Vous êtes sur la page 1sur 5

Robust Gamma-Stable Controller Design for a Two-Degree-of-Freedom Robot

Arm

László Lemmer Bálint Kiss


Department of Control Engineering Department of Control Engineering
and Information Technology and Information Technology
Budapest University of Technology and Economics Budapest University of Technology and Economics
Magyar Tudósok krt. 2, H-1117, Budapest Magyar Tudósok krt. 2, H-1117, Budapest
Hungary Hungary
lemmer@seeger.iit.bme.hu bkiss@iit.bme.hu

Abstract – This paper presents the robustness analysis of two-degree-of-freedom robot arm with an uncertain load
a two-degree-of-freedom robot arm based on the notion mass and moment of inertia. Section III presents the design
of Gamma-stability. The model of the robot arm is given of the aforementioned conventional PD like compensation.
first, and the model parameters, such as load mass and
inertia, which have to be considered as uncertain parameters In Section IV, the Γ-stability analysis of the feedback
are enumerated. A simple linear controller is designed for control system is discussed. Section V applies the method
the nominal parameters to provide position control without to a numerical example where the Γ-stability boundaries
overshoot and the Gamma-stability of the closed loop system are projected into the parameter space.
is studied.

I INTRODUCTION II DYNAMICAL MODEL


The accurate and robust control of the free motion of a
Consider the schematic representation of the two-degree-
robot arm is of practical interest in many robotic applica-
of-freedom robot arm in Fig. 1. The motion of the robot is
tions, e.g. for transfer of workpieces in a factory. The design
restricted in the vertical plane. We assume that the center of
of a robot arm controller is a robustness problem since
gravity of the load is located on the mid-plane of the gripper
tracking error has to be kept small even for large variations
and on the line determined by the joint axis of Segment 2
in the moment of inertia of the load, in its mass, and in the
and its center of gravity. The notation of the parameters are
location of its center of gravity. Moreover, the equations
summarized in Table I.
describing the dynamics of the robot contain nonlinearities
and often significant effects have to be neglected to give a TABLE I
linear approximation that makes it possible to apply linear NOTATION OF THE PARAMETERS, i = 1, 2
techniques in controller synthesis. Robustness analysis has
to be carried out to cope with the model uncertainties due Ji moment of inertia of segment i
mi mass of segment i
to the parameter uncertainties and simplifications. li length of segment i
To control the free motion of a robot, an adequate method lci distance between the center of gravity and the joint
is position control in joint coordinates. In this paper, we axis of the segment i
qi angular displacement of joint i
present a Γ-stability analysis of a closed loop position JL moment of inertia of the load
control for each joint. To obtain the closed loop, we use mL mass of the load
a PD like compensation technique based on the simplified
linear dynamical model of the two-degree-of-freedom robot
To obtain the equations of motion, we use the Lagrange
arm with load. Note that a linear parameter varying (LPV)
function [2], [3]:
model would be also reasonable.
Polynomial dependency on the parameters of the coef- d ∂L ∂L
ficients of the characteristic polynomial of a closed loop − = τi for i = 1, 2 (1)
dt ∂ q̇i ∂qi
control system can be handled by constructing stability
and performance regions in the parameter space [1]. This Here, qi and τi (i = 1, 2) are the generalized coordinate
robustness analysis is called Γ-stability analysis and it and torque at the joint i, respectively. L stands for the
allows us to conclude about the (performance) robustness Lagrangian which is the difference of the kinetic and the
of the position control with respect to specified parameter potential energy.
changes such that not only the stability of the closed We transform the obtained equations of motion into a
loop system can be studied for different operating domains form more familiar in robotics which reads in our case
in the parameter space, but the satisfaction of additional
· ¸µ ¶ µ ¶ µ ¶
constraints on the location of the poles determining a so- H11 H12 q̈1 h1 τ
called Γ-stable region of the complex plane. + = 1 , (2)
H21 H22 q̈2 h2 τ2
The remaining part of the paper is organized as follows.
In Section II, we consider the mathematical model of the where the elements of the symmetric inertia matrix H read
way to control the motion of the robot (with the controlled
variables q1 and q2 ) for given joint angle references is
to consider the segments separately and use independent,
decentralized position control algorithms to both of them.
Some nonlinear terms, where the parameters are known,
will be fed forward, other terms will be reduced to the
input and will be considered as disturbances.
A Decentralized Linear Model
We suppose a dc-motor to actuate the joint but with a
small electrical time constant with respect to the mechanical
time constant, so we can perform a crude approximation by
neglecting the armature inductance. We introduce Jmot,i ,
bmot,i , ϕmot,i and Kmot,i for the rotor inertia of the motor
i, the viscous friction coefficient of motor i, the angular
Figure 1
displacement of the motor shaft i, and the gain of the ith
Two-degree-of-freedom robot arm moving in a vertical plane
amplifier–motor system, respectively. Since the motor shaft
is coupled to the joint through a gear train with a gear ratio
Gi , ϕmot,i = Gi qi , the total inertia and friction coefficient
reduced to the joint are
Ji,eff = Hii + G2i Jmot,i (5a)
2
H11 = J1 + J2 + JL + m1 lc1 + (3a) bi = bsi + G2i bmot,i (5b)
¡2 2
¢
+ m2 l1 + lc2 + 2l1 lc2 cos q2 +
¡ ¢ where Hii is defined as in (3) and bsi stands for the viscous
+ mL l12 + l22 + 2l1 l2 cos q2 friction coefficient of joint i. So the differential equation of
H12 = J2 + JL + m2 lc2 (l1 cos q2 + lc2 ) + (3b) the system can be written as
+ mL l2 (l1 cos q2 + l2 ) Ji,eff q¨i + bi q˙i = Gi Kmot,i ui − uf i − di for i = 1, 2 (6)
H21 = H12 (3c)
2
where uf i contains the terms which can be feed-forwarded
H22 = J2 + JL + m2 lc2 + mL l22 . (3d) (up to a gain Gi Kmot,i ) and di contains all cross-coupling
The terms h1 and h2 containing the gravity effect and the terms, and the terms containing uncertain parameters im-
terms being quadratic with respect to the joint velocities peding their feed-forward
can be also detailed uf 1 = g{m1 lc1 cos q1 + m2 l1 cos q1 (7)
h1 = g{m1 lc1 cos q1 + (4a) + m2 lc2 cos(q1 + q2 )} − 2l1 m2 lc2 sin q2 q̇1 q̇2
+ m2 [l1 cos q1 + lc2 cos(q1 + q2 )]+ uf 2 = gm2 lc2 cos(q1 + q2 ) + l1 m2 lc2 sin q2 q̇12 (8)
+ mL [l1 cos q1 + l2 cos(q1 + q2 )]}− di = Hij q̈j + hi − uf i for j 6= i (9)
− 2l1 (m2 lc2 + mL l2 ) sin q2 q̇1 q̇2 − The cos q2 term in H11 will be considered as a parameter
− mL l1 l2 sin q2 q̇22 uncertainty.
h2 = g[(m2 lc2 + mL l2 ) cos(q1 + q2 )]+ (4b) The transfer function between Qi (s) and Ui (s) is ob-
tained by taking the Laplace transform of both sides of (6)
+ l1 (m2 lc2 + mL l2 ) sin q2 q̇12 , with zero initial conditions:
where g is the gravitational acceleration. Qi (s) Gi Kmi
= (10)
Let us distinguish two sets of model parameters. This Ui (s) s (Ji,eff s + bi )
classification is the basis of the later Γ-stability analysis The inertia Hii is parameter and joint coordinate varying
and the definition of the space of the uncertain parameters. in a nonlinear way but in order to apply linear controller
1) Parameters belonging to robot arm segments (J1 , J2 , design techniques, a linear model has to be considered.
m1 , m2 , l1 , l2 , lc1 , lc2 ) are determined by the well Therefore we use a constant nominal value H̄ii for Hii
known geometry and material properties of the robot and similarly
itself.
2) Other parameters (JL , mL ) come from the properties J¯i,eff = H̄ii + G2i Jmot,i . (11)
and geometry of the load so they can vary in relatively B Compensator Design
wide ranges and hence considered to be the uncertain
The plant transfer function (10) is an integrator plus a
parameters.
first order term. By the feedback of such a system modified
III CONTROLLER DESIGN with an adequate preamplifier gain, we obtain a second-
The objective of the closed loop control is to have the order closed loop system.
2
joint variables, q1 and q2 track some references, denoted ωni
W (s) = (12)
by q1r and q2r , respectively. The simplest and common 2
s2 + 2ξωni s + ωni
control [4]: Leave a slow system slow and a fast system
q- ei
ir e- ui Gi Kmot,i q̇i- 1 qi- fast. Hence we may choose
Kpi - e-
+6 − + 6− Ji,eff s + bi s bi
ωni = ¯ , (17)
Ji,eff
Kti ¾ where the nominal value of the effective inertia, namely
J¯i,eff are used again.
IV Γ-STABILITY ANALYSIS
For a given characteristic polynomial whose coefficients
Figure 2
depend polynomially on uncertain parameters, we might
Block diagram of the compensated closed loop system
determine whether all the roots of the polynomial are
located in a prescribed region Γ in the complex plane for
all admissible parameter values.
Mapping the boundary ∂Γ of Γ into the parameter space
The closed loop natural undamped frequency ωni and the allows determining the subset of the parameter values for
damping factor ξi determine where one has to move the which the closed loop poles remain in the Γ domain and
nonzero pole of the plant with the controller. Similarly, ωni hence satisfy the prescribed closed loop dynamical behav-
and ξi gives also the way to set the loop gain. ior. Recall however, that Γ-stability analysis cannot say
The replacement of the nonzero pole has the physical anything about the uncertainties that do not appear in the
interpretation of modification the damping in the system. characteristic polynomial coefficients. Consequently, it does
It can be achieved by a tachometer (i.e. velocity) based not take into account the disturbance rejection. Usually,
feedback with gain Kt . The loop gain can be changed these disturbances can be dealt with the introduction of
by a preamplifier gain Kp . Fig. 2 shows the analytical a feed forward term with the help of commonly used
block diagram of the model with this compensation for estimation techniques (e.g. Luenberger observer or Kalman
joint i, where the compensation gains are also denoted with filter with load change).
a subscript i. In the case of this PD type linear controller, In the closed-loop transfer function, the uncertain param-
the control signal is calculated using eters enumerated in Section II affect the inertia parameter
Ji,eff , which is both parameter and joint variable dependent.
Recall that we used a constant nominal value J¯i,eff to
ui = Kpi (qir − qi ) − Kti q̇i . (13)
determine the coefficients Kti and Kpi . Using (14) and
The closed-loop transfer function of the compensated plugging in the expression of Kpi and Kti from (16), we
system reads obtain the characteristic polynomial

Gi Kmot,i Kpi
Pi (s) = Ji,eff s2 + J¯i,eff 2ξi ωni s + J¯i,eff ωni
2
. (18)
Ji,eff
W (s) = (14) As we have already stressed out, one wants to avoid
bi +Gi Kmot,i Kti G K Kpi
s2 + Ji,eff s + i Jmot,i
i,eff
overshoots. This condition is satisfied if all roots of the
characteristic polynomial are real, i.e. the left half of the
We can read out from (12) and (14) the following two real axis is the robust Γ-region. The roots of the character-
identities: istic polynomial (18) are real if the discriminant
¡ ¢
2 Gi Kmot,i Kpi D = 4J¯i,eff ωni
2
J¯i,eff ξi2 − Ji,eff (19)
ωni = (15a)
Ji,eff is not negative.
bi + Gi Kmot,i Kti As all quantities in (19) are nonnegative, we obtain the
2ξωni = . (15b)
Ji,eff inequality
Ji,eff ≤ J¯i,eff ξi2 (20)
From (15), the controller parameters Kpi and Kti can be
expressed as which explains how the choice of ξi influences the Γ-
robustness.
J¯i,eff ωni
2 Since the real roots must be negative, one has another
Kpi = (16a) constraint
Gi Kmot,i √
2ξi ωn J¯i,eff − bi D < J¯i,eff 2ξi ωi . (21)
Kti = , (16b)
Gi Kmot,i Constraint (21) is always satisfied for physically reason-
able values for which (20) holds also true.
where the nominal values of the effective inertia, namely With the expressions given by (5a), (11), and (3), the
J¯i,eff are used. characteristic polynomial can be rewritten as
For satisfactory choice of the design parameter ξi , we
have to consider that in robotics one usually wants to Pi (s, r) = Ji,eff (r)s2 + J¯i,eff 2ξi ωni s + J¯i,eff ωni
2
. (22)
avoid overshoots. For fast transients, the natural undamped We see from this equation that only Ji,eff depends
frequency ωni should be chosen high but the bounded on the vector of the real uncertain parameters r =
actuating signal limits the closed loop bandwidth. We will (JL , mL , cos q2 ) ∈ P . In the succeeding subsections we
choose ωni according to the basic rule of thumb of robust project constraint (20) to the parameter space P .
A Parameter Space Approach at Segment 2 The cosine function is bounded, therefore |r3 | ≤ 1.
The effective inertia of Segment 2 depends on the uncer- For all r3 ∈ [−1, 1], F (r1 , r2 , r3 ) = 0 determines a line
tain parameters JL and mL . We choose the nominal value boundary in the (r1 , r2 ) plane in the parameter space P as
as the sum of the terms that do not contain these parameters: in the previous discussion for the case of Segment 2. Using
the fact that
J¯2,eff = J2 + m2 lc2
2
+ G22 Jmot,2 (23) ∂
F (r1 , r2 , r3 ) ≡ 0 ⇒ F (r1 , r2 , r3 ) ≡ 0, (32)
Than the real effective inertia reads ∂r3
it is easy to show that for all r3 , the section of these
J2,eff = J¯2,eff + JL + mL l22 . (24)
boundaries is the unique point
It can be seen that the the only uncertain parameters ¡ ¡ ¢ ¡ ¢ ¢
(r1 , r2 ) = J¯1,eff ξ12 − 1 − l12 + l22 m2 , m2 . (33)
involved in the Γ-robustness study here are (r1 , r2 ) =
(JL , mL ), hence one can obtain In Equation (31), the coefficient of r1 is constant and
the coefficient of r2 is linear in r3 . Accordingly, the strict
J2,eff = J¯2,eff + r1 + l22 r2 . (25)
boundaries in the (r1 , r2 ) plane introduced by (30) are the
Let us apply now constraint (20) to (25) lines with r3 = ±1.
¡ ¢ For the choice of ξ1 , similar remarks can be made as
r1 + l22 r2 ≤ J¯2,eff ξ22 − 1 (26) for the choice of ξ2 , since the Γ-stability region in the
The stability boundary in the parameter subspace parameter space P can be again enlarged if the damping
(r1 , r2 )⊂ P is given by the line determined in (26) if we factor is increased as shown in Fig. 4 for the numerical
consider it as an equation, see Fig. 3 for the numerical values presented in the next section.
values presented in Section V. The inequality can be V NUMERICAL STUDY
interpreted as a limitation to the moment of inertia of
the load reduced to the second joint. Note that with the To present the method graphically in an example, we take
choice ξ2 = 1, undamped behavior is assured only for the physical parameters of the second and third segments of
JL = mL = 0, which is the nominal case without load. a PUMA 560 robot arm according to [5] and the references
If we want the system to endure heavier loads without therein. For completeness, the reported values are also given
overshoot, we can modify the deign parameter ξ2 but it here in Table II. We examine the effects of changing the
may result an even more overdamped behavior in closed tuning parameters ξi on the boundaries of the Γ-stable
loop. region in the parameter space P .
TABLE II
B Parameter Space Approach at Segment 1
PHYSICAL PARAMETER VALUES
The effective inertia of Segment 1 depends also on the
uncertain parameter cos q2 in addition to JL and mL . The Parameter Value Unit
nominal value of the effective inertia is the sum of the J1 0.539 kg · m2
Jm1 200.000 kg · mm2
constant terms without the uncertain parameters m1 17.400 kg
¡ ¢ b1 8.000 N ms/rad
J¯1,eff = J1 +J2 +m1 lc1
2
+m2 l12 + lc2
2
+G21 Jmot,1 . (27) c1 12.000 Nm
lc1 0.068 m
The real effective inertia (including the terms which involve l1 0.432 m
uncertain parameters) of Segment 1 reads K1 0.219 1
G1 107.360 1
J1,eff = J¯1,eff + 2m2 l1 lc2 cos q2 + J2 0.086 kg · m2
(28) Jm2 200.000 kg · mm2
+ JL + mL (l12 + l22 + 2l1 lc2 cos q2 ). m2 4.800 kg
b2 4.000 N ms/rad
Observe that q2 occurs only as the argument of the cos c2 6.000 Nm
function. Here, all elements of the parameter space P are lc2 0.070 m
involved r = (r1 , r2 , r3 ) = (JL , mL , cos q2 ). So we can l2 0.432 m
K2 0.202 1
write G2 53.690 1
J1,eff = J¯1,eff + r1 + (l12 + l22 )r2 +
(29)
+ 2m2 l1 lc2 r3 + 2l1 lc2 r2 r3 . Fig. 3 shows the root boundaries of constraint (26) for
Here, the coefficient J1,eff is no longer affine with respect three different values of ξ2 . Note that for ξ2 = 1.0 the
to the uncertain parameters due to the term 2l1 lc2 r2 r3 . Let closed loop system of Segment 2 is robustly Γ-stable only
us apply constraint (20) again to (29) in order to obtain an without load. Higher ξ2 enlarges the Γ-stable operating
inequality of the form domain in the parameter space P . Fig. 4 represents the root
boundaries of constraint (30). Here, two lines belonging to
F (r1 , r2 , r3 ) ≤ 0, (30) the extrema (±1) of r3 build up together the root boundary
for one value for ξ1 in the (r1 , r2 ) plane of the parameter
where
¡ ¢ space P . Observe that at some values of r3 = cos q2 ,
F (r1 , r2 , r3 ) = J¯1,eff 1 − ξ12 + 2m2 l1 lc2 r3 + the closed loop system may not be robustly Γ-stable for
¡ ¢ (31)
+ r1 + l12 + l22 + 2l1 lc2 r3 r2 . ξ1 = 1.0. However, ξ1 = 1.5 enables a large enough
12 20
ξ2 = 1.0
10 ξ2 = 1.5
ξ2 = 2.0 15
8

6
10
r2 = mL [kg]

r2 = mL [kg]
4

5
2

0
0
−2

−4 −5
−1 −0.5 0 0.5 1 1.5 2 2.5 3 −2 0 2 4 6 8
r1 = JL [kg m2] r1 = JL [kg m2]

Figure 3 Figure 5
Root boundaries for Segment 2 in the (r1 , r2 ) plane of the parameter The two-degree-of-freedom robot arm is robustly Γ-stable in the hatched
space operating domain for ξ1 = 1.5 and ξ2 = 2.0

40 the parameter space can be made. We represented the


ξ1 = 1.0
35
ξ1 = 1.5
boundaries in the parameter space using the obtained con-
30 ξ1 = 2.0 straints to project the Γ region from the complex plain to
the parameter space. With this representation technique, it
25
became easier to chose the admissible values for the tuning
20 parameters ξi (i = 1, 2) in the position control loops of
r2 = mL [kg]

15
a two-degree-of-freedom robot arm. The method can be
generalized to structures having more degrees of freedom
10
and with considerations to other uncertain parameters. A
5 limitation is that only two dimensional graphical represen-
0 tation is practical.
−5 VII ACKNOWLEDGEMENT
−10 This research was partially founded by the Hungarian
−5 0 5 10 15 20
r1 = JL [kg m2] Science Fund under grant OTKA T 042634 and by the
Advanced Vehicles and Vehicle Control Knowledge Center
Figure 4 under grant RET 04/2004.
Root boundaries for Segment 1 in the (r1 , r2 ) plane of the parameter
VIII REFERENCES
space
[1] J. Ackermann, D. Kaesbauer, and R. Muench, “Robust
gamma-stability analysis in a plant parameter space,”
Automatica, vol. 27, no. 1, pp. 75–85, 1991.
domain for the load inertia and for the load mass for all [2] G. Béda and A. Bezák, Kinematika és dinamika. Bu-
admissible values of r3 . As for the Segment 2, one may dapest: Műegyetemi Kiadó, 1999.
choose ξ2 = 2.0. With these choices of ξi (i = 1, 2) [3] B. Lantos, J. Somló, and P. T. Cat, Advanced Robot
the resultant robustly Γ-stable operating domain, which is Control. Budapest: Akadémiai Kiadó, 1997.
the intersection of the Γ-stable operating domains for the [4] J. Ackermann, A. Bartlet, D. Kaesbauer, W. Sienel, and
separated segments, is shown in Fig. 5. R. Steinhauser, Robust Control: Analysis and Design
of Linear Control Systems with Uncertain Physical
VI CONCLUSION Parameters. London: Springer, 1993.
[5] P. Corke and B. Armstrong-Hélouvry, “A search for
In this work we presented a robotic application where Γ-
consensus among model parameters reported for the
stability analysis techniques were applied to prove special
PUMA 560 robot,” in Proc. IEEE Int. Conf. Robotics
dynamic behavior properties in a changing environment.
and Automation, San Diego, May 1994, pp. 1608–1613.
The method is based on the analysis of the locations of
the roots of the characteristic polynomial of the closed
loop system. We have shown how a compromise between
fast transients and broad Γ-stable operating domain in

Vous aimerez peut-être aussi