Professional Documents
Culture Documents
Jha Prabhudutt
Mathematical programming formulation for nonlinear optimal robot control has been attempted using
geometric programming. The objective is to minimise the error subject to stability and torque constraints. In
particular, the condensation approach of geometric programming along with Persaoal’s theorem has been used
for modeling. The relationships of sampling time versus the proportional gain (K,), integral gain (K,), and the
derivative gain (K, I are studied. The control parameters (K,, K,, KiI decrease with time before it becomes
flat. The stabilityand torque constraints effect the control gains substantially and uniquely It is also concluded
that geometric programming is an eficient mathematical technique for the nonlinear control of robotic system.
0 1996 by Elsevier Science Inc.
control has the advantage of reaching the target point tional controller for a teleoperator. It required a position
smoothly. However, the torque is not controlled and it is sensor and a torque sensor on the input shaft as well as on
not suitable for tasks that require a constant torque or the output shaft. The resulting double-loop control system
force. The selected control approach is dependent on the yielded a complicated stability problem.
application of environment in which the robot arm oper- Many of these theories of optimal control lack applica-
ates. bility. The existing methods do not guarantee that the
The purpose of the controller is to compare the actual singular configuration will be avoided while optimising the
output of the plant with the input command and to error function. It may lead to large joint torques and joint
provide a control signal that will reduce the error to zero accelerations. It is also observed that motions with opti-
or as close to zero as possible. mal joint torques are usually associated with large joint
The PID controller can be represented by speeds and joint accelerations. To overcome this problem,
this article proposes the minimisation of angular position
under torque and stability constraint.
In this paper a mathematical model for optimal control
of a robot is developed. The solution of the optimal
and the transfer function is control model is carried out through a nonlinear program-
ming technique. The geometric program (GP) is a class of
nonlinear program. The papers14-16 show that geometric
(2) programs have certain intrinsic characteristics not found
in other optimisation techniques. GP can be effectively
PID control provides quick response, good control of used to resolve the difficulties of optimal control of a
system stability, and low steady-state error. Therefore prescribed end-effector motion and a prescribed initial
optimum control of the manipulator becomes very impor- manipulator configuration. The optimal control problems
tant and useful for accuracy and for improving perfor- inherently involve optimisation of dynamic systems. The
mance. The optimal control will allow us to determine the dynamic system solutions are computationally difficult and
system configuration or design that will optimize system involve repetitive solution of vector differential equations
performance with respect to selected criteria. during each step of a gradient-based search procedure.
During the last two decades several control methods For these reasons the user sometimes prefers to employ
have been developed for robotic manipulators.2-4 In liter- classical frequency response methods for the design of a
ature concerned with the dynamic control of manipula- single-input/single-output system.17 The main disadvan-
tors, the complexity of nonlinear dynamics is emphasised, tage of the classical approach is the lack of systematic
and various methods that compensate all nonlinear terms procedures for choosing compensators to the satisfaction
in dynamics in real time are developed to reduce the of real-world constraints. The modern design approach to
complexity of control systems.5m7 However, these methods dynamic compensation is based on optimisation and simu-
require a large amount of complicated calculation, so that lation using digital computers. The professional control
it is difficult to implement these methods with low-level engineer should be familiar with both classical and mod-
controllers such as microcomputers. For most industrial ern control modeling techniques.
robots, each joint of the manipulator is independently The main thrust of this paper is to derive an optimum
controlled by simple linear feedback.6 However conver- relation between system and compensator parameters. It
gence to a target position is difficult for a general nonlin- is demonstrated how geometric programming can offer
ear mechanical system. Totani et a1.7 discussed the motion the designer a clear insight into the optimum pole place-
of the transferred object itself, and a mathematical model ment problem and that it has the added advantages of
for the object motion is derived from an optimal control modest computational and data processing requirements.
point of view. Such advantages are also important for real-time imple-
To control the manipulator motions reasonably, opti- mentation of adaptive control in space and hazardous
mal control theory may be applied.8m’” Bohn’ used a waste disposal, an area of considerable interest.
process of condensation to develop a geometric program- Geometric programming has been shown to be an
ming approach to the necessary conditions for a con- effective design tool for a special class of nondynamic
strained optimum in an elementary way for trajectory system. Much effort has gone into overcoming the original
planning. Shin et al.” used the dynamic programming GP restrictions so that it would be applicable to a broader
technique for the general case where 1) the actuator class of systems. 18~20 The development of efficient com-
torque limits are dependent one another, 2) the cost putational algorithms has, however, lagged considerably
functions can have an arbitrary form, and 3) there is behind theoretical developments. GP applications at pres-
constraint on the jerk or derivative of the acceleration. ent are mainly in the area of nondynamic systems.
They have shown that the dynamic programming solution Straightforward extensions of quasistatic dynamical sys-
converges as the grid size decreases. tems have been made.20 Hence the aim of the paper is to
Thompsoni presented an improved procedure for de- develop a mathematical model for an optimal compen-
signing multivariable PID controllers for an unidentified sator, subject to a time-delay response constraint, and
plant. It is assumed that the open-loop plant is stable, and then solve it. By expressing the performance index and
responses to step inputs are basically nonoscillatory. constraint in posynomial form, it is shown that the GP
Hollerback13 suggested a design of a force-reflecting posi- approach results in simple explicit equations for the opti-
form=1,2,3 ,..., M
forX,>O,n=1,2 ,..., N
2. Mathematical programming technique
Most optimal control problems are inherently nonlinear in where N is the number of total variables.
nature. In real-world applications, only constrained non- The above GGP problem has a dual problem associ-
linear mathematical programming techniques such as the ated with it.*’
reduced gradient method, dynamic programming, and geo-
metric programming are found to be useful for the opti-
mal design of control systems. The constrained nonlinear Dual problem. Maximize
program is usually described in the following manner.
Minimize
(8)
f(x), x=[x,,x, )..., XJEP (31
2. I. Geomettic programming
(11)
Geometric programming is a recently developed tech-
nique for solving algebraic nonlinear functions subject to W,, 2 0 and definitions of Signum functions are
linear or nonlinear constraints. The structure of geometric
programming is more appealing to engineers. The design,
manufacturing, and control problems are more fully ex- um =fl, m=0,1,2 ,..., M; %t = fl,
plainable through this technique. Geometric programming
(GP) has special features that are noteworthy for investi- t= 1,2,..., T,,m=0,1,2 ,..., M (12)
gating the optimum design of robot control systems. Other
nonlinear programming techniques involving repetitive so- Once the dual variables W,, are known, the corre-
lutions do not produce a globally optimal solution for sponding primal values x, can be determined. The geo-
prescribed end-effort motion. There are many different metric programming algorithm described below is based
geometric programming algorithms. However because of on a constraint derivative concept. The algorithm gener-
its recent development, no standard software is readily ates the number of equations necessary to find unique
available. values of dual variables. The complex transfer function
A general geometric program (GGP) is formulated2” as will be represented by a matrix vector equation described
follows. later. The algorithm proceeds as follows:
1. Partition the dual variables vector W into two mutu-
ally exclusive sets y and z, i.e., W = (y,z). The vector y is
Primal problem. Minimize designed as a set of (N + 1) dependent or decision vari-
ables, and z is a set of (T-N - 1) independent or state
T0 N variables. Any (N + 1) among the T independent vari-
g,(x) = c a(,tC(J, n xn”“‘n (6) ables could be chosen as the decision variable. Let M = N
n=l
t=1 + 1.
:(T-N- 1
dh,/dz, dh,/dz, ... dh,/dz,-,
=
Ix(h4x
dhj;dz, dh,/dz,
1))
... dh,/dz,-,
(14)
3.1. Analysis of the optimal robot control problem
The degree of difficulty can be reduced either by reducing (s2K,K, +sK,K, + K,K,)Od,(s)
the total number of terms (T) or by increasing the num- O,(s) =
s3R,J,,+s2(R,f,,+K,K,+K,K,)
ber of primal variables (X). The reduction in the number +sK,K, + K,K,
of terms can be accomplished by partial condensation of
the signomial function g,(X). The condensation point X (21)
need not be explicitly specified. If need be, the signomial
function could be fully condensed to a monomial.‘8*‘9 After including the disturbance D(S) in the transfer
8. The number of nonlinear equations is now solved by function O,(s)/&), we get
parameter estimation to get the dual variables.
9. Then find the primal variables X:
O,(s) -snR,
-=
D(s) s~R,J,~~+s’(R,~,~~+K,K~+K,K,)
wo,oogo(x)~ t= 1,2,...,T, +sK,K, + K,K,
(171 (22)
Therefore torque is
From (21) and (23) and using the superposition princi- TV I Ki (32)
ple, we get
where Ki is a maximum torque for the i-joint.
(s*K,K, + sK,K, + K,K,)Od,(s) Consequently, a mathematical model for the robot
-snR,D(s) control including an objective function, stability con-
G,(s) = s3R J straints, and torque constraints is developed. By using the
a eff constrained GP optimisation technique, the optimal value
+s2(R,f,,+K,K,+K,K,)+sK,Kp+K,K, of x is obtained by minimising the error objective function
(performance index) under the real-world constraints of
(24)
stability and torque.
Let p = st,, where t, is a time-scaling design parame-
ter. Use of t, allows the characteristic equation for the 3.2. Objective finction
closed-loop system poles to be expressed in the nondimen-
sional form, For a step input of magnitude A, i.e., Of(s) =,4/s and
disturbance D(s) = B/s, the error function E(s) is devel-
(25) oped as follows:
d,= 1, t, =x4,
are tabulated for linear systems22 and can be easily con-
verted into a ratio of signomials and expressed in terms we get
of x.
We can obtain i-constraints for stability by using the c*s*+ CIS + cg
Routh-Hurwitz criterion: E(s) = (35)
d3s3 + d2s2 + d,s + d,
gi(x) > O, i=1,2,3 ,..., i (30) Now it is necessary to evaluate integrals of the form
-I
theorem converts I into the corresponding frequency do- x2x4 -x1 -‘x3x,220 (46)
main form as
x3xp 2 0 (47)
I= &Jr Hs)E(-s)ds...(g)
a
(37)
q(s) =s3 +x,x,‘? +x,x,% +x,x,3 (44) +m212ki2+ +m212fi,2sin e2 + +m,gl cos(0, + 0,)
SK, (52)
By using the Routh stability criterion” we get three
stability constraints as follows:
Now /3,, (b,), (e,), (O,), (i,), (6,) must be calculated.
Equation (24) is converted to a t-domain function O(t)
x,x,’ 2 0 (45)
from the S domain by Laplace inverse transformation.
e,(t) = N, + N2t + G,e(“l” + G2ecrzt) + G3ecS3’) The performance index based on the minimisation of
positional error under the stability and torque constraints
(62) is developed. The specifications of a two-link robot are
presented below.
e,<(t) = N2 + G,s,e(S?‘) + G3s3e(r3r) (63)
1
i,(t) = G,sfecSl’) + G2s~ecsz’) + G3s_:e(S’f) (64) J, = 12 X 10e4 kg-m-s2/rad,
a,==
0.274x,x,4 + 2520.66~~~;
I= (68) g,(x) = (0.274~~~;~ + 2520.661x,x~(x,x,x, -x,2)-’
x1x2x3 -xi
subject to constraints.
subject to constraints (451, (461, (471, (59), (60), and (61). In an optimal design an equality constraint can be
Two different optimal control solutions are presented replaced by two inequality constraints of 5 0 and 2 0.
by separating the mathematical model in two forms: From the formulation of the problem it is clear that the
1. The optimal control design in the environment of sta- 2 0 constraint is proper:
bility constraint.
2. The optimal control design in the environment of both gr(x> =x5” +x,x,‘x,2 +x,x,2x, +x,x,” 2 0
stability and torque constraints.
g,(x) =x,‘+x,x,‘x,2 +x;‘x,2x, +x,x,3 2 0
This separation demonstrates the degree of difficulty
associated with numerical optimisation of control prob- g3(x) =x; +x,x,,x; +x,x,2x, +x;‘x,3 r 0
lem. It was discovered that convergence is easier for the
stability constraint. To solve the above problems by geo- g4(x) =x2 -x,x3 2 0
I 1 constraint. It is also noted that the standard geomet- uO’wOl + ul’w,’ + u31w31 + a,,w41 - u42w42 = o
ric programming format requires only the I 1 constraint. (73)
g,(x) =x*-x,x3 I1 There are seven dual variables, and so seven equations
are required for a unique solution. As described earlier,
g,(x) = O.O1389(G’x; + G,x,2 + G,x;) the seventh equation is generated by the algorithm pre-
+ O.O208(G’x, + G,x, + G,x,)x, sented in Section 2. The solutions obtained are
+ .8167x’, I 1
W,, = 0.016, W,, = 0.9833, W” = 1.99,
g,(x) = x;‘x; ‘x3 +x; ‘x; ‘x3 ‘xs I 1
w,, = 0.001 w,, = 0.01, W,, = 0.08,
where
w,, = 2.1, w,, = 0.001
difficulty. The dual geometric program (DGP) of Problem ahlW”l + ~l,Wl, + a24w2.4 - %Ws4 - u42w,2
2 is presented below.
+ %lw,l - obzw,, = 0 (80)
Maximise
- 4aolKl + 5u02W02 - u12W12 - 2ulFl,
(82)
(83)
(84)
-
~O,W”l - ~02Wo2 + %2W62 = 0 (85)
Table 2. Results under stability and torque constraints elusion drawn earlier when only the stability constraint is
t present. The graph of sampling time versus K,, K,, and
KP K” K;
Ki under stability and torque constraints is shown in
0.3054 0.7958 0.0755 0.048 Figure 2. Both graphs confirm the conclusion drawn ear-
0.4045 0.4507 0.04 0.02 lier that optimal gain values decrease as the sampling time
1.9980 0.0189 0.0063 0.0035 increases.
2.93791 0.0094 0.00724 0.008973
The effect on the frequency of sampling is quite differ-
4.98529 0.00378 0.008 0.009575
ent. The system under the stability constraint could work
5.98221 0.00283 0.007927 0.0045
6.97935 0.0023 0.00798 0.002869 for a frequency of sampling as low as 0.001 set, but under
9.97331 0.002 0.008044 0.00124 both the stability and torque constraints the solution
12.014 0.0014 0.0081 0.00110 could not be obtained below a 0.30.54-set time sampling.
15.0458 0.00108 0.00811 0.001 After closely examining the lowest value of sampling time
20.0821 0.0010 0.00814 0.00114 as applied to Problems 1 and 2, two types of conclusions
23.1 19 0.001 0.00815 0.001072 could be drawn. The first conclusion is related to numeri-
cal optimisation, which means that in the case of Problem
2 the convergence could not be achieved below some
equality constraints, analytical closed-form solutions could initial value. The convergence as well as the feasible
not be obtained. Geometric programming has been used results both are affected by initial values.
to design the control system parameters iteratively. The other conclusion regarding the minimum value of
The results are presented for two different forms of sampling time could be that a complex mechanical system
problems. The results of Problem 1 for the minimisation like a robot could not be driven as fast as we want. This
of error function under inequality stability constraints are could happen for many reasons, like inertia, backlash,
presented in Table I. The results show various solutions gravity, change in compliance, and deflection of the ma-
for different sampling times. Under close examination of nipulator arm. In must be stated clearly that solutions may
the results it is apparent that as the sampling time in- not always exist for all models. Sometimes numerous
creases the control values decrease. This makes sense. analyses may be necessary to determine the effect of
When the sampling time is large the system dynamics parameter modification in nonlinear control. However the
allow the robot to adjust to stability, and a large gain is overall conclusion that the servo values proportional gain
not required. And when sampling the time is low, the (K,), integral gain (K;), and derivative gain (K,) decrease
gains are high, bringing the system to stability. Results are as the sampling time increases holds true analytically as
plotted in Figure 1. well as physically.
Similarly, when stability and torque constraints are In an on-line control the time of computation of the
applied together, results show the consistency of the con- control gains are also important. In the case of Problem 1
0 5 IO 15 20 25
SAMPLING TIME
Figure 1. Sampling time versus control gain values under stability constraint.
10.’
5 12 22
the time of computation is 4.97 set, and in the case of viscous-friction coefficient of the motor referred
Problem 2 the time of computation is 114.68 sec. The to the load shaft (kg *m . s/rad)
computation was carried on a VAX-5380 Open-VMS sys- fin Viscous-friction coefficient of the motor referred
tem. It is obvious that for Problem 2 the control algorithm to the motor shaft (kg. m . s/radl
would take more time, as the optimisation model has F(W,,) dual objective function
many constraints. Hence Problem 1 could be easily imple- gjCx) inequality constraint
mented for on-line control. h,(x) equality constraint
The numerical algorithm could be further improved to J Jacobian matrix
lower the time of computation. It was observed that the Jeff effective moment of inertia of the combined
number of iterations for convergence in Problem 1 is
motor and load referred to the motor shaft
much lower than the number of iterations required for
JL moment of inertia of the motor referred to the
convergence in Problem 2. However these compensator
load shaft (kg. m . s’/rad)
values can be used to adjust the effect of controllers to
J,
moment of inertia of the motor referred to the
achieve an overall desired system behavior. Authors want
motor shaft (kg. m .s2/radl
to experiment with different numerical optimisation tech-
The motor-torque proportional constant (kg.
niques on these two nonlinear optimal control problems
and compare computation time and convergence. It is m/A)
observed that those nonlinear programming techniques, Ktl proportional constant (V . s/radl
which iterate the solution by linearising the model, may Ki error integral feedback gain
not be suitable for highly nonlinear control problems; KP
error proportional feedback gain
however the approach presented in this paper should be K’ error derivative feedback gain
helpful for further research. L, armature inductance (Hem-y)
n gear ratio
R real number
R, armature resistance (ohms)
Nomenclature t, time-scaling parameter
A step input L number of terms in the mth constraints (m =
C rnf the coefficient of the tth term of the mth con- l,...,M)
straint T0 number of terms in the objective function
D(s) disturbance w,, dual variable of the tth term of the mth con-
angular error straint
effective viscous friction coefficient of the com- X design variable
bined motor and load referred to the motor Y state variable
shaft z decision variable