically [15].

Page 2 of 35
30 Several research groups have favored hydraulic operation of McKibben mus-
cles for the higher possible efficiency and stiffness [14, 16, 17, 15, 18, 19].
When powering McKibben muscles hydraulically, some of the considerations

t
ip
that must be made when selecting the type of fluid include cost, availability,
system chemical compatibility, viscosity, bulk modulus, specific gravity, freez-

cr
35 ing/boiling points, and environmental impact [13, 12, 20]. The most common
type of hydraulic fluid is petroleum oil-based because they tend to be low cost,

us
have good lubricity, and are readily available [20, 12]. The main drawbacks
of these fluids include higher flammability, and that spills result in slip safety
hazards and are less environmentally friendly than other alternatives [20, 12].

an
40 Water-based hydraulic fluids are another relatively common option, and these
are generally less flammable, more environmentally friendly (depending on the
additives used), and have a higher bulk modulus and significantly lower viscos-
M
ity [13, 12, 20]. The lower viscosity reduces pressure losses around the system,
however, it is also why water-based hydraulic systems suffer poor lubrication
45 and accelerated component wear [12, 20]. Some additional challenges associ-
d

ated with water-based hydraulic systems include corrosion and the increased
te

risk of algae and bacteria growth [20, 14].

Other means of McKibben muscle actuation have been explored as well.
p

Takashima et al. impregnated the McKibben mesh with a shape memory poly-
50 mer so that the muscle can hold a fixed position without the need for continuous
ce

control [21]. Sangian et al. developed thermally activated, paraffin filled McK-
ibben muscles to eliminate the need for a bulky prime mover [22, 23]. Ball et
al. replaced the braided mesh of McKibben muscles with a tubular-knit sleeve
Ac

to attain more skeletal muscle-like force-stroke profiles with free contractions of

55 over 50% [24]. Recently, variable recruitment of parallel McKibben muscles has
gained popularity as a way to increase energy efficiency [25, 26, 27, 28].
There has been significant controller development for pneumatically ac-
tivated McKibben muscles spanning classical control to far more advanced
and nonlinear techniques, which will be discussed in more detail in Section
60 3 [6, 7, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38], however there has been very little

Page 3 of 35
 control system consideration for hydraulic artificial muscles (HAMs) [14, 19, 28].
This paper investigates the angular position control system development for a
robotic arm actuated with oil-hydraulic McKibben muscles lifting a fixed load.

t
ip
The main objective was to achieve good angular position tracking over a range
65 of frequencies up to 1 Hz, which is similar to what is experienced in anthro-

cr
pomorphic robotic and orthotic applications [48]. Simpler control architectures
were emphasized, therefore only augmentations to classical control techniques

us
were employed.
The primary contribution of this work is the development and successful
70 implementation of a model-based feedforward term to improve the position

an
tracking of both a proportional-integral (PI) controller as well as a cascaded
controller with proportional-integral control for the outer position loop and
proportional control for the inner pressure loop (PI-P). This feedforward in-
M
cludes the expected volumetric flowrate demands derived from the McKibben
75 muscle kinematics as well as the valve flow gain. The McKibben muscle pa-
rameters required for this feedforward compensation are identical to those used
d

in modeling its force output, and the flow gain is usually provided in a valve’s
te

datasheet, therefore only a small additional effort is required to realize a rea-

sonable first attempt at the feedforward term. If the best agreement between
p

80 the reference trajectory and feedback are desired, as was in this work, then
these parameters can be identified more accurately in hardware. All angular
ce

position controllers were experimentally assessed by performing frequency and

step response experiments. The results demonstrate that the developed feed-
forward compensation significantly improves position tracking by reducing the
Ac

85 phase lag of the proportional-integral controller, and the addition of an inner

pressure loop further reduces rise and settling times.

2. System definition

The hydraulic McKibben muscle controllers were developed for the robotic
arm testbed pictured in Fig. 1. The arm can lift a variety of loads, but 4.54 kg

Page 4 of 35
90 (10 lb) was used throughout the experiments presented in this paper. At the
system supply pressure of 7.6 bar (110 psi), a maximum load of 9.07 kg (20 lb)
can be lifted by the arm over its full range of motion, hence half of this maximal

t
ip
load was chosen to serve as a representative load use case. This 4.54 kg (10 lb)
load results in approximately a 578 N (130 lbf) static tension on the muscles

cr
95 due to the geometry of the arm.

us
an
M
d

Figure 1: The robot arm experimental test rig.

te

Six parallel McKibben muscles actuate the arm, and these are powered hy-
p

draulically for the efficiency and control benefits over pneumatic operation. The
McKibben muscles were fabricated using a silicone bladder with an inner diam-
ce

eter of 9.53 mm (3/8 in), an outer diameter of 12.7 mm (1/2 in), and a shore
100 hardness of 35 A. This bladder was chosen for its hydraulic oil compatibility, as
Ac

well as its relatively low shore value so as to minimize elastic energy storage.
A Kevlar braided mesh with an expansion range of 12.7 mm (1/2 in) to 22.2
mm (7/8 in) in diameter was used; Kevlar was chosen over the more common
polyethylene or nylon due to its more durable fibers. These components are
105 held together with Eaton Weatherhead 06Z-R06 steel o-ring boss swivel crimp
fittings. Each HAM has a 17.78 cm (7 in) active length, measured between the
inside edge of the end fittings. Once assembled, the effective initial mesh braid

Page 5 of 35
 angle of the six parallel McKibben muscles was identified as 29.7◦ .
A Haldex Barnes 4F686A hydraulic power unit (HPU), capable of producing
51.7 bar (750 psi) and 15.1 Lpm (4 gpm), supplies hydraulic power to the

t
110

ip
system. All six HAMs are fluidically connected to one Moog G761 four-way
electrohydraulic servovalve (EHSV) with a rated flow of 62.5 Lpm (16.5 gpm)

cr
at a supply-to-return pressure drop of 69 bar (1000 psi). The EHSV directs the
flow from the HPU to the HAMs during pressurization, and from the HAMs

us
115 to the reservoir during venting. A petroleum-based oil was chosen as the fluid
medium largely due to the fact that the majority of available hydraulic system
components are intended for use with this type of fluid, including our accessible

an
pump and servovalve. Using a petroleum-based oil did result in taking special
care when choosing the McKibben muscle bladder material since natural rubbers
120 that are often utilized would quickly degrade when exposed to such a fluid [15].
M
A Delta Computer Systems VC2124 two axis voltage-to-current converter
powered the EHSV. Since the HAMs are single-acting and used in a gravity-
return configuration, the servovalve B-port was plugged. The pressure of all six
d

HAMs is monitored with one Measurement Specialties MSP-300-250-P-4-N-3

te

125 17.24 bar (250 psi) pressure transducer. A 2500 count per revolution US Digi-
tal E6 kit optical encoder measured the angle of the arm. Data acquisition and
p

control commands were carried out on a Quanser QPIDe I/O board with a sam-
pling rate of 1 kHz, using the MATLAB/Simulink environments and Quanser
ce

QUARC real-time control software. A diagram of the robot arm, McKibben

130 muscles, hydraulic circuit, and sensing and control elements are shown in Fig.
2.
Ac

3. Controller development

Many research groups have investigated control system development for

pneumatic artificial muscles, however very few performed feedback control with
135 HAMs. Before beginning our controller development, a brief survey of existing
controllers for McKibben muscles and traditional hydraulics was carried out, and

Page 6 of 35
 t
ip
cr
us
Figure 2: Schematic of the whole system including the robot arm, McKibben muscles, hy-
draulic system, and sensing and control.

an
is presented in Table 1. Simpler PAM control system development involves clas-
sical control techniques such as PI and proportional-integral-derivative (PID)
M
control [37, 19], as well as augmenting these classical approaches by adding
140 dead time compensation for delays in valve operation and fluid transport lag
[38], cascaded loops for improving speed and disturbance rejection [32, 36, 26],
d

or feedforward terms to address nonlinearities such as hysteresis [39, 36].

Caldwell et al. noted that when utilizing a PID with feedforward controller
te

for their pneumatic artificial muscle driven system [39], the closed-loop perfor-
145 mance displayed high sensitivity to errors in their feedforward term, primarily
p

due to fluctuations in temperature and supply pressure, as well as pipe length

ce

and other actuator characteristics; these limitations drove them to pursue adap-
tive control [29]. The intrinsic compliance and sensitivity to changes in temper-
ature make accurate position control of pneumatic systems challenging [13, 12];
Ac

150 largely due to these reasons, coupled with the added nonlinearities of McKibben
muscles, many research groups have required more sophisticated control tech-
niques to improve positioning accuracy [29, 30, 14]. Some examples of these
approaches include sliding mode control [30], neural network-based nonlinear
PID control [46], and parameter self-adjust model-free adaptive control (PSA-
155 MFAC) [40].
On the other hand, the high bulk modulus of hydraulic systems results in

Page 7 of 35
 far more accurate positioning while employing simpler control algorithms when
compared to its pneumatic counterpart [13, 12]. Traditional hydraulic cylin-
der control usually stems from classical control schemes [41, 42, 43, 44], often

t
ip
160 augmented with velocity reference feedforward (ff), and sometimes acceleration
reference ff [42, 44]. While there are very few studies on hydraulic McKibben

cr
muscle control, most are classical control-based. Xiang et al. used PID to con-
trol their water and air valves supplying their combined hydraulic/pneumatic

us
McKibben muscle [19]. Focchi et al. focused on the performance comparison of
165 water vs. air McKibben muscles (not the control), however a control loop was
needed to perform their experiments; they implemented a cascaded architecture

an
with an outer position loop (gain-scheduled PI-controller) and inner pressure
loop (P-controller) [14]. Kobayashi et al. developed a model predictive con-
troller (MPC) with adaptive parameter estimation through a recursive least
M
170 squares (RLS) algorithm to better handle different load cases in their water-
hydraulic McKibben muscles [45]. Meller et al. investigated the energy savings
of variable recruitment hydraulic McKibben muscles; since position tracking
d

performance was not the center of this work, PI-control was employed for their
te

oil-filled HAMs, and approximately 60◦ of phase lag was present in their baseline
175 position tracking experiment [28].
p

While many of the aforementioned advanced pneumatic artificial muscle con-

trol schemes would presumably work well for the system under consideration in
ce

this paper, they are likely more complex than necessary because our working
fluid is oil. A main objective of this paper is to minimize the complexity of
the hydraulic artificial muscle controller, however using PI-control alone has in-
Ac

180

dicated that a more intricate approach is necessary to avoid significant phase

lag [28]. Traditional cylinder-driven hydraulic systems have exhibited accurate
position tracking with the use of classical control and feedforward terms [42, 44].
Our proposed methodology extends this feedforward approach by taking into
185 account the nonlinear flow behavior of the muscle. Focchi et al. attained good
trapezoidal position tracking at 0.5 Hz with water hydraulic McKibben muscles
controlled by a cascaded architecture [14]. Similar nested controllers are popular

Page 8 of 35
 with PAMs [32, 36, 26]; their success with this relatively simple approach rooted
in classical control also influenced our choice of control loops to investigate.

t
Table 1: Survey of existing controllers for McKibben muscles and traditional hydraulics.

ip
Author(s) Actuator Fluid Sensing Req’d Controller
This work McKibben Oil Angle, pressure Cascaded PI-P w/ ff
Meller et al. [28] McKibben Oil Angle PI

cr
Xiang et al. [19] McKibben Water/air Angle PID
Kobayashi et al. [45] McKibben Water Position MPC with RLS
Focchi et al. [14] McKibben Water Position, pressure Cascaded PI-P w/ gain sched.
Thanh et al. [46] McKibben Air Angle Nonlinear PID w/ neural net.
Yang et al. [40] McKibben Air Angle PSA-MFAC

us
Woods et al. [38] McKibben Air Angle P w/ dead time comp.
Tondu et al. [30] McKibben Air Angle Sliding mode
Robinson et al. [26] McKibben Air Angle, pressure Cascaded PI-PI
Vo Minh et al. [36] McKibben Air Position, pressure Cascaded PI-PI w/ hysteresis ff
Caldwell et al. [39] McKibben Air Angle PID w/ ff

an
Caldwell et al. [29] McKibben Air Angle Adaptive linear
Schröder et al. [32] McKibben Air Angle, torque Cascaded PI-PI
De Volder et al. [37] McKibben Air Position PI
Manring [43] Cylinder Oil Position PI
Vickers [41] Cylinder Oil Position PID
Nachtwey [44] Cylinder Oil Position PID w/ vel. (and accel.) ff
M
Johnson [42] Cylinder Oil Position PIV w/ vel. (and accel.) ff
Sohl et al. [47] Cylinder Oil Position, pressure Feedback linearization

190 We consider three different controllers for our hydraulic McKibben muscle-
d

actuated robotic arm. These angular position tracking control laws determine
te

the commanded voltage to the electrohydraulic servovalve throttling the fluid

power to and from the McKibben muscles. It was desired to attain acceptable
p

tracking over a range of frequencies up to 1 Hz since this is a common frequency

195 range in human motion [48], and therefore anthropomorphic robots and assistive
ce

devices. Another goal was to keep the control architectures reasonably simple
by only making improvements to classical control techniques, and leveraging
Ac

information McKibben muscle users should already have at hand. This way,
the improvements are more accessible and versatile, and can be applied to a
200 range of systems and applications.

3.1. PI-control

The first closed loop angular position controller tested, serving as the base-
line of comparison, is a classical proportional-integral controller. The error be-
tween the desired (θd ) and measured (θ) angles was sent into the PI-controller,

Page 9 of 35
205 which then output the commanded valve voltage, uv . A block diagram repre-
sentation of this is given in Fig. 3. This attempts to deliver the pressure and
flowrate of oil required to track the reference trajectory. PI-control was chosen

t
ip
as the baseline because it has shown to be the simplest useful control technique
spanning both PAMs and hydraulics [37, 28, 42, 43]. Additionally, including a

cr
210 derivative term typically adds little to no performance enhancement in electro-
hydraulic servo-systems [42]. The PI gains were chosen in hardware by following

us
the Tyreus-Luyben tuning rules, which are based upon the Ziegler-Nichols ulti-
mate gain method, but yield less overshoot [49]. To experimentally determine
the gains, the system is first set in proportional-control only mode and while

an
215 commanding new set-points, the proportional gain is increased until marginal
stability is reached [49]. The proportional gain at this condition is the ultimate
gain (Ku ), and the period of the sustained oscillations is the ultimate period
M
(τu ). These two quantities are then used to calculate the proportional and in-
Ku
tegral gains of the controller, where KP = 0.3125 × Ku and KI = 0.1420 × τu

220 respectively [49].

d
p te

Figure 3: Block diagram for PI-control.

ce

3.2. PI-control with model-based feedforward

The second control architecture developed adds a feedforward term to the

Ac

baseline PI-controller (the same PI gains are used). Feedforward can be utilized
to 1) minimize the effects of known disturbances (typically requires a sensor to
225 detect the disturbance and a model), or 2) it can provide anticipative actions in
tracking applications (typically requires the reference trajectory and a model)
[50]. The second definition of feedforward is utilized in this work.
In the traditional hydraulic control methods that use feedforward in Table
1, the most commonly implemented form consists of a gain multiplied with the

10

Page 10 of 35
230 time derivative of the position reference signal [42, 44]. This method is used
primarily for reducing following error, and is often presented as an additional
gain to tune where the user blindly increases the gain while performing ramp

t
ip
commands until this error is minimized. Fortunately this gain is a forgiving
parameter to tune and it does not affect loop stability [42].

cr
235 Upon further investigation, a simple feedforward gain multiplied with the
time derivative of the position reference signal likely works so well in hydraulic

us
cylinder control because in steady load conditions, the flowrate out of the valve
is effectively directly proportional to the commanded current (we command a
voltage that is linearly mapped to a current through a voltage-to-current con-

an
240 verter) [51, 41, 52], and the flowrate demands are directly proportional to the
velocity of the cylinder [41, 42]; hence all of this information can be lumped into
one gain. It was desired to use a similar feedforward methodology to provide an-
M
ticipative action thereby reducing the tracking error in this work. Kinematically
speaking, since McKibben muscles are highly nonlinear and their flowrate is de-
245 pendent on both position and velocity (derived in Eq. (3)), whereas cylinder
d

flowrate is only a function of velocity, we decoupled the typical feedforward gain

te

into a nonlinear predicted volumetric flowrate model of the McKibben muscles

and a linear flow gain of the EHSV.
p

The volumetric flowrate demands of the McKibben muscles were modeled

250 following common assumptions that the HAMs maintain a cylindrical shape
ce

throughout contraction, the Kevlar fibers are inextensible but flexible, and the
Kevlar mesh follows the pantograph opening principle [8]. While some take
into account the semi-conical end effects [53, 17, 54], it was experimentally
Ac

shown in previous work that for McKibben muscles with similar construction
255 and slenderness ratios as are used in this paper, the cylindrical model provides
an extremely accurate estimate of the liquid volume delivered to the muscle over
its full contraction range [15, 28]. For simplicity, we also assume that all six
parallel McKibben muscles are manufactured identically so the volume equation
is multiplied by a factor of six. Taking all this into account yields an equation for
260 the volume of the HAMs as a function of the current stroke and initial geometry

11

Page 11 of 35
 " #
3
(l0 − x) (l0 − x)
Vm = 6 · πr02 − , (1)

t
sin2 (α0 ) l02 tan2 (α0 )

ip
where r0 is the initial radius, l0 is the initial resting length, α0 is the initial

cr
angle the mesh fibers make with respect to the longitudinal axis of the muscle,
and the stroke is defined as

us
x = l0 − l = R (sin(θ) − sin(θ0 )) , (2)

an
where l is the current length of the HAMs, R is the distance between the at-
265 tachment point of the McKibbens and the robot arm “elbow” joint, θ is the
M
angle the robot arm makes with respect to the horizontal, and θ0 is the initial
arm angle when the HAMs are at their initial longest length. Taking the time
derivatives of Eqs. (1) and (2) and substituting yields the volumetric flowrate
d
model of the McKibben muscles
te

" #
2
3 (l0 − x) 1
Qm = 6 · πr02 ẋ − . (3)
l02 tan2 (α0 ) sin2 (α0 )
p

In developing a model of the EHSV, we first assume that the dynamics

ce

270

of the torque motor and flapper are fast enough to be negligible, hence the
commanded valve voltage is directly proportional to the spool position. Others
Ac

have made this assumption as well [43, 47]. This is a reasonable assumption
since the maximum closed-loop frequency tested is 1 Hz and previous testing
275 of this EHSV when piloted at 17.24 bar (250 psi) yielded a bandwidth of 10
Hz (normally 40 Hz when piloted at 207 bar (3000 psi) [52]). While the pilot
pressure is 17.24 bar (250 psi), the supply pressure to the main stage of the valve
(to the McKibben muscles) is 7.6 bar (110 psi); this allows for improved valve
performance while minimizing the risk of the McKibben muscles bursting. We

12

Page 12 of 35
280 also assume that the valve has a critically centered lap condition with square
port geometry (based on the valve datasheet [52], and validated in Fig. 7 (a) of
Section 4.2), and that the flow obeys the orifice equation, which gives

t
ip
√

cv uv PS − Pm for uv > 0

cr





Qv = 0 for uv = 0, (4)

−c u √P − P



us

v v m T for uv < 0

where cv is the valve orifice coefficient, PS is the pump supply pressure, Pm is

an
the pressure in the McKibben muscles, and PT is the tank pressure [51, 43, 47].
285 In order to estimate the flowrate through the valve at a given voltage ap-
plied, without the need for updating estimates or measurements of the system
M
pressures, a linearization about the null conditions was used; this is when the
valve spool is centered, there is no flow, the tank pressure is assumed to be zero
for simplicity, and the actuator pressure is assumed to be one half the supply
d

290 pressure. In taking the Taylor series expansion for the linearization, we must
te

take the partial derivatives of Eq. (4) with respect to the valve voltage and mus-
cle pressure. For critically centered valves, we know that the partial derivative
of Eq. (4) with respect to the muscle pressure evaluated at the null condition
p

is zero, hence we only need to consider the partial derivative of Eq. (4) with
ce

295 respect to the valve voltage, which is also known as the flow gain [51, 43]
Ac

 r
∂Qv  1
Kq0 = = cv PS . (5)
∂uv 0 2

The assumptions made about the actuator and tank pressures in this derivation
are an acceptable approximation of the system behavior, and they allow the
flow gain to reduce to one value for flow both into and out of the actuators [43].
Hence, the linearized flow equation is given by

13

Page 13 of 35
 Qv,lin ≈ Kq0 uv . (6)

t
ip
300 If we assume that the compressibility and pressure dynamics of the fluid
are negligible, then we can state that the linearized valve flow in Eq. (6) must

cr
equal the volumetric flowrate into the muscle given in Eq. (3). Rearranging
this equation for the commanded valve voltage yields the feedforward term in

us
our control law

an
uff = Qm (θd , θ̇d )/Kq0 . (7)

Assuming negligible pressure dynamics is reasonable since our maximum load

M
305

dynamics are on the order of one second while the pressure dynamics are typ-
ically on the order of milliseconds [43]. This assumption would also hold for a
d
water-based hydraulic system since the pressure dynamics are dependent on the
fluid’s bulk modulus, and water-based hydraulic fluids tend to have a slightly
te

310 greater bulk modulus than those that are petroleum-based [51]. Significantly
higher frequency operation or much larger loads on the system would likely re-
p

sult in the need to consider the effects of fluid compressibility, however [41].
ce

This feedforward term in Eq. (7) is shown in the block diagram representing
the overall combined feedforward/feedback control law in Fig. 4.
Ac

Figure 4: Block diagram for PI-control with model-based feedforward.

14

Page 14 of 35
315 3.3. Cascaded PI-P-control with model-based feedforward

Several groups have implemented a cascaded PI-PI architecture when con-

t
trolling PAMs [32, 36, 26]. Cascade control has the added benefit of improved

ip
disturbance rejection to the primary controlled variable; however, it requires
the addition of a sensor to measure another variable, and that the inner loop

cr
320 dynamics are much faster than the outer loop dynamics [55]. The last scheme
implemented in this paper adds an inner control loop with pressure feedback to

us
the second developed architecture from Fig. 4. The outer PI-control loop is for
angular positioning; instead of its output being the feedback controller portion
of the commanded valve voltage as in Section 3.2, its output is now the pressure

an
325 reference for the inner secondary loop. The inner proportional (P) control loop
tracks the desired pressure set by the outer loop. P-control was chosen over PI-
control to avoid potential competing dynamics of the cascaded loops, and the
M
integral action of the outer loop is sufficient for eliminating steady-state error of
the arm angle. The gain tuning process began with the inner loop as suggested
in Marlin’s book [55]. The proportional gain was chosen to be the inverse of
d
330

the valve’s pressure sensitivity. Then the inner loop was fixed and treated as if
te

it were part of the plant, and the outer loop was tuned as described in Section
3.1. The block diagram representing this control scheme is shown in Fig. 5.
p
ce
Ac

Figure 5: Block diagram for cascaded PI-P-control with model-based feedforward.

15

Page 15 of 35
 4. Parameter identification for feedforward compensation

335 4.1. Effective initial braid angle of the parallel McKibben muscles

t
In Section 3.2, the developed feedforward term is heavily dependent on the

ip
initial braid angle of the six parallel McKibben muscles. Hence, it was desired to
obtain an accurate estimate of the effective initial braid angle from experimental

cr
data. The procedure for determining this involved measuring the volume of oil
340 delivered to the McKibben muscles as a function of robot arm angle. An Omega

us
FPD2003 0.114-26.5 Lpm (0.03-7 gpm) range positive displacement flowmeter
measured the volume delivered, while the aforementioned encoder measured

an
the arm angle. As stated previously, we assume that each McKibben muscle is
fabricated identically and therefore all six HAMs are operating with the same
345 effective initial braid angle. We can subtract the initial volume from Eq. (1) to
M
get an expression for the volume delivered to the muscles as the arm contracts
from its initial condition, given as
d

" #
3
(l0 − x) (l0 − x)
∆Vm = 6 · πr02 − − l0 . (8)
te

sin2 (α0 ) l02 tan2 (α0 )

To get Eq. (8) in terms of robot arm angle, Eq. (2) is substituted; using this
p

result and experimental data, a least squares fit was performed to obtain the
ce

350 effective initial braid angle of 29.7◦ . The model and experimental data is plotted
in Fig. 6 showing good agreement.
Ac

4.2. Electrohydraulic servovalve flow gain

In most applications, a reasonable estimate of the flow gain can be made

from an EHSV’s datasheet and using Eq. (5). However, we are supplying the
355 valve with a pressure of 7.6 bar (110 psi) – about 3.7% of the valve’s rated supply
pressure of 207 bar (3000 psi) – and our EHSV can allow more flowrate than our
pump can supply. Therefore in order to obtain a more accurate representation of
the flow gain of our combined valve-pump system, several tests were performed.

16

Page 16 of 35
 100

80

Volume Delivered [ml]

t
60

ip
40

cr
20
Data
Model
0
−40 −30 −20 −10 0 10 20
Robot Arm Angle [deg]

us
Figure 6: Volume of oil delivered as a function of robot arm angle for the six parallel McKibben
muscles and the associated model for identification of the effective initial braid angle.

an
First, the A and B ports of the valve were connected directly. The same Omega
360 FPD2003 flowmeter measured the volumetric flowrate supplied to the valve.
M
The commanded valve voltage was sinusoidally cycled from -10 to 10 V at 0.25
Hz for five different supply pressures, ranging from 3.45 bar (50 psi) to 17.24
bar (250 psi) in 3.45 bar (50 psi) increments.
d

Fig. 7 (a) shows the EHSV no-load flowrate as a function of the commanded
valve voltage for a supply pressure of 17.24 bar (250 psi). This plot is included
te

365

for two reasons: 1) to show that the assumption of a critically centered lap
condition in the valve spool is reasonable, marked by the minimal deadband
p

exhibited around the null condition, and 2) to demonstrate how the flow gain
ce

at a given pressure is determined by taking the average voltage at ± 40% of the

370 saturation flowrate (maximum flowrate attained at maximum valve command
for that supply pressure) and finding the slope between these two points, shown
Ac

as the red dots and red dashed line respectively [56]. The flow gain found from
Fig. 7 (a) yields the far right data point in Fig. 7 (b). Fig. 7 (b) shows the
other pressures tested as well as the model of the flow gain as a function of
375 supply pressure using Eq. (5). A least squares fit was used to determine an
estimate of cv in this equation; a value of 1.34 Lpm·V−1 ·bar−1/2 was identified,
whereas 1.07 Lpm·V−1 ·bar−1/2 was calculated from the datasheet. Expressing
the model of the flow gain over a range of pressures is convenient for allowing

17

Page 17 of 35
 the selection of different system supply pressures. The supply pressure of our
380 system averaged 7.6 bar (110 psi), hence the flow gain at this pressure was used
throughout.

t
ip
5
10 Flowrate at PS = 17 bar
Average Flow Gain, Kq0
Valve No−Load Flow [Lpm]

cr
5

Flow Gain [Lpm/V]

Kq0 3
1 +/− 40%
0 Q
sat

us
2

−5
1
Flow Gain Data
−10 Flow Gain Model
0

an
−10 −5 0 5 10 0 5 10 15 20
Commanded Valve Voltage [V] Pressure [bar]

(a) (b)
Figure 7: (a) Experimentally determined average null flow gain at 17 bar (250 psi), and (b)
null flow gain as a function of pressure and associated model.
M
5. Angular position tracking experiments for developed controllers
d

In order to evaluate the performance of the developed controllers, a series

te

of experiments was conducted on the robot arm. First, we consider the angular
385 position tracking abilities following a sinusoidal reference trajectory. Then, we
p

do this over a range of frequencies from 1/8 to 1 Hz in 1/8 Hz increments to

ce

obtain a frequency response. Lastly, a smooth approximation of a step response

test is conducted.
Ac

5.1. Sinusoidal angular position tracking

390 In observing the sinusoidal angular position tracking of each controller in

detail, we consider 1/4 Hz as the lower frequency and 1 Hz as the higher fre-
quency tested in our range of interest. Experiments at these two frequencies
with a 10 degree peak-to-peak amplitude are shown for each of the developed
controllers. In addition to the reference and measured angles, the instantaneous
395 error is plotted.

18

Page 18 of 35
 Beginning with the baseline of comparison, the PI-controller tracking of the
1/4 Hz sine wave is shown in Fig. 8 (a), and its tracking of the 1 Hz sine wave
is shown in Fig. 8 (b). At 1/4 Hz, the peak-to-peak tracking is reasonable,

t
ip
however there is a significant phase lag of approximately 67 degrees, resulting
400 in the large instantaneous error. At 1 Hz, the attenuation is so severe that it is

cr
only reaching about 28% of the reference amplitude, and the phase lag is even
worse than in Fig. 8 (a).

us
Robot Arm Angle, θ [deg]

Robot Arm Angle, θ [deg]

θd θd
−5 −5
θPI θPI

−10 −10

an
−15 1/4 Hz
−15 1 Hz

0 5 10 15 0 1 2 3 4
Error [deg]

Error [deg]

5 5
M
0 0
−5 −5
0 5 10 15 0 1 2 3 4
Time [s] Time [s]

(a) (b)
d

Figure 8: Sinusoidal angular position tracking for PI-control at (a) 1/4 Hz, and (b) 1 Hz.
te

Next, the same two tracking tasks were tested using the feedforward term
added to the PI-controller as shown in Fig. 4. In Fig. 9 (a), it is clear that the
p

405 1/4 Hz sine wave tracking reaches the desired 10 degree peak-to-peak amplitude,
ce

like with the PI-controller in Fig. 8 (a), however the phase lag has been entirely
removed. At 1 Hz, the tracking shown in Fig. 9 (b) is drastically improved over
Fig. 8 (b) as both the attenuation and phase lag have been almost completely
Ac

eliminated. There is a slight lag, especially at the bottom of the trajectory, but
410 overall the shape of the reference is mostly preserved in the response.
Testing the last developed controller, with the addition of the inner pressure
feedback loop to the second architecture as shown in Fig. 5, the same tracking
tasks were once again performed. This is shown in Fig. 10, and tracking in both
(a) and (b) is quite similar to that seen in Fig. 9 (a) and (b). The main difference
415 being that the shape of the reference trajectory appears to be slightly better

19

Page 19 of 35
 Robot Arm Angle, θ [deg]

Robot Arm Angle, θ [deg]

θd θd
−5 −5
θ θ
PI,FF PI,FF

t
−10 −10

ip
−15 1/4 Hz
−15 1 Hz

0 5 10 15 0 1 2 3 4
Error [deg]

Error [deg]

cr
5 5
0 0
−5 −5
0 5 10 15 0 1 2 3 4
Time [s] Time [s]

us
(a) (b)
Figure 9: Sinusoidal angular position tracking for PI-control with a model-based feedforward
term at (a) 1/4 Hz, and (b) 1 Hz.

an
preserved with a little less overshoot. Therefore the addition of the cascaded
architecture seems to have minimally changed the tracking of Fig. 9, but is
M
significantly better than Fig. 8. The benefits of the this nested pressure loop
will become more clear in Section 5.3.
Robot Arm Angle, θ [deg]

Robot Arm Angle, θ [deg]

d
θd θd
−5 −5
θPI−P,FF θPI−P,FF
te

−10 −10

−15 1/4 Hz
−15 1 Hz
p

0 5 10 15 0 1 2 3 4
Error [deg]

Error [deg]

5 5
0 0
ce

−5 −5
0 5 10 15 0 1 2 3 4
Time [s] Time [s]

(a) (b)
Ac

Figure 10: Sinusoidal angular position tracking for cascaded PI-P-control with a model-based
feedforward term at (a) 1/4 Hz, and (b) 1 Hz.

420 5.2. Frequency response

For the frequency response experiments, the previous tests in Section 5.1
were repeated to span the frequency range from 1/8 Hz to 1 Hz in 1/8 Hz
increments. To quantify the performance differences between the three control

20

Page 20 of 35
 architectures, a Bode plot and root mean square (RMS) error plot are included
425 in Fig. 11 (a) and (b) respectively.
It is clear in the Bode magnitude plot that solely PI-control is insufficient

t
ip
since the −3 dB bandwidth occurs at about 3/8 Hz. Additionally, there is a
significant overshoot at 1/8 Hz, and as frequency is increased, the attenuation

cr
quickly becomes unacceptable (up to −11 dB at 1 Hz). The phase lag is quite
430 large over the full frequency range tested as well. Both the PI-control with

us
feedforward and cascaded PI-P control with feedforward show similar improved
Bode characteristics, with the exception that the former shows slightly more
overshoot in the mid-range of frequencies tested, and the latter remains slightly

an
closer to 0 dB throughout. Similar results are seen in the RMS error plot,
435 where the PI-controller quickly jumps to over 4 degrees of error, while both the
PI-control with feedforward and cascaded PI-P-control with feedforward stay
M
below 1 degree of error for the majority of frequencies tested.

3 5
Magnitude [dB]

0
d
−3
4
−6 PI
PI,FF
RMS Error [deg]

−9
te

PI−P,FF
−12 3
1/8 1/4 3/8 1/2 5/8 3/4 7/8 1 PI
PI,FF
PI−P,FF
2
0
Phase [deg]

−25
p

−50 1
−75
−100
ce

−125 0
1/8 1/4 3/8 1/2 5/8 3/4 7/8 1 1/8 1/4 3/8 1/2 5/8 3/4 7/8 1
Frequency [Hz] Frequency [Hz]

(a) (b)

Figure 11: (a) Bode plot, and (b) RMS error plot of all controllers tested tracking a 10 degree
Ac

peak-to-peak amplitude reference.

5.3. Step response

The last test used to evaluate the different controllers was a step response.
440 Since a true step function would demand infinite flowrate from the feedforward
term, a ‘smooth’ step approximation was used. Sohl et al. utilized a 7th or-
der polynomial step approximation [47], and Weisstein presented various ways

21

Page 21 of 35
 including exponentials or a hyperbolic tangent function [57]. We chose to em-
ploy a hyperbolic tangent approximation of a 10 degree step with a 10 to 90%
rise time of 0.088 s. This represents a rapid command, but does not saturate

t
445

ip
the valve command signal (although it is in the highly nonlinear region); this
approximation is given by

cr
  
1 t−5
θd (t) = 10 · tanh − 10 . (9)

us
2 0.04

Using Eq. (9) as the reference, the smooth step response of the three con-

an
trollers is shown in Fig. 12 (a). Visually comparing these, the PI-controller, in
450 the top portion of this plot, has the slowest response in terms of both rise and
settling times, and has the most overshoot. This makes sense since it has the
M
lowest commanded valve voltage initially, seen in Fig. 12 (b). The PI-controller
with feedforward compensation, whose response is given in the middle of Fig.
12 (a), has a faster rise time due to the anticipative action of the feedforward
d

455 term, noted by the large spike in control effort initially in Fig. 12 (b), but has
a similar settling time. The overshoot is also smaller, due to the integrator
te

not winding up as much because the feedforward term reduces the initial error.
Lastly, the cascaded PI-P-controller with feedforward is presented in the bottom
p

of this figure. This controller clearly has the fastest rise time and settling time
ce

460 when looking at both Fig. 12 (a) and (b), but has some additional oscillation
near the step event.
Ac

6. Discussion

While the feedforward implemented in this work provides anticipative ac-

tion in response to changing command profiles, it has zero robustness in terms
465 of changing loads, different operating conditions (such as temperature), or servo-
loop degradation [42]. The feedforward is intended to provide most of the con-
trol effort, while the feedback control portion handles the remaining errors and
unpredictable system changes.

22

Page 22 of 35
 −5 8
uv,PI
−10 6 u
θ θ 0.8 v,PI,FF
Robot Arm Angle, θ [deg]

d PI
−15 4 uv,PI−P,FF

Valve Voltage, uv [V]

0 2 4 6 8

t
0.6
2
−5

ip
0.4 0
−10 θ θ
d PI,FF 0 1 2
−15
0 2 4 6 8 0.2

cr
−5 0
−10 θ θ
d PI−P,FF
−15 −0.2
0 2 4 6 8 0 2 4 6 8
Time [s] Time [s]

us
(a) (b)
Figure 12: (a) Step response of all controllers tested (from top to bottom: PI, PI with
feedforward, and cascaded PI-P with feedforward), and (b) associated control inputs.

470
One concern of using this feedforward term is its inability to manage chang-
ing loads. Varying loads correspond to varying pressures, which in turn affects
an
M
the flowrates out of the valve and degrades tracking. Therefore it was desired to
observe the pressure ranges experienced in the most demanding tracking tests
above. Even though we considered a fixed load, since we are using McKibben
d

muscles, the pressure must change to achieve different robot arm angles.
The ratio of the measured actuator pressure to the measured supply pres-
te

475

sure for the 1 Hz tracking tasks with all controllers tested is given in Fig. 13
(a). This ratio is presented instead of the raw actuator pressure since it better
p

represents the operating range and proximity to saturation. It is clear that a

ce

large pressure range was spanned while the tracking with the feedforward term
480 was still acceptable. The pressure range experienced was wider than the recom-
mended range by Merritt for control authority purposes [51], showing that the
Ac

feedforward is still helpful even though the pressure largely deviates from the
1
2 PS linearization point.
The commanded valve voltages for all tested controllers during the 1 Hz
485 sinusoidal tracking task are presented in Fig. 13 (b). The calculated feedforward
portion is given as well. The fact that the magnitude of the commanded voltages
do not exceed 1.5 V demonstrates that the voltages stayed within the linear
region of the valve flow-voltage relationship; hence from this perspective, the null

23

Page 23 of 35
 flow gain used is an accurate representation of this relationship. It is also noted
490 how closely the feedforward voltage matches the ultimately commanded valve
voltages for the PI with feedforward, and cascaded PI-P with feedforward. This

t
ip
shows that the feedforward term is a good approximation of the demands of the
system, and the feedback components are making relatively small contributions

cr
to compensate for the small errors mainly due to unmodeled dynamics such as
495 acceleration.

us
100
PI PI,FF PI−P,FF PI PI,FF PI−P,FF FF
2
80

Valve Voltage, uv [V]

1
Pm / PS × 100 [%]

an
60
0
40
−1
20
M
1 Hz −2 1 Hz
0
0 1 2 3 4 0 1 2 3 4
Time [s] Time [s]

(a) (b)
d

Figure 13: (a) The ratio of actuator pressure to supply pressure, and (b) the commanded
valve voltages of all controllers tested at 1 Hz.
te

A summary of some key performance metrics comparing the three controllers

p

is presented in Table 2. The PI-controller is the simplest and cheapest to im-

plement, but also has the worst dynamic performance of the controllers tested.
ce

With the addition of the model-based feedforward term, the tracking perfor-
500 mance was significantly improved. It is important to note that this controller
has the same disturbance rejection as the PI-only controller. The last control
Ac

scheme, which added the nested pressure feedback loop within the second archi-
tecture, showed overall the best dynamic performance due to improved settling
time, however this controller has the most gains to tune and requires an addi-
505 tional sensor (pressure transducer).

24

Page 24 of 35
 Table 2: Comparison of controllers tested.

PI PI,FF PI-P,FF

t
Rise Time (0→100%) [s] 0.88 0.61 0.27

ip
Delay Time [s] 0.26 0.13 0.16
Settling Time (5%) [s] 7.52 7.38 1.81
Max Overshoot [%] 22.8 13.8 20.3

cr
−3 dB Bandwidth poor good good
Phase Lag poor good good
Disturbance Rejection fair fair good
No. of Gains to Tune 2 3 4

us
No. of Sensors 1 1 2

an
7. Conclusions

This study experimentally investigated the control of HAMs, demonstrating

that hydraulic actuation and control of McKibben muscles at supply pressures
M
less than 4% of the 207 bar (3000 psi) of traditional hydraulics is feasible. As
510 a baseline, we first showed that using solely PI-control results in poor tracking,
d
marked by a large phase lag at low frequencies and severe attenuation at higher
frequencies. Next, we added to the PI-controller a novel model-based feedfor-
te

ward term that takes into account the predicted nonlinear volumetric flowrate
demands of the hydraulic McKibben muscles, as well as the linearized flow gain
p

515 of the electrohydraulic servovalve; the phase lag and attenuation exhibited by
ce

the PI-controller were almost completely eliminated over the frequency range
tested. Lastly, an inner pressure feedback loop was included in the previous
scheme to create a cascaded PI-P-controller with the feedforward term; this
Ac

architecture decreased the rise and settling times, and while not tested, also
520 theoretically allows for improved disturbance rejection due to the additional
sensing of the faster inner pressure loop dynamics.
These augmentations of classical control techniques were satisfactory for
our tracking purposes. By leveraging parameters that McKibben muscle users
should have available, these improvements are accessible and readily realizable.
525 While the flowrate demands dominated the feedforward signal in our applica-

25

Page 25 of 35
 tion, due to the versatility of this framework, extensions could easily be made to
include the effects of other nonlinearities such as friction or the impact load has
on flowrate. If performance is limited by system force capability, including an

t
ip
antagonistic pair, additional McKibben muscles, or increasing operating pres-
530 sure will yield faster response times. If an asymmetric load is experienced, the

cr
valve flow gain in the feedforward branch could be split into two values: one for
raising the arm, and one for lowering. Improved tracking is also possible with

us
the addition of an acceleration feedforward, which is a relatively common next
step in hydraulic cylinder control, however it too would need to be adjusted
535 for the nonlinear behavior of the McKibben muscles. Further performance en-

an
hancement may also be possible by taking into account the pressure dynamics
and using feedback linearization, or moving to an adaptive control methodology.
M
Acknowledgments

The authors gratefully acknowledge the financial support provided by the

d

540 Defense Advanced Research Projects Agency (DARPA) through the Maximum
Mobility and Manipulation (M3) Program under the direction of Dr Gill Pratt
te

(grant number W31P4Q-13-1-0012).

p

References
ce

[1] Shadow Robot Company, The Shadow Biped, http://www.shadow.org.

545 uk/projects/biped.shtml, accessed: 2016-09-07.
Ac

[2] B. Tondu, S. Ippolito, J. Guiochet, A. Daidie, A seven-degrees-of-freedom

robot-arm driven by pneumatic artificial muscles for humanoid robots, The
International Journal of Robotics Research 24 (4) (2005) 257–274. doi:
10.1177/0278364905052437.

550 [3] FESTO, Airic’s arm: Robot arm with fluidic muscles, https://www.
festo.com/net/SupportPortal/Files/42058/Airics_arm_en.pdf, ac-
cessed: 2016-09-07.

26

Page 26 of 35
 [4] D. Büchler, H. Ott, J. Peters, A lightweight robotic arm with pneumatic
muscles for robot learning, in: 2016 IEEE International Conference on
Robotics and Automation (ICRA), IEEE, 2016, pp. 4086–4092. doi:10.

t
555

ip
1109/ICRA.2016.7487599.

[5] D. P. Ferris, J. M. Czerniecki, B. Hannaford, U. of Washington, An ankle-

cr
foot orthosis powered by artificial pneumatic muscles, Journal of applied
biomechanics 21 (2) (2005) 189.

us
560 [6] T.-J. Yeh, M.-J. Wu, T.-J. Lu, F.-K. Wu, C.-R. Huang, Control of McK-
ibben pneumatic muscles for a power-assist, lower-limb orthosis, Mecha-

an
tronics 20 (6) (2010) 686–697. doi:10.1016/j.mechatronics.2010.07.
004.

[7] Y.-L. Park, B.-r. Chen, N. O. Pérez-Aranciba, D. Young, L. Stirling, R. J.

M
565 Wood, E. C. Goldfield, R. Nagpal, Design and control of a bio-inspired
soft wearable robotic device for ankle-foot rehabilitation, Bioinspiration &
d
Biomimetics 9 (1) (2014) 016007. doi:10.1088/1748-3182/9/1/016007.
te

[8] B. Tondu, Modelling of the McKibben artificial muscle: A review, Journal

of Intelligent Material Systems and Structures 23 (3) (2012) 225–253. doi:
10.1177/1045389X11435435.
p

570

[9] C.-P. Chou, B. Hannaford, Measurement and modeling of McKibben pneu-

ce

matic artificial muscles, Robotics and Automation, IEEE Transactions on

12 (1) (1996) 90–102. doi:10.1109/70.481753.
Ac

[10] M. F. Gentry, N. M. Wereley, Effects of braid angle on pneumatic artificial

575 muscle actuator performance, in: ASME 2008 Conference on Smart Ma-
terials, Adaptive Structures and Intelligent Systems, American Society of
Mechanical Engineers, 2008, pp. 617–623. doi:10.1115/SMASIS2008-539.

[11] N. Tsagarakis, D. G. Caldwell, Improved modelling and assessment of pneu-

matic muscle actuators, in: Robotics and Automation, 2000. Proceedings.

27

Page 27 of 35
580 ICRA’00. IEEE International Conference on, Vol. 4, IEEE, 2000, pp. 3641–
3646.

t
[12] A. Parr, Hydraulics and pneumatics: a technician’s and engineer’s guide,

ip
Boston, MA: Butterworth-Heinemann, 2010.

[13] J. F. Blackburn, G. Reethof, J. L. Shearer, Fluid Power Control, The Tech-

cr
585 nology Press of M.I.T., 1960.

us
[14] M. Focchi, E. Guglielmino, C. Semini, A. Parmiggiani, N. Tsagarakis,
B. Vanderborght, D. G. Caldwell, Water/air performance analysis of a flu-
idic muscle, in: Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ

an
International Conference on, IEEE, 2010, pp. 2194–2199. doi:10.1109/
590 IROS.2010.5650432.
M
[15] M. A. Meller, M. Bryant, E. Garcia, Reconsidering the McKibben mus-
cle: energetics, operating fluid, and bladder material, Journal of Intelli-
gent Material Systems and Structures 25 (18) (2014) 2276–2293. doi:
d

10.1177/1045389X14549872.
te

595 [16] M. Mori, K. Suzumori, M. Takahashi, T. Hosoya, Very high force hy-
draulic McKibben artificial muscle with a p-phenylene-2, 6-benzobisoxazole
p

cord sleeve, Advanced Robotics 24 (1-2) (2010) 233–254. doi:10.1163/

016918609X12586209967366.
ce

[17] R. Tiwari, M. A. Meller, K. B. Wajcs, C. Moses, I. Reveles, E. Garcia,

600 Hydraulic artificial muscles, Journal of Intelligent Material Systems and
Ac

Structures 23 (3) (2012) 301–312. doi:10.1177/1045389X12438627.

[18] D. Sangian, S. Naficy, G. M. Spinks, B. Tondu, The effect of geometry

and material properties on the performance of a small hydraulic McKibben
muscle system, Sensors and Actuators A: Physical 234 (2015) 150–157.
605 doi:10.1016/j.sna.2015.08.025.

[19] C. Xiang, M. E. Giannaccini, T. Theodoridis, L. Hao, S. Nefti-Meziani,

S. Davis, Variable stiffness McKibben muscles with hydraulic and pneu-

28

Page 28 of 35
 matic operating modes, Advanced Robotics 30 (13) (2016) 889–899. doi:
10.1080/01691864.2016.1154801.

t
610 [20] E. Corporation, Eaton Mobile Hydraulics Manual, Maumee, OH: Eaton

ip
Hydraulics Training Services, 2010.

cr
[21] K. Takashima, J. Rossiter, T. Mukai, McKibben artificial muscle using
shape-memory polymer, Sensors and Actuators A: Physical 164 (1) (2010)
116–124. doi:10.1016/j.sna.2010.09.010.

us
615 [22] D. Sangian, S. Naficy, G. M. Spinks, Thermally activated paraffin-filled
McKibben muscles, Journal of Intelligent Material Systems and Structures

an
(2016) 1045389X16633766doi:10.1177/1045389X16633766.

[23] D. Sangian, J. Foroughi, S. Farajikhah, S. Naficy, G. M. Spinks, A bladder-

M
free, non-fluidic, conductive mckibben artificial muscle operated electro-
620 thermally, Smart Materials and Structures 26 (1) (2016) 015011.

[24] E. J. Ball, M. A. Meller, J. B. Chipka, E. Garcia, Modeling and testing

d

of a knitted-sleeve fluidic artificial muscle, Smart Materials and Structures

te

25 (11) (2016) 115024. doi:10.1088/0964-1726/25/11/115024.

[25] M. Bryant, M. A. Meller, E. Garcia, Variable recruitment fluidic artificial

p

625 muscles: modeling and experiments, Smart Materials and Structures 23 (7)
ce

(2014) 074009. doi:10.1088/0964-1726/23/7/074009.

[26] R. M. Robinson, C. S. Kothera, N. M. Wereley, Variable recruitment test-

Ac

ing of pneumatic artificial muscles for robotic manipulators, IEEE/ASME

Transactions on Mechatronics 20 (4) (2015) 1642–1652. doi:10.1109/
630 TMECH.2014.2341660.

[27] S. A. DeLaHunt, T. E. Pillsbury, N. M. Wereley, Variable recruit-

ment in bundles of miniature pneumatic artificial muscles, Bioinspiration
& Biomimetics 11 (5) (2016) 056014. doi:10.1088/1748-3190/11/5/
056014.

29

Page 29 of 35
635 [28] M. Meller, J. Chipka, A. Volkov, M. Bryant, E. Garcia, Improving ac-
tuation efficiency through variable recruitment hydraulic McKibben mus-
cles: modeling, orderly recruitment control, and experiments, Bioinspira-

t
ip
tion & Biomimetics 11 (6) (2016) 065004. doi:10.1088/1748-3190/11/
6/065004.

cr
640 [29] D. G. Caldwell, G. A. Medrano-Cerda, M. Goodwin, Control of pneumatic
muscle actuators, IEEE control systems 15 (1) (1995) 40–48. doi:10.1109/

us
37.341863.

[30] B. Tondu, P. Lopez, Modeling and control of McKibben artificial muscle

an
robot actuators, Control Systems, IEEE 20 (2) (2000) 15–38. doi:10.
645 1109/37.833638.

[31] R. W. Colbrunn, G. M. Nelson, R. D. Quinn, Design and control of a

M
robotic leg with braided pneumatic actuators, in: Intelligent Robots and
Systems, 2001. Proceedings. 2001 IEEE/RSJ International Conference on,
Vol. 2, IEEE, 2001, pp. 992–998. doi:10.1109/IROS.2001.976298.
d

650 [32] J. Schröder, D. Erol, K. Kawamura, R. Dillman, Dynamic pneumatic

te

actuator model for a model-based torque controller, in: Computational

Intelligence in Robotics and Automation, 2003. Proceedings. 2003 IEEE
p

International Symposium on, Vol. 1, IEEE, 2003, pp. 342–347. doi:

ce

10.1109/CIRA.2003.1222113.

655 [33] J. H. Lilly, P. M. Quesada, A two-input sliding-mode controller for a pla-

nar arm actuated by four pneumatic muscle groups, Neural Systems and
Ac

Rehabilitation Engineering, IEEE Transactions on 12 (3) (2004) 349–359.

doi:10.1109/TNSRE.2004.831490.

[34] A. Hildebrandt, O. Sawodny, R. Neumann, A. Hartmann, Cascaded con-

660 trol concept of a robot with two degrees of freedom driven by four artificial
pneumatic muscle actuators, in: Proceedings of the 2005, American Con-
trol Conference, 2005., IEEE, 2005, pp. 680–685. doi:10.1109/ACC.2005.
1470036.

30

Page 30 of 35
 [35] E. A. Bubert, Highly extensible skin for a variable wing-span morphing
665 aircraft utilizing pneumatic artificial muscle actuation, ProQuest, 2009.

t
[36] T. V. Minh, T. Tjahjowidodo, H. Ramon, H. Van Brussel, Cascade position

ip
control of a single pneumatic artificial muscle–mass system with hystere-
sis compensation, Mechatronics 20 (3) (2010) 402–414. doi:10.1016/j.

cr
mechatronics.2010.03.001.

[37] M. De Volder, A. Moers, D. Reynaerts, Fabrication and control of miniature

us
670

McKibben actuators, Sensors and Actuators A: Physical 166 (1) (2011)

111–116. doi:10.1016/j.sna.2011.01.002.

an
[38] B. K. Woods, Y.-T. Choi, C. S. Kothera, N. M. Wereley, Control system
development for pneumatic artificial muscle-driven active rotor systems,
Journal of Guidance, Control, and Dynamics 36 (4) (2013) 1177–1185. doi:
M
675

10.2514/1.56528.

[39] D. G. Caldwell, A. Razak, M. Goodwin, Braided pneumatic muscle actua-

d

tors, IFAC Proceedings Volumes 26 (1) (1993) 522–527.

te

[40] H. Yang, C. Xiang, L. Hao, L. Zhao, B. Xue, Research on psa-mfac for a

680 novel bionic elbow joint system actuated by pneumatic artificial muscles,
p

Journal of Mechanical Science and Technology 31 (7) (2017) 3519–3529.

ce

[41] S. Skinner, R. Long, Closed loop electrohydraulic systems manual,

Rochester Hills, MI: s Incorporated Training Center, 1998.
Ac

[42] J. L. Johnson, Designer’s handbook for electrohydraulic servo and propor-

685 tional systems, IDAS Engineering, 2000.

[43] N. Manring, Hydraulic control systems, John Wiley & Sons Incorporated,
2005.

[44] P. E. Nachtwey, Practical design for fluid power motion control, Delta Com-
puter Systems, Inc., Battle Groung, WA, 2013.

31

Page 31 of 35
690 [45] W. Kobayashi, K. Ito, S.-i. Yamamoto, Displacement control of water
hydraulic mckibben muscles with load compensation, JFPS International
Journal of Fluid Power System 8 (2) (2014) 107–112.

t
ip
[46] T. D. C. Thanh, K. K. Ahn, Nonlinear pid control to improve the control
performance of 2 axes pneumatic artificial muscle manipulator using neural

cr
695 network, Mechatronics 16 (9) (2006) 577–587.

[47] G. A. Sohl, J. E. Bobrow, Experiments and simulations on the nonlinear

us
control of a hydraulic servosystem, IEEE transactions on control systems
technology 7 (2) (1999) 238–247. doi:10.1109/87.748150.

an
[48] J. K. Moore, S. K. Hnat, A. J. van den Bogert, An elaborate data set
700 on human gait and the effect of mechanical perturbations, PeerJ 3 (2015)
e918. doi:10.7717/peerj.918.
M
[49] B. D. Tyreus, W. L. Luyben, Tuning PI controllers for integrator/dead
time processes, Industrial & Engineering Chemistry Research 31 (11) (1992)
d
2625–2628. doi:10.1021/ie00011a029.

[50] J.-J. E. Slotine, W. Li, et al., Applied nonlinear control, Vol. 199, prentice-
te

705

Hall Englewood Cliffs, NJ, 1991.

p

[51] H. E. Merritt, Hydraulic Control Systems, John Wiley & Sons, Inc., 1967.
ce

[52] Moog, Moog g761/761 series flow control servo valves, http:
//www.moog.com/literature/ICD/Moog-ServoValves-G761_
710 761Series-Catalog-en.pdf, accessed: 2016-02-02 (2014).
Ac

[53] N. Tsagarakis, D. Caldwell, G. Medrano-Cerda, A 7 DOF pneumatic mus-

cle actuator (pMA) powered exoskeleton, in: Robot and Human Interac-
tion, 1999. RO-MAN’99. 8th IEEE International Workshop on, IEEE, 1999,
pp. 327–333. doi:10.1109/ROMAN.1999.900361.

715 [54] E. G. Hocking, N. M. Wereley, Analysis of nonlinear elastic behavior in

miniature pneumatic artificial muscles, Smart Materials and Structures
22 (1) (2012) 014016.

32

Page 32 of 35
 [55] T. E. Marlin, Process control, McGraw-Hill New York, 2000.

[56] Electrohydraulic valves... a technical look, Tech. rep., Moog, Inc.

t
ip
720 [57] E. W. Weisstein, Heaviside step function, from MathWorld–a Wolfram web
resource, http://mathworld.wolfram.com/HeavisideStepFunction.

cr
html, accessed: 2016-09-07.

us
an
M
d
p te
ce
Ac

33

Page 33 of 35
 Biographies

Michael Meller received dual B.S. degrees in mechanical

t
engineering and aerospace engineering from the University

ip
at Buffalo, Buffalo, NY, USA, in 2010. He received the M.S.
and Ph.D degrees in mechanical engineering from Cornell
University, Ithaca, NY, USA in 2014 and 2017 respectively.

cr
Dr. Meller is a systems engineer in the industry currently fo-
cused on hydraulic actuation systems. His research interests
include bioinspired actuation and control to produce highly
efficient motion, ranging from ground robots to aerial vehicles.

us
Boris Kogan received the B.S. degree in mechanical engi-
neering from Technion - Israel Institute of Technology, Haifa,

an
Israel, in 2005, and the M.S. degrees in mechanical engi-
neering and aerospace engineering from Cornell University,
Ithaca, NY, USA in 2014 and 2015 respectively. Boris con-
tinues to fill various roles in the industry as a mechanical
engineer in the fields of precision mechanics, robotic actu-
M
ation and controls. His research interests include precision
engineering, actuation of various scales, mechatronics and control.

Matthew Bryant received the B.S. degree in mechanical

d

engineering from Bucknell University, Lewisburg, PA, USA

in 2007 and the M.S. and Ph.D. degrees in mechanical en-
te

gineering from Cornell University, Ithaca, NY, USA in 2011

and 2012 respectively. Dr. Bryant is an Assistant Professor
in the Department of Mechanical and Aerospace Engineering
p

at North Carolina State University, Raleigh, NC, USA. His

research interests include the dynamics and control of smart
ce

structures and mechatronic systems with applications to novel actuation tech-

niques, energy harvesting devices, fluid-structure interactions, and robotics in
the land, air, and underwater domains.
Ac

Ephrahim Garcia received the B.S., M.S., and Ph.D. de-

grees in aerospace engineering from the University at Buf-
falo, Buffalo, NY, USA, in 1985, 1988, and 1990 respectively.
He was a Professor in the Sibley School of Mechanical and
Aerospace Engineering at Cornell University, Ithaca, NY,
USA when he passed away in 2014. His wide array of research
interests included mechatronics, smart materials utilized in
energy harvesting and actuation, morphing and perching air-
craft, and bioinspired artificial muscle actuators.

34

Page 34 of 35
 Model-based feedforward and cascade control of hydraulic McKibben
muscles
Michael Meller, Boris Kogan, Matthew Bryant, Ephrahim Garcia

t
ip
Highlights
• Experimentally showed PI-control of hydraulic McKibben muscles yields large error.

cr
• Exhibitedsignificantly enhanced tracking with simple augmentations to PI-control.
• Developed feedforward compensation comprised of McKibben and valve flowrate models.
• Demonstrated increased bandwidth and reduced phase lagwith developed feedforward.

us
• Improved rise and settling times with addition of an inner pressure feedback loop.

an
M
e d
pt
ce
Ac

Page 35 of 35