Professional Documents
Culture Documents
Abstract . A new adaptive control scheme for direct control of manipulator end-effector to achieve
trajectory tracking in Cartesian space is developed in this paper. The control structure is obtained
from lin ear multi variable theory and is com posed of simple feedforward and feedb ack controllers a nd
an auxiliary input. The direct adaptation laws are derived from moclel reference adapt ive control
theory and are not based on parameter estimation of the robot model. The utili zation of adapti ve
feed forward control and the inclusion of auxiliary input are novel features of the present scheme
and result in improved dynamic performance over existing adaptive control schemes, The adaptive
controller does not require the complex mathematical model of the robot dynamics or any knowledge
of th e robot parameters or the payload; and is computationally fast for on-line impleme ntation with
high sampling rates,
Keywords. Adaptive control; feedback ; feedforward; Lyapunov methods ; multivari able control
systems; position control; robots; tracking systems.
289
290 H. Seraji
X(t) = A(~) (2) The role of the ru:n feedback controller K($) is to provide
a stable closed-loop system with poles at desired locations in
the complex plane. The feedback controller also ensures that
where the ru:l vector A(~) is a nonlinear function which em-
any error between 3<.(t) and r(t) is reduced to zero asymptoti-
bodies the manipulator geometry. Equations (1)- (2) consti-
cally.
tute the nonlinear mathematical model of the end- effector
dynamics. For the purpose of stabilization and pole placement, it is
sufficient to apply feedback from the position vector 3<.(t) and
Suppose that the initial condition of the end-effector in the velocity vector ~(t); since this corresponds to full state
the Cartesian space is denoted by Xo and corresponds to the feedback for the linear model (7) . Hence, an appropriate choice
joint angle vector ~o and the joint torque vector 'L0 . Let for the feedback controller is
the operating point representing this initial condition of the
manipulator be denoted by P. Now, suppose that the joint (9)
torque vector is perturbed slightly by 1:(t), that is 'L(t) =
'L 0 +1:(t); and let the resulting perturbation in the joint angle where Kp and K. are constant ru:n position and velocity feed-
vector and the end-effector position vector be ~(t) and 3<.(t) back gain matrices respectively. The feedback control law is
respectively, i.e. ~(t) = ~o + ~(t) and X(t) = Xo + 3<.(t). In therefore given by
a recent study [20], it is shown that ~(t) is related to 1:(t) by
the linear differential equation
Equation (3) is obtained by linearizing the nonlinear equation L1 (t) = Kp~(t) + K.§.(t) (11)
(1) about the operating point P; and -4, B, and 6 are constant
ru:n matrices which depend on P. Now, consider the differen- in the time-domain, where ~(t) = r(t) - 3<.(t) and ~(t) = i(t)-
tial transformation from the joint space {~} to the Cartesian ~(t) are the nxl position and velocity tracking-error vectors
space {X}. The joint angle perturbation ~(t) is related to the respectively. Note that no differentiation is involved in imple-
end-effector perturbation 3<.(t) by [22-23] menting the feedback controller since i( t) and ~(t) [= Jo~( t)]
are directly available; and hence K($) is realizable.
(4)
2. 2 FeedCorward Controller
where J o = J(S;l0) = [1l~~;U]p is the ru:n Jacobian matrix
The function of the ru:n feedforward controller Q ($) is
evaluated at the operating point P. Assuming P is not a sin- to cause the position vector 3<.(t) track the reference trajectory
gular point of the Jacobian matrix, equation (4) can be solved vector 1:(t) . From Refs. [24-25], the feedforward controller
for ~ as Q(s) is chosen as the minimal-order inverse of the end-effector
(5) model (8); that is
Now, let !:(t) denote the ru:l "incremental reference tra- 1:($) = J~ [[1($) + [2($)]
jectory vector" in Cartesian space, that is, the desired value (15)
of the end-effector position vector 3<.(t) . Then, in order for the = J~ [Kp + K.$]~($) + J~ [c + B$ + A$2] r($)
position vector 3<.( t) to track the reference trajectory vector
r( t), the control system must have two independent controllers in the frequency-domain; or by
[24] as shown in Figure 1; namely the feedback controller K($)
and the feedforward controller Q ($ ). Each controller is now 1:(t) = J~ [Kp~(t) + K.§.(t) + Cr(t) + Bi(t) + Ai(t)] (16)
Cartl'si;1I1 Control oj Robotic \Llniplll ;ltors 291
in the time-domain. On applying the control law (16) to the adaptation ensures that the robot control system is tuned on-
end-effector model (7). we obtain line such that a good dynamic performance is achieved despite
coefficient variations due to changes of the robot geometric
configuration. speed of motion or the payload.
Noting that ~ = f. - i;. and £: = t -~. we obtain the (2n)th order Consider the nonlinear model of the end-effector dynam-
error differential equation ics written as [1 .2.12]
Equation (17) describes the dynamic behavior of the position where A". B". and C" are nxn matrices whose elements are
tracking-error ~(t) in terms of the feedback gains Kp and K •. complex nonlinear functions of X.K. and the payload. Let us
It must be noted that by choosing the feedforward controller now consider the linear time-varying robot control law
as the minimal inverse of the robot. we have ensured that the
error differential equation (17) is homogeneous; i.e .• the right- £:(t) =£.(t) + [Kp(t)g:(t) + K.(t)t(t)]
hand side of (17) is zero. The solution of equation (17) can be (23)
expressed as + [C(t)K(t) + B(t)E.(t) + A(t)E(t)]
where g:(t) = K(t) - X(t) is the nx1 "total" position tracking-
~(t) = [:f: G;e",p('1;t)] ~(O) + [:f: H;e",p('1;t)] ~(O) (18) error vector. Equation (23) is a generalization of the "total"
,:;;::1 1::;:1 control law (21) developed in Section 2 based on the incremen-
tal analysis . As in equation (21). this control law is composed
where ~(O) and ~(O) are the initial position and velocity errors. of three components; namely the "auxiliary input" £.(t). the
G; and H; are some constant nxn matrices. and '1; is a root of
the (2n)th order error characteristic polynomial [24]
feedback term [Kp(t)g:(t) + K.(t)t(t)]. and the feedforward
.le(s) = IAs2 + (B + K.)s + (C + Kp)1 (19) term [C(t)K(t) + B(t)R(t) + A(t)E(t)] . The time-varying sig-
nal £.(t) corresponds to the operating point term £:" in equa-
tion (21) and will be synthesized by the adaptive scheme. On
It is noted that the feedforward controller Q(s) does not
applying the total control law (23) to the nonlinear end-effector
affect the error characteristic polynomial .le( s) and hence the
model (22) as shown in Figure 3. we obtain the closed-loop
dynamics of the tracking-error ~(t) is independent of Q(s) .
Since the gain matrices K p and K. can be chosen such that model
'11 •.. ·.'12n have negative real parts [24]. the tracking-error vec-
tor ~(t) in equation (18) will tend to zero asymptotically; i.e .• A"X(t) + B"K(t) + C"X(t) = £(t) + Kpg:(t) + K.t(t)
~(t) -> Q or ~(t) -> 1:(t) as t -> 00 for all ~(O) and ~(O). In + CK(t) + BR(t) + AE(t)
fact. by placing the roots '11 •.. .• '72n at desired locations. the (24)
transient behavior of the tracking-error can be shaped at the This equation can be written in terms of the total tracking-
designer's discretion. error g:( t) as
3. Direct Adaptive Control of the End-Effector
Motion A"t(t) + (B" + K.)t(t) + (C" + Kp)g:(t) = -£.(t)
(25)
+ (A" - A)E(t)+ (B" - B)R(t) + (C" - C)K(t)
In Section 2. the structure of multi variable controllers to
control the incremental motion of the end-effector was devel- It is seen that 1£( t) and £.( t) appear as forcing functions on the
oped. This control scheme works well for small movements of right-hand side of the error differential equation (25) . There-
the end- effector about the nominal operating point P. The fore. if the gains of the feedforward controller Q(s) and the
implementation of the incremental control law on the robot auxiliary input £.(t) are fixed. the solution of equation (25)
manipulator is shown in Figure 2. It is seen that the "total"
for the tracking-error g:( t) will no longer tend to zero asym-
control law is the sum of two components . The first compo-
totically. and will depend on K(t) and £.(t). As a result. it
nent is the value of the end-effector force vector at the nominal
is essential to adapt the feedforward gains and the auxiliary
operating point P; namely :£:'. The second component is the
input to cope with variations of the operating point so as to
contribution due to the incremental controllers K(s) and Q(s).
accomplish trajectory tracking. The feedback gains will also be
Thus. from Figure 2. the "total" control law is given by
adapted to ensure closed-loop stability with desired transient
£:(t) =£:0 + [(t) = :£:' + Kp [!:(t) - ~(t)] performance.
(20)
+ K. [f.(t) - i;.(t)] + C!:(t) + Bf.(t) + Af(t) Now. let us define the 2nx1 position-velocity tracking-
Now. let the "total" reference trajectory vector be K(t) = K° +
1:(t) and the "total" position vector be X(t) = XO +~(t). where error vector ;.(t) ~ (.i~g) and rewrite equation (25) in the
in practice K° = XO . Substituting these in equation (20) gives standard state-space format
the "total" control law in terms of the "total" variables as
where
where Dl = diag.(wl) and D2 = diag.(2{.w.) are constant n.xn
diagonal matrices and the subscript 'm' denotes the reference
model. Equation (28) can be put in the standard state-space
format is an nx1 vector . Let us now consider the reference model state
(29) vector ~m(t) = ( t~g) . The solution ofthe homogeneous
reference model (29) can be expressed as
+ Q1.pd(
di ~~
')(02In
0 {32In
0) where the nx1 vector 'l.(t) is defined as
+ l'
parison with the adaptation scheme. In order to cancel out
A· in the adaptation laws, the matrices in equation (30) are Kp(t) = Kp(O) + 02'l.(t).E'(t) 01 'l.(t).E'(t)dt (47)
chosen as
6. C.S.G. Lee and M.J. Chung: "An adaptive control strat- 23. J .J . Craig: "Robotics: Mechanics and Control,"
egy for mechanical manipulators," IEEE Trans . Aut . Addison- Wesley Publishing Company, Reading, MA,
Control, AC29(9), pp. 837-840, 1984. 1986.
7. T .H. Guo and A.J. Koivo : "On a linearized model and 24. H . Seraji: "Line"! multivariable control of robot manip-
adaptive controller implementation for manipulator mo- ulators," Proc. IEEE Intern. Conf. Robotics and Au-
tion," Journal of Robotic Systems, 1(2), pp . 141-156, tomation, pp. 565-571, San Francisco, 1986.
1984.
25. H. Seraji: "In verses for multi-link robots with applica-
8. A.J. Koivo: "Self-tuning manipulator control in Carte- tion to trajectory tracking," Proc. 19th Asilomar Conf.
sian base coordinate system," AS ME Jour. Dyn. Sys- Circuits, Systems and Computers, pp. 139-143, Pacific
tems, Measurement and Control, Vol. 107, pp. 316-323, Grove, 1985.
1985.
26. H. Seraji: "A simple method for model reference adaptive
9. C.S.G. Lee and B.H. Lee: "Resolved motion adaptive control," Submitted for publication, 1986.
control of mechanical manipulators," ASME J. Dyn. Sys-
tems, Measurement and Control, Vol. 106, pp . 134-142,
1984.
f2
Q(s) • C + Bs + As2
r-- ,..---
fl + f
r + -o-!. K(s) • K + K s
P ~()--!-. J' r!-.. MANIPULATOR
~ Jo x
,- v + 0
'---
DYNAMICS
-
C + Bs + As 2
K +K s
P V
L__ _ _ _ _ _ _ _ _ ---1
R +
""0
.....
<
""..........z
~
:>
X
""
0
..... X
u
.....
< R
""
.....
z
<
V>
~
""
<
u
R +