Professional Documents
Culture Documents
Closed Loop Control With Single Integral Action
Closed Loop Control With Single Integral Action
Bertrand Tondu
CNRS ; LAAS ; 7 Avenue du colonel Roche, F-31077 Toulouse Cedex 4, France
University of Toulouse, UPS, INSA, INP, ISAE ; UT1, UTM, LAAS ; F-31077 Toulouse Cedex 4, France
bertrand.tondu@insa-toulouse.fr
Abstract—Artificial muscles constitute a large class of non-linear we propose a simple non-linear model of a contractile artificial
actuators characterized by their own stiffness and damping. We muscle including a damping factor from which are derived
analyze in the case of an axial contraction artificial muscle the open-loop and closed-loop positioning simulations; a
relevance of a single integral action closed-loop controller. The preliminary stability analysis is then developed in section III
basic idea of our approach consists in taking advantage of before reporting in section IV experiments performed with a
artificial muscle own stiffness and damping in order to substitute pneumatic McKibben artificial muscle.
for a classic PID-controller an I-controller with as a consequence
only one parameter to tune. Step and tracking responses
performed with a pneumatic textile-braided McKibben muscle II. MODELLING AND SIMULATION OF OPEN-AND-CLOSED
are reported showing the practical efficiency of the method to LOOP CONTROL OF A TYPICAL ARTIFICIAL MUSCLE
combine accuracy and load robustness performances.
A. A simple phenomenological model and its positioning
Keywords: Artificial muscle closed-loop control; McKibben muscle performances in open-loop
The notion of artificial model is defined in analogy with the
I. INTRODUCTION phenomenological behaviour of the skeletal muscle. If we don’t
Artificial muscles are a large class of actuators take into account the passive elasticity, a simple model of the
characterized by an open-loop positioning controllability in force generated by the skeletal muscle can be written as
analogy with skeletal muscle behaviour. The artificial muscle follows:
control variable plays the role of the neural activation making
possible to adapt muscle stiffness as the natural muscle does. x 0 ≤ u ≤ 1
F = uFmax (1 − ) − Fdamp ( x& ) with (1)
Numerous technologies with various physical or chemical xmax 0 ≤ x ≤ xmax
control variables were developed in the last decade with the
hope to substitute for classic actuators these ‘soft actuators’ in a where x is the shortening length of the muscle supposed to be
large range of applications where lightness, compactness and inextensible since no passive elongation is considered and u is
natural compliance are necessary. However any artificial a normalized control variable. The first term of the model
muscle technology is generally based on the use of soft and/or catches the static behaviour of the skeletal muscle by
shape changing materials inducing highly non-linear linearizing the spring-like character of the muscle with a
phenomena with, as a consequence, a real difficulty to get an stiffness proportional to u; as a consequence, Fmax represents
accurate positioning of the artificial muscle in closed-loop. the maximum isometric muscle force while xmax represents the
This highly non-linear character has often led researchers to maximum length shortening independent of u. At our
consider non-linear or fuzzy logic based control approaches, knowledge, N. Hogan [4] was among the first to show the
eventually combined with internal models, to try to accurately relevance of this model which however does not express the
control technical devices – essentially robots – actuated by variation of maximum contraction with neural activation. The
artificial muscles (as, for example, in [1]-[3] for limiting second term of the model represents the natural damping of the
ourselves to pneumatic artificial muscles), with as a drawback muscle without which the muscle will contract with endless
a large complexity in the control parameter tuning. The oscillations. This damping factor is a consequence of friction
originality of our approach consists in reconsidering this phenomena inside the muscle. In the framework of this paper,
problem by dealing with a single artificial muscle whose end- we will consider two friction models: the first one is a linear
position must be accurately controlled in closed-loop. To reach viscous friction model of constant coefficient we will note fv :
this goal we do the assumption that a single ‘external’
Fdamp ( x& ) = f v x& (2)
integrator combined with ‘internal’ muscle stiffness and
damping will be able to satisfy both the accuracy requirement
The second one is a non-linear friction model combining a
and the robustness requirement for load variations as manual
static friction whose constant coefficient will be noted µs with a
perturbations. The paper is organized as follows: in a section II
X-Position (cm )
where µk_lim corresponds to a limit kinetic friction coefficient rM= 0.3
and x& k is a speed constant determining the quickness with rM= 0.4
which the current kinetic friction varies towards its limit rM= 0.5
(Fig.1). If µs>µk_lim with a high value of x& k the model can be rM= 0.6
considered as approaching a classic static-kinetic friction rM= 0.7
model. If µs<µk_lim with a moderate x& k -value, this model rM= 0.8
approaches a typical textile-physics friction model rM= 0.9
characterized by a kinetic friction increasing with speed. This
is this model we privilege in our study because, on the one
hand, we think it is able to approach the dynamic behaviour of Time (s)
natural muscle [5] and, on the other hand, it makes possible to (a)
simply modelize the dynamic behaviour of textile-braided
pneumatic McKibben muscle, as we will show it in section IV. rM= 0
rM= 0.1
rM= 0.2
rM= 0.3
X-Position (cm)
rM= 0.4
Friction c oefficient
rM= 0.5
rM= 0.6
Textile-physic s inspired
kinetic friction rM= 0.7
rM= 0.8
719
curves deduced from our two simulated friction models: while
the curve is almost a straight-line in the linear viscous friction k=
I 200
case with (a/Fmax) ≈ 0.8, in the non-linear friction case the k=
I 100
corresponding curve with a (a/Fmax) ≈ 0.35 belongs to the k=
I 50
physiological range . k=
I 20
X-position (cm)
Quic k-release data and approac hed k=
I 10
veloc ity-forc e relation with (a/Fmax)= 0.8
in linear visc ous fric tion c ase k=
I 5
Quic k-release data and approac hed k=
I 4
veloc ity-forc e relation with (a/Fmax)= 0.35
V/Vm ax-ratio
Time (s)
(a)
rM= 0.02
rM= 0.05
Ratio rM= Mg/Fm ax
Figure 3. Comparison of Hill’s curves derived from our muscle model with rM= 0.1
linear viscous friction model and non-linear kinetic friction model (see text).
X-Position (c m) rM= 0.2
We will check in section IV the relevance of the non-linear rM= 0.25
friction model to capture typical McKibben artificial muscle
open-loop response and the possibility given by our closed-
loop controller to preserve this robust damping. rM= 0.3
∫
−t / T
k I ( ( xd − x)dt )(1 − e )[1 − ( x / xmax )] − rdamp ( x& ) − rM
(7)
X-Position (cm)
720
and 4.b the linear friction model is considered: an optimal now possible to apply to (13) the classic Routh-Hurwitz
kI=10 m-1s-1 value is empirically determined in Fig. 4.a for stability criteria peculiar to a third order system in the form
rM=0.1 value while Fig. 4.b shows the effects of rM-variations λ3 + Aλ2 + Bλ + C = 0 : the system is stable if and only if A>0,
for this constant kI -value. In Fig. 4.c the non-linear friction B>0, C>0 and AB>C. In our case these conditions can be
model is considered with constant kI=18 m-1s-1. Let us note that gathered in the following relationship:
although all our efforts in differential equation programming
the convergence to the desired position was generally not Int ( xd ) > ( xmax − xd ) / rv (14)
achieved. We will check it further in experimental conditions
(section IV). We can however remark the absence of which is always verified if:
overshooting during simulated contraction whatever the load. Int ( xd ) > xmax / rv (15)
III. CLOSED-LOOP STABILITY ANALYSIS This is a surprising stability condition in which kI does not
explicitly appear, for which we propose the following
From closed-loop system analysis developed in section II,
interpretation; first of all let us emphasize the fact that Int(xd) is
the desired xd-position is the alone equilibrium point which
obligatory positive due to the definition of the u-muscle control
moreover appears stable in reported simulations. In order to
belonging to the range [0-1]; secondly, Int(xd) implicitly
better understand the closed-loop stability conditions we
depends on the single kI-term control: lower will be kI, higher
propose to apply to it a classic linearization around this
will be Int(xd) since the u-muscle control is proportional to kI.
equilibrium point. We will limit in a first step our analysis to
As a consequence, for a given muscle specified by its three
the linear viscous friction case. Moreover, we consider that in
parameters (xmax, Fmax, fv) one can reasonably think to be able to
the close neighborhood of the equilibrium point no delay-effect
experimentally determine a maximum kI-value verifying (15)
due to the control transmission has to be considered. We note
and so the closed-loop system would be stable whatever the xd-
rv=fv/Fmax. From (5) we derive:
value. Moreover, if this kI-limit corresponds to the lower
(du / dt )[1 − ( x / xmax )] − u ( x / xmax ) − rv &x& = (rm + rM )&x&& / g (10) considered load, it can also be thought that a higher load would
favor, for the same kI-value, the stability by increasing Int(xd).
Let us define the state variables as follows: Lastly, the delay induced by the control transmission – at least
x1 = &x& , x2 = x& and x3 = x − xd . As a consequence the under its linear first order considered form – appears as a
linearization will be realized around a zero-equilibrium point stability factor. What happens now in the case of the non-linear
x1d= x2d=x3d=0 in such a way that x1= x1d+ε1, x2= x2d+ε2, x3= kinetic friction model? A simple reasoning consists to
x3d+ε3 . Let us define the following integral term Int(x3): substitute to the constant viscous coefficient fv the slope of the
t t
non-linear friction model at zero-velocity:
( µ k _ lim − µ s )
∫ ∫
Int ( x3 ) = ( xd − x)dt = − x3dt (11) dFNL − damp
( x& = 0) = (16)
0 0 dx& x& k
corresponding to the integration of the position-error between In particular, a textile-inspired friction model appears as a
initial state at zero-time and current state x3(t). From (10) the stability factor since its slope is maximal at zero-velocity. For
following state representation is deduced in which the term in example, in simulations reported in first section a relatively
x32 corresponding to a ε 32 -term was neglected: high ratio rv of 0.5 was considered to be compared for closed-
loop step responses with similar response times to the ratio
g k I Int ( x3 ) x corresponding to initial slope of our non-linear friction model
x&1 = ( r + r )[− rv x1 − x x2 − k I (1 − d ) x3
x equal from (16) to 3.6.
m M max max
x& 2 = x1 (12) Although our stability analysis is not able to give explicit
x& = x conditions of stability for the kI-term, it justifies, according to
3 2 us, our intuition to privilege a single integral action closed-loop
controller. We now apply this controller to the position control
In a classic way, we can deduce the matrix relationship of a classic McKibben artificial muscle.
ε& = Jε in which ε = [ε1 ε 2 ε 3 ]T and the ‘characteristic
polynomial’ corresponding to det( J − λI 3 ) = 0 where I 3 is the IV. APPLICATION TO ROBUST AND ACCURATE POSITIONING
3 × 3 unit-matrix and λ a complex variable. We get: OF MCKIBBEN PNEUMATIC MUSCLE
We consider a single pneumatic artificial muscle designed
λ3 + rv λ2 + (k I Int ( x3 ) / xmax )λ + k I [1 − ( xd / xmax )]) = 0 (13) at the laboratory and tested on the experimental set-up shown
in Fig. 5: at the muscle fixed tip, the pressurized air is
Analyzing the system stability from this equation has however
controlled by the intermediate of a SAMSON
a meaning only if all coefficients are constant. This can be
Intensity/Pressure converter while the muscle free tip can lift a
made from a simple physical interpretation of Int(x3): let us
given load. The muscle is made of a thin butyl rubber inner
assume that the system transitory state takes a time td to put the
tube surrounded by a rayon-braided weave. Its active initial
system from its initial position to the neighborhood of xd with length is about 28cm. In order to limit the non-linear effects
the ε-accuracy, and that this time is long enough in order that due to the I/P converter working pressure was restrained in the
its change resulting from the ε–vector is negligible. Let us note [0-3bar] range. At 3 bar we get a maximum force of about
Int(xd) this corresponding almost-constant value of Int(x3). It is 50dN and a maximum contraction length equal to about 11cm.
721
M= 0.2, 0.5, 1, 3, 5 kg
8 kg
10 kg
12 kg
15 kg
X-position (cm)
18 kg
20 kg
25 kg
Figure 5. Experimental set-up for open-and-closed loop position control of a Time (s)
pneumatic McKibben artificial muscle (see text).
(a)
We will not try to modelize the dynamic behaviour of
McKibben muscle – see, for example, our recent study [7] and 1 kg
associated references. However in order to justify our non-
linear friction model inspired by textile physics in the 10 kg
understanding of McKibben muscle damping, we developed a
X-position (cm)
simple comparison between our own specimen controlled in 15 kg
open-loop at 3 bar for various loads (Fig. 6.a) and the model
proposed in section 1 with following parameters: Fmax=50dN, 20 kg
xmax=0.11m, µs=4dN, µk_lim=40dN, x& k =0.6m/s, T=0.1s (a pure model
delay of 0.05s was also added) and m=50g. As it can be seen in real
Fig. 6.b although the model is not able to capture the full non-
linear character of the actuator, a relative good concordance
between the model and the real muscle is obtained with an
accuracy better than 10% all along the dynamic contraction.
In a second stage, we tested our single integrator-controller
Time (s)
sampled at 10ms. We show in Fig. 7 the experimental result for
step responses between x=2cm and x=9cm, corresponding to a (b)
large step, and for step responses between x=3cm and x=7cm Figure 6. Justification of a textile-physics inspired friction model for explaining
corresponding to a more limited step. In both cases a similar the McKibben artificial muscle damping by means of a comparison between
real muscle open-loop step responses performed at 3 bar (a) and section 1
value of kI has been chosen – equal to 0.0075 bar.m-1s-1. During muscle model including the non-linear kinetic friction model (b).
the large step ascending motion, the real position never
overshoots the desired one with a 95% response time between at the lower position and released it after some seconds. It can
about 0.9s and 1.9s (Fig. 7.a); during the middle-range be checked in Fig. 8.b that no instability occurred although a
ascending motion, overshooting of the desired position is brutal deviation from the desired trajectory, and that the real
limited to less than 5% with a 95% response time between 0.6s trajectory rapidly converges again on the desired one.
and 1.3s (Fig. 7.b). It is worthy to note that in both cases,
during the descending motion, due to the single effect mode of V. CONCLUSION
the artificial muscle – its obligatory positive control variable
corresponds to a simple contraction – the I-controller only To some extent we validated in our study the possibility to
becomes active after the desired position. An overshooting control non-linear artificial muscles thanks to a single integral
results however quickly reduced; the first peak could even be action. Our theoretical stability analysis is limited and must be
reduced with a smaller kI-gain however not adapted to a quick extended but it suggests, for a given muscle, roughly
response ‘against gravity’, suggesting the interest of some characterized by its maximum contraction length, its
adaptive approach with respect to motion sense as mentioned maximum isometric force and its zero-velocity friction
in conclusion. The tracking of a sinus-wave was also tested as coefficient, that a maximum kI-parameter can be
shown in Fig. 8.a with a constant load of 10kg moving from a experimentally determined. Simulations even suggest a larger
lower 2cm to an upper 9cm x-position alternatively with and stability domain. The results reported with a single McKibben
against gravity and a constant kI-gain equal to 0.015 bar.m-1s-1: artificial muscle show the practical relevance of the controller
the tracking is relatively good with a maximum delay of about both for step and tracking responses with a real robustness to
0.8s. In a last experiment, we tested the ability of the controller load variations as to manual external perturbations.
to reject a manual perturbation: during the sinus-wave tracking
the operator manually caught the load when it was passing
722
Real traj. for
desired M= 1,2,5,
sinus. traj. 10,12.5 kg
M= 1,2,3,
4,5,7,10
X-position (cm)
X-position (cm)
12,15 kg
M= 1,2,3,
4,5,7,10
12,15 kg
desired Real
sinus. traj. trajec tory
M= 1,2,3,
4,5,7,10
X-position (c m)
X-p osition (c m )
12,15 kg
M= 1,2,3,
4,5,7,10
12,15 kg Start of
manual
perturbation
723