532 IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. 4, NO.

5 , OCTOBER 1988

Dynamic Modeling and Control by Utilizing an

Imaginary Robot Model

A bstract-A new dynamic model which represents an exact lineariza- nonlinear feedback, and then proceeds to develop the corres-
tion scheme with a simplified nonlinear feedback is presented in this ponding dynamic equations. Obviously, the resulting equa-
paper. To realize this model for robotic systems, the output functions
tions would still be nonlinear and complicated.
should be chosen so that a special decomposition of the total inertial
matrix is satisfied. The concept of an imaginary robot is utilized to A new dynamic model for robotic systems has been
achieve the formulation and to solve the realization problem. Two proposed and developed in [ 151 and [ 181. It has the form of
illustrative examples are given in the paper, one for the Stanford arm and linear system equations obtained from the exact linearization
the other for a PUMA type of robots. An optimal robotic physical design control scheme and the simple nonlinear feedback relation
and a control system design based on the new model are also discussed.
obtained from the operational space formulation. In this paper,
the realization problem and other aspects of the model
I. INTRODUCTION developed in [15] will be investigated. The concept of
imaginary robot will be redefined, and will lend further insight
R OBOT DYNAMICS and system modeling have been
investigated and developed for about twenty years. A
number of symbolic formulations and algorithms for numeri-
into the approach. Two examples of real robots will be
presented to illustrate the realization of imaginary robots.
cal calculation has been created and proposed, such as the They are followed by a discussion on the optimal design of
classical Lagrangian formulation [ 11, recursive Lagrangian robot manipulators associated with imaginary robot model.
formulation [2], Newton-Euler formulation and its recursive Finally, an overall control scheme based on the new dynamic
method [3], [4], and dual-number Lagrangian formulation [ 5 ] , model with a proportional-derivative (PD) control law will be
Kane’s partial velocity [6], and compact dynamic formulations presented.
[7], [SI. All these methods and algorithms have made
11. DYNAMIC
MODELING
OF ROBOT
MANIPULATORS
important contributions to the problems of robot dynamics
computations and robotic control system modeling. For a robot manipulator, we may apply a Lagrangian
Due to high nonlinearity, multi-input and multi-output equation to build up a complete dynamic model. Suppose the
characteristics, and interaxis coupling in robot dynamics, the robot has n joints, a joint position, and joint velocity are
applications of most of the existing advanced control schemes defined as q = (ql q,)T E R” and q = (ql qn)TE
are usually constrained by computational complexity. Further- R “ , respectively. According to [8], [19], the total kinetic
more, various approximations do not often alleviate the energy K of the robot can be written as
problem. Thus in robotic system modeling, one is faced with
the need of building a model which is simple and yet
adequately reflects the real system.
The exact linearization control scheme [9]-[ 111 provides a where W is an n by n positive-definite symmetric matrix and
new approach to modeling robotic control systems. In essence, is called the total inertial matrix of the robot. The formulation
this scheme involves skillful nonlinear transformation so that of Wcan be given in terms of “subjacobian matrices” Jj’s and
the resulting transformed system can be described by a set of generalized mass matrices Uj’s [8], [151, [19], i.e.,
linear differential equations. However, the scheme is compli-
cated by nonlinear feedback. It can be seen that the nonlinear
feedback control, which is also consistent with the result given
,=I
by the resolved acceleration control approach [12], contains at
least an entire robotic dynamic equation and may pose where Uiis defined by
computational problems. Khatib proposed an operational
space formulation [ 131, [ 141, i.e., the Cartesian space formu-
lation. This approach, in essence, first assumes a simple

Manuscript received November 20, 1986; revised January 25, 1988.

mjI miC/’
rJ 1 , j=l, . . e ,

and m, is the mass of link j , Ciis a 3 by 3 skew-symmetric

n

The authors are with the School of Engineering and Computer Science, matrix of cross-product operator for the centroid vector of link
Center for Robotics and Advanced Automation, Oakland University, Roches-
ter, MI 48063. j with respect to thejth coordinate frame, I is a 3 by 3 identity
IEEE Log Number 8822825. matrix, r, is an inertial tensor of link j . The first three

0882-4967/88/1OOO-0532$01.00 O 1988 IEEE

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 GU AND LOH: DYNAMIC MODELING AND CONTROL 533
subjacobian matrices are listed in Fig. 1 for the Stanford arm
and PUMA-like robot. Obviously, matrix W only depends on ai ai cosai sinal
q , and its computation method will be addressed in Section IV
-900 0 0 -1
of the paper.
If we further define an n by n matrix 90' 0 0 1

0' 0 1 0

-900 0 0 -1

900 0 0 1

00 0 1 0

(3)

ai ai cos?ii sin?ii

-900 0 0 -1
which is a function of both q and q and where €3 denotes
0' ap I Ea2
Kronecker product, then the dynamic equation of robot in
90' 0 0 1
Lagrangian formulation can be written as [8], [15]
-900 0 0 -1

900 0 0 1

00 0 1 0

where r is the joint torque and rg = -dN/dq is the joint

torque caused by gravity, and N is the gravitational energy of
the robotic system.
It can be seen that (4) is a nonlinear equation in q and q, but
it is linear with respect to the matrix W , namely, if

m
W=C w;, for m > ~
i= I
J, =

then there must be m uj's such that

m
uj=r+rg r -d2c2 0 . i
i= I

and

where
( - ;-)
W;q+ WT-- W; q = u i , i=l, e . . , m (5b)
J, = I
d,s,,+a,c,
-d,s23
-stl

c2 1
0
-a,c,
0
1
0
o I

(d)
Fig. 1 Denavit-Hartenberg tables and subjacobian matrices. (a) D-H table
for the Stanford robot. (b) D-H table for the PUMA-like robot. 07, = a,+
ea,is the dual angle and e 2 = 0, see [17]. (c) List of the subjacobian
and matrices for main frame of the Stanford arm. (d) List of the subjacobian
matrices for main frame of the PUMA-like ann. s, = sin O,, c, = cos O,, sz3
= sin (0, + 03), and e23 = cos (0, + OJ.

2 %.=W.
i= I contains all centrifugal and Coriolis forces/torques which
always exist as the robot moves. However, since a Cartesian-
We refer to (5) as the linearity of (4) with respect to W . level output function is usually required and the state vector is
From the linearity (3, we can also see that in robotic transformed into a corresponding state vector in modeling a
dynamic equation (4), a dominant factor is the total inertial robot control system, we may take advantage of the nonlinear
matrix W .Thus to simplify robotic dynamic model for optimal transformation to eliminate the second term of (4) by an
control design, we should devote our attention to the matrix additional term due to a particular selection of the output
W. function. From this point of view, let us develop the modeling
It may be observed that one could hardly find a suitable way problem to show the details of simplification.
to simplify the second term in the left side of (4), because it Suppose we choose Q = h(q) as the output function, and

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 534 IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. 4, NO. 5 , OCTOBER 1988

the mapping h : R" R" is a diffeomorphism. Then We now wish to find a condition under which the second
term in (13a) can cancel its third term, i.e., a ( x ) = 0.
Q=
(3*
- 4=G' (7) Specifically, we try to explore a relationship between G and W
such that
1
where G = (ah/dq)T is the Jacobian matrix of h , and
according to the well-known Inverse Function Theorem [ 161,
wG-lGq=(w'-i w) q (14)
G - ' is the Jacobian matrix of the inverse function q =
A-'(@, i.e., whereby (13b) can be reduced to

U= wG-'v=p(x)v. (15)
G-l=(!z)T. (8) In [I51 and [18], it has been rigorously proven that if a
diffeomorphism h is so chosen that its Jacobian matrix G
Now, since becomes a factor of the following decomposition:

Q=Gq+Gq and q=G-'(Q-&) (9) W=GTG (16)

substituting (9) into (4), we obtain then (14) is satisfied, and (15) can be further reduced to
u=GTV (17)
WG-IQ- WG-'Cq+
in which p(x) = GT.
The proof in [ 151 has been given both in a direct way and by
where U = r + rg E R" is defined as the input function of the means of the well-known Hamiltonian canonical transforma-
original system. This result is consistent with the previous tion theory.
work on resolved acceleration control [12] and also on the Therefore, the linear equation (12) and simplified nonlinear
exact linearization control with output decoupling [9], [IO]. feedback (17) become a new compact model for the original
Let the acceleration vector associated with the output robotic system described by (4). In fact, this model is also
function be a new input function, i.e., applicable to any nonlinear dynamic system, as long as its
v = Q. system equation can be written in the same form as (4) and the
decomposition (16) holds.
Then, (1 1) shows a linear system equation with the Brunovsky The following simple example with a scalar input and a
canonical form if the new state vector is defined by X = scalar output provides a demonstration of the new model. If a
(Q' Q')' E RZn,namely, dynamic system has the following differential equation:

X=AX+BV &+l.
( e + 2 4 + i)g+--- eZ=u, 8>0
where 2 4

A=[: 0'3 .=[:I.

and
then comparing with (4)

w=e+24+1=(4+1)2
Substituting (11) into (lo), we obtain a relation between the
old input U and new input V. and

w=e(a w / a e ) = ( 4 + i ) & 4 .
( - ;-)
WG-IV- W G - ' G q + W T - - W q=u. (13a)
When the output function is chosen as
Since all terms in the left side of (13a) depend on the old state
vector x = ( q T q T ) T ,(13a) also represents a nonlinear
feedback in the exactly linearized system given by (12). Let us
y=(;~+i) e (19)
define
so that G = dy/a8 = & + 1 and G 2 = W , it leads to a scalar
case of decomposition (16). Then, it can be readily verified
a(x)= - wG-'Gq+ that

and a ( w ) = O and /3(x)=G=&+ 1 (20)

/3(x)= WG-1 whenever x = (6 e) T. Equation (20) gives a simple nonlinear
feedback
so that

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 GU AND LOH: DYNAMIC MODELING AND CONTROL 535

The system equation (18) has been transformed to a robotic physical design problem for matching the above
conditions. The detailed discussion on design problems will be
y= V (22) addressed in Section V .
and defining X = ( y -9) T , it becomes (12). For a number of existing robots, each link does not exactly
Obviously, the main problem is now to seek an appropriate satisfy the conditions of diagonal inertial tensor and only one
diffeomorphism h such that its Jacobian matrix G satisfies (16) nonzero coordinate of centroid. Then the remainder matrix
for a given dynamic system described by (4). This is also a A Wmay have a more complicated form than the constant one,
realization problem associated with the new compact model and the nonlinear feedback may accordingly contain more
described by (12) and (17). terms than (24) but the decomposition (23) is always valid.
However, if some robots have link parameters which are close
to satisfying the above conditions, then we may take them to
111. REALIZATION AND IMAGINARY ROBOT
approximately realize the decomposition (23) with the constant
According to the analysis in the last section, there are two AW. The robustness for parameter deviations in robotic
possible methods for realizing the new model for a given control with the new model has been verified by computer
robotic system, namely: simulation and will be discussed in Section V I .
1) Find a decomposition for the positive-definite symmetric Due to the fact that the output Q = h ( q ) which leads to the
matrix W, i.e., W = G TGby some mathematical method and realization of (23) with constant A W is, in general, not
make sure that the resulting G is a Jacobian matrix. pointing to the wrist of the robot arm, its arrow may be located
2) Find an appropriate output function Q = h ( q ) such that somewhere in the robot workspace. We need an imaginary
its Jacobian matrix G = (8h/8q)Tsatisfies robot model for ease of analysis. The imaginary robot will be
so chosen, in contrast with its associated real robot, that the
W = G T G +A W (23) position vector of the imaginary robot wrist is just the choice
of Q = h ( q ) ,whose Jacobian matrix G satisfies (23) with the
with a simple remainder matrix A W .
constant A W . In the next section, two typical examples are
For the first method, although it is not very difficult to find
given to illustrate the approach to selecting a suitable
a decomposition in the form of W = G T G ,to require that G
imaginary robot.
be a Jacobian matrix, which is a total derivative of some vector
In order to justify the simplification of the nonlinear
function, may not be simple.
feedback formulation, let us compare the computational
For the second method, although the decomposition (23)
complexity between the conventional method (13) and the new
can always be realized by any choice of the output function,
model (24). Because the linearized model is the same for both
according to the linearity (5) the robotic dynamic equation is
cases, we only need to find the number of multplications
now divided into two portions: one has the total inertial matrix
required to compute (13) and (24). Reasonably, we assume
GTG and other has AW. It is obvious that if Q = h ( q ) is
that the matrices W , G,and G-I involved in the two formulas
chosen arbitrarily, the second portion with A W may give rise
can be obtained in symbolic form, and in each updating period
to the same complexity of formulation as the original system,
they do not need computer time. Then, a comparative table for
no matter how simple the GTGportion is. Therefore, in order
the number of multiplications in computing the nonlinear
to achieve a global simplification, we have to choose an
feedback between (13) and (24) for a robot arm with n joints is
appropriate output Q = h ( q ) so that the A W portion has a
as follows:
simple dynamic relation. For example, if A W is a constant
matrix, then the dynamic equation of the A W portion has only
one term A Wq, and its centrifugal and Coriolis terms are all AWq m(x) @(x)V Amount If
zero. Since the GTG portion can be transformed into the by (3) by (9) by (13) by (13) Number n = 3
_ _
exactly linearized system (12) with the simplified nonlinear
feedback (17), thus if an output Q = h ( q )can be found so that u = m + p v
(13) n3 0 5n2 2n2 n3+7n2 90
AW is a constant matrix, we then obtain U = G ‘Y + A Wq
(24) 0 3n2 0 n2 4n2 36
U =GT Q +A wq =GTV+ A wq. (24)

Equation (24) has the form of a dynamic system equation The table indicates that the nonlinear feedback (24) based on
having only Newton’s forces/torques and no centrifugal and the imaginary robot model reduces computer time signifi-
Coriolis terms. cantly. If we adopt an approximate interpolation algorithm to
A method of realization for the decomposition of (23) with update q instead of (9), then (24) will be further simplified.
constant AW has been proposed in [15]. If we consider Particularly in real-time control, even for each entry of all
position as the output function, and each link of a robot has at concerned matrices, its computational time should be taken
least two dimensions being symmetric, i.e., all inertial tensors into consideration. Thus the new model provides a much faster
rj’s are diagonal matrices and the centroid of each link has algorithm, because the total inertial matrix W is not involved
only one nonzero coordinate component with respect to its in (24) and W which does appear in the conventional nonlinear
own coordinate frame, then the realization can be accom- feedback (13) usually contains very long terms in most of its
plished for many such robotic manipulators. This also leads to entries.

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 536 IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. 4, NO. 5 , OCTOBER 1988

IV. ILLUSTRATIVE
EXAMPLES
Let us now study the realization problem for the decomposi-
tion (23) with the constant AW. A general procedure for
finding a corresponding imaginary robot for a given robot is as
follows:
Procedure I-Finding the Imaginary Robot
1) Using ( 2 ) , calculate W .
2) Determine an appropriate output function Q = h ( q ) ,
and compute its Jacobian matrix G = ( d h / d q ) T and W - wrist of real robot
G TG. W'- wrist of imaginary robot
3 ) Find AW = W - GTG. Let Q = h ( q ) contain some
unknown parameters, then determine these parameters
(a)
so that A W becomes a constant matrix. The imaginary
robot is constructed accordingly by utilizing these
parameters. lY2
Example 1: Stanford Robot: Suppose we consider the first
three joints of the Stanford robot. All the first three subjaco-
bian matrices Jj's and the D-H table of the Stanford arm are
listed in Fig. 1. If each of the first three links has only nonzero
z coordinate of the centroid with respect to its own frame, i.e.,
the skew-symmetric matrix of the cross-product operator for
the centroid vector is given by

cj=[;
0
8
-zj

:]
0
, j=l,2,3 (25)
L W - wrist of real rob0 t
W'- wrist of imaginary robot
and the inertial tensor of each link is a diagonal matrix (b)

rj=mj
[: 1 ;]
0 kjj 0 j = 1 , 2, 3

where k:j, k i j , and kij are the gyration radii of link j with
respect to the x, y , and z axes of the jth frame, respectively,
(26)
Fig. 2 Real robot and imaginary robot. (a) Stanford arm and its imaginary

given by
robot. @) PUMA-like arm and its imaginary robot.

and Zj, j j j , i;. are the coordinates of centroid of linkj projected

on theJth frame, see Fig. 2 , then, three Wj'scan be found by
Wj = J,'ujjJ,, j = 1, 2, 3 , i.e.,
we replace d2and d3by new parameters d2and d3(actually, d3
di+2.Z1d2+kL1 0 0
Wl=ml
[::
di+k:2si+kt2ci
0O O I
0

-Z2d2c2 0
(274
and d3 are joint variables) to define the wrist W' of the
imaginary robot as

W2= m2 - Z2d2~2 ki2

0 0 0

- d2d3~2- dz.Z3c2 - d2~2

dZZ3c2 d: + 2d3& + kj3
- d2s2 0 1

where ~ 3 1 1= di + d:si + 2d3&si + k$si + k;,ci. Then, the Jacobian matrix G of h is determined as
The total inertial matrix is the sum of Wj's -d2c1-d3Sls2 d3cIc2 CIS2
w = W ] +w2+ w3 (28) d3slc2 SlS2

Since the position vector of the Stanford arm's wrist W is -d3~2 ~2

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 GU AND LOH: DYNAMIC MODELING AND CONTROL 537

and rn3G2d2=(rn2+m3)a2d2+m2.f2d2 (364

d; +ais; - d2d3~2 - d2~2 J2J3 = d2 d3 + Z3 d2 (36e)

-d2d3~2 J: (32) then the remainder matrix is a constant matrix given by
r
1i-
1
- &s2 0 O 1
Now let us compare (32) with (28). It can be seen that if the
wll 0 0 1
Aw= m2b:;;,b: m3b: (37)
following conditions hold:
m3b:
J2=d2 and J 3 = d 3 + Z 3 (33a) where

Z2 =0 and Z3 = - Jk:, - ki3+ m2/m3(k:,- k i 2 ) (33b) wll = rnl (d: + 2Z1d2+ k:,) + rn2(k22+ d:)
then the remainder matrix AW becomes constant, i.e., + m3(kt3+ d: - a:)

[sl : :]
b;=k:,+ k:2- k t 2
AW= 0 ~2 0 (34) and

b:= k:3 + k f 3- k:3.

where
A feasible solution for (36a)-(36e) is as follows:
w 1= rn, (d: + 2Z1d2+ k;,) + m2(d ;+ k i 2 )+ m3k f 3
and .f2 = - ( k:2 - k;,)/a2 (384

w2=rn2(k:,+k i 2 - k:,)+m3(ki3+k i 3 - k i 3 ) .
(38b)
The two constraints which must be satisfied in (33b) are
reasonable conditions for the Stanford robot, while (33a)
specifies the location of imaginary robot wrist W' . Since Z3 <
Z3=(p2- l)d3-pd(p2- l)d:+k:,-ki, (38c)
0 and IZ3J < total length of link 3, the new wrist W' should be
at some medium point on the prismatic link, i.e., link 3, as d2 = pd2, 62 = pa2, and d3= (z3+ d3)/p. (38d)
shown in Fig. 2(a).
Similarly, (38a)-(38c) specify the constraints to be satisfied
ExarnP1e 2: PUMA-Like A PUMA-1ike robot
which has been discussed in [15] will be considered in this by the physical robot, and the imaginary robot wrist Wt is
determined by substituting (38d) into (35). Also, the con-
example. For many other industrial robots, such as the
straints (38a)-(38c) are reasonable for the PUMA-like robot.
Cincinnati Milacron 776 and GMF S-380R, their main frames
As a numerical example, let d2 = 15 cm, a2 = d3 = 43 cm,
may be considered as similar to the PUMA-like robot with the
k 2 = 1400cm2, k:2 = 220cm2, k:3 = 680cm2, ki3 = 40
second joint offset d2 = 0. Thus the following discussion is
c s , and rn2/m3 = 2, then according to (38a)-(38d), the
applicable to a number of industrial robots with articulated
numerical solution is
configuration.
We can apply (2) to calculate W = W l + W2 + W3for the X 2 = -27.44 cm p= 1.3
PUMA-like robot, as derived in [15]. Using the same method
as used for the Stanford robot, the output function is chosen as Z3= -27.23 cm
follows:
and
Q = h ( 4)= Jm3p 0"
& = 19.5 cm li2=55.9 cm d3=12.13 cm.
=Jm,
[
62cl c2 + 23c]s23 - d's1
li2S,C2+&S1S23+d2CI
- 62s2 + d3 e23
] . (35) The schematic diagram of the PUMA-like robot and its
imaginary robot in Fig. 2(b) is drawn in proportion based on
the above numerical solution.
Likewise, the three parameters GI,&, J3 are chosen so that Once (38a)-(38d) are all satisfied and the corresponding
A W is a constant matrix in the decomposition (23). output function (35) is chosen, the PUMA-like robot system
After G and GTG have been determined for (35) it follows can be transformed into (12) with the nonlinear feedback (24).
that if the following five quadratic equations simultaneously Since in the output equations (30) and ( 3 3 , all function forms
hold: of Oi's are unchanged, the imaginary robots are ensured to
have simple relations with their individual real robots, as
shown in Fig. 2.
From the above two examples, the determination step of
output function in Procedure 1, i.e., step 2, can be embodied
in the following

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 538 IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. 4, NO. 5 , OCTOBER 1988

subroutine: robot model, the nonlinear feedback (24) still has a term
A Wq, and it needs to update the joint acceleration vector q at
1) Determine the position vector POWfor the robot wrist each sampling point. An accurate formula for 4 is based on
with respect to the base coordinates (9), and may be time-consuming. We can utilize a feedforward
2) The output function Q = h ( q ) is obtained by replacing method, i.e., let q = qd, or an approximate interpolation in
each link length Or joint offset in the expression Pow terms of the neighboring 4’s. In order to effectively simplify
by an unknown parameter and then PremultiPIYing the computation for A Wq by an approximation algorithm, one
fi3 to the position vector. has to reduce the effect of AW and take it into the design
In fact, this subroutine of output determination as well as consideration.
It can be shown that the entries of the remainder matrix A W
Procedure 1 can apply to a number of industrial robot arms
with either articulated, spherical, cylindrical, or Cartesian either in (34) or in (37) cannot be all zero. However, we may
redistribute the inertia moments for each link under the
configuration, as long as their link parameters approximately
have the certain condition of simple mass and inertia distribu- optimal design so that all the positive entries of A W become as
tions, such as (25) and (26). Also, the determination of small as possible. Then, a tradeoff between the maintenance of
imaginary robot and the remainder matrix A W by following an accurate dynamic model and the further simplification of
up Procedure 1 with the subroutine is required to be carried nonlinear feedback computation could have a more feasible
solution.
out only once for each given robot arm. Then, the determined
imaginary robot and simplified nonlinear feedback (24) can be VI. CONTROL
SYSTEM
DESIGN
permanently implemented in the control system of the given
robot arm for executing any task. Having exactly linearized the robotic system equation and
simplified the nonlinear feedback function, the remaining
V. OPTIMAL
DESIGN
of ROBOT
MANIPULATORS work is the conventional control design. All optimal control
methods suitable for designing linear control systems can be
According to the previous discussion, the realization of an
applied to the new model ( X , V, Y) in (12) with the output
imaginary robot model requires a physical optimal design for
defined as Y = Q .
robot manipulators. Robotic optimal design problems have
For a linear and time-invariant system, the optimal control
been investigated in recent years. Asada and Youcef-Toumi
law relative to a quadratic performance measure has a state
modified the arm structure and mass distribution to simplify
feedback form [20]
the robotic dynamics [2 11. Then, they explored the possibility
for decoupling of the total inertial matrix as well as making it AV*= - K A X (39)
invariant [22]. Also, Yang and Tzeng redistributed the inertia
moments of the robot arm to realize dynamics simplification where K is an n by 2n gain matrix, and AX =
and linearization [23]. Such a class of robotic design based an (AQTA Q T ) T ,in which AQ = Qd - Q , the error function
mass balancing, force balancing, and inertia redistribution between a desired output Qd and an actual output Q . If K =
provided a decoupled and simple dynamic model. Under the (Kl Kz), where each Ki is an n by n matrix, then
balanced condition, a solution to the time optimal control
problem was presented in [24].
A V* = - K~AQK
- ~AQ. (40)
In this section, an optimal design problem is defined
Equation (40) is known as a proportional-derivative (PD)
alternatively based on the realization of imaginary robot
control law.
model. If all the link parameters of a robot main frame are
Since
adjusted so that
A V * = Vd- V * = Q d - v * (41)
1) each link has a centroid with only one nonzero coordi-
nate, such as (25) or any x or y coordinate, and a (40) can be rewritten as
diagonal inertial tensor (26),
2) some parameters satisfy a set of constraints determined V* = 0, + K1 A Q + KzA e. (42)
by Procedure 1, such as (33b) for the Stanford arm and
(38a)-(38c) for the PUMA-like arm, Often, an additional term is inserted into (42), i.e.,

then we refer to the adjustment as the optimal design

associated with imaginary robot model.
I
V * = Q d + K , A Q + K , A Q + K , AQ dt. (43)

It can be seen that in this optimal design, the total inertial This is a typical proportional-integral-derivative(PID) con-
matrix W is not required to be diagonalized, or even to be troller, and K , , K 2 , K3 are the coefficients of proportion,
invariant. Thus it relaxes design conditions for simplification derivative, and integral terms, respectively. The common
of the robotic dynamic model in comparison with any previous criteria of selections of K , , K 2 , and K3 are to guarantee the
mass balancing or inertia redistribution design criteria [22] , stability of the linearized system, quick convergence, and
[23], and provides potentially wide applications to robotic small steady-state error.
physical design. To show the global stability of the entire control system with
Under the optimal design associated with the imaginary the PD control law (42) and nonlinear feedback (24), a

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 GU AND LOH: DYNAMIC MODELING AND CONTROL 539

Lyapunov function can be defined as

1 1
VL=- Q T Q + - AQTKIAQ
2 2
1 1 1
= - q T W q - - q T AW q + - A Q T K I A Q (44)
2 2 2
where Q = Gq and (23) have been applied, and VL > 0
except Q = 0 and Q = Q d if K 1 is a positive-definite gain
matrix. Using the same method as discussed in [25] and [19],
substituting (24) and (42) into the derivation, and noticing that
Fig. 3 A block diagram of nonlinear feedback control with imaginary robot
AW is constant, we obtain the time derivative of VL as model.
follows :

=@TU - q T A wq - A Q T K ] Q
+ wq)- q T~ wq - A Q T
= q T ( ~ T ~ A* QI
~

= - QTK2Q

matrix. In (45), we fixed the desired output Qd so that &d =

- 0. It now remains to show that when
Qd - 0
(45)
which is nonpositive if K2 is also chosen as a positive-definite

= 0, the error
AQ = 0. This is true if G is nonsingular [19], [26]. Therefore,
the control law (42) and nonlinear feedback (24) asymptoti-
cally stabilize the entire robotic control system with imaginary
robot model.
In (43) or (42), Q d , Q d , and & are determined by the
desired trajectory. We may use differential motion which
follows the Jacobian equation in robotic kinematics, i.e.,
cc
J
v

-2 /-

-12
.i? ,\ .k .b 1 112 1!4 116 118 1 2!2 2 .I4 216 2.8
I

t (sec.)

Fig. 4 Three components of output error vector versus time in simulation

study of PUMA-like robotic system.

link parameters are purposely set with up to k 10-percent

deviations between the plant and the model, and also, q in the
nonlinear feedback (24) is updated by approximation methods
of either the first-order interpolation or feedforward loop with
where U and o are robot hand linear and angular velocities,
(46b) in computer simulation. The result has the same
respectively, Jd is a Jacobian matrix for the robot hand at the
convergence behavior and speed as obtained without the
desired q d . Having found q d , q d , and qd from (46) at each
parameter deviations. It can be concluded that the robustness
sampling point, we substitute them into the formulas of Qd, and global asymptotic stability for the control scheme with an
Q d , and &. All the determinations are executed off line. A
imaginary robot model can effectively reject any disturbance
block diagram of the robotic control system by utilizing due to unmodeled dynamics, parameter uncertainty, and
imaginary robot model and exact linearization scheme with the
payload change, provided that those external factors are
simplified nonlinear feedback is shown in Fig. 3.
bounded within a certain range. On the other hand, the optimal
A computer simulation study has been made in the case of physical design which has been discussed in Section V as well
the PUMA-like robot with first three joints. The proportional as the controller hardware design may also have to involve all
gain matrix K 1 and derivative gain matrix Kz were chosen as possible considerations to avoid or reduce the unmodeled
361 and 121, respectively, where I is a 3 by 3 identity matrix. dynamics and parameter uncertainty, so as to ensure that the
The desired trajectory was a parabolic curve with one- proposed control scheme works successfully for real-time
dimensional acceleration. A diagram of the error vector AQ robotic system.
versus time t is shown in Fig. 4. The resulting plot illustrates a
VII. CONCLUSION
quick convergence behavior of AQ within 1-1.5 s as external
disturbance interfered. The concept of an imaginary robot has been utilized and
In order to verify the robustness of the control scheme, most further discussed. To realize the new robotic dynamic model

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.
 540 IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. 4, NO. 5 . OCTOBER 1988

which can be described by a set of linear system equations with [I61 B. A. Dubrovin, A. T . Fomenko, and S . P. Novikov, Modern
a simplified nonlinear feedback, an appropriate selection of Geometry Methods and Applications. New York, NY: Springer-
Verlag, 1984.
output functions is required. The output and its Jacobian ~ 1 7 1 y. L. G ~ “Analysis
, of robots with redundancy,” Ph.D dissertation,
matrix have to satisfy a decomposition of the robotic total Purdue University, W. Lafayette, IN, May 1985.
inertial matrix. This leads to the introduction ofthe concept of [I81 D. E. Koditschek, “Robot kinematics and coordinate transforma-
tions,” in Proc. 24th IEEE Conf. on Decision and Control (Ft.
an imaginary robot. The solutions for constructing the Lauderdale, FL, Dec. 11-13, 1985). pp. 1-4.
imaginary robot for the Stanford robot and the PUMA-like 1191 H. Asada and J-J. E. Slotine, Robot Analysis and Control. New
robot have been presented. Furthermore, the robotic physical York, NY: Wiley, 1986.
[20] H. Kwakernaak and R. Sivan, Linear Optimal Control Systems.
optimal design associated with the imaginary robot model has New York, NY: Wiley-Interscience, 1972.
been developed. The optimal control for the entire robotic (211 H. Asada and K. Youcef-Toumi, “Analysis and design of a direct-drive
system based on the imaginary robot model and simplified arm with a five-bar-link parallel drive mechanism,” ASME J . Dyn.
Sysf. Meas. and Contr., vol. 106, no. 3, pp. 225-230, 1984.
nonlinear feedback as well as the justification of global [22] K . Youcef-Toumi and H. Asada, “The design of arm linkages with
asymptotic stability and the verification of robustness have decoupled and configuration-invariant inertial tensors: Part 1 : Open
also been discussed. It is concluded that the proposed new kinematic chains with serial drive mechanisms; Part 2: Actuator
relocation and mass redistribution,” in Robotics and Manufacturing
model could be implemented in robotic systems to realize real- Automation (Winter Annual Meeting of ASME, Miami, FL, Nov.
time control. Relying on its robustness and global stability, the 1985), pp. 145-161.
imaginary robot model applies to optimal control of many [23] D. C. H. Yang and S. W. Tzeng. “Simplification and linearization of
manipulator dynamics by the design of inertia distribution,” Inc. J .
industrial robots. If a robot arm is designed with the proposed Robotics Res., vol. 5 , no. 3, pp. 120-128. 1986.
optimal design criteria, then real-time control can perfectly be [24] W. S . Newman and N. Hogan, “The optimal control of balanced
realized by the new model for the robot arm. manipulators,” in Proc. ASME Winter Annual Meet. (California,
Dec. 1986). pp. 177-184.
(251 S . Arimoto and F. Miyazaki, “Stability and robustncss of PID feedback
REFERENCES control for robot manipulators of sensory capability.” presented at the
J. J. Uicker, “On the dynamic analysis of spatial linkages using 4 x 4 3rd Int. Symp. of Robotics Research, Gouvieux, France, July 1985.
matrices,” Ph.D dissertation, Dept. of Mech. Eng. and Astronaut. [26] J. I. Craig, Adaptive Control of Mechanical Manipulators.
Sci., Northwestern Univ., Evanston, IL, 1965. Reading, MA: Addison-Wesley, 1988.
J. M. Hollerbach, “A recursive Lagrangian formulation of manipulator
dynamics and a comparative study of dynamic formulation complex-
ity,” IEEE Trans. Syst., Man, Cybern., vol. SMC-10, no. 11, pp. You-Liang Gu (M’86) was born in Shanghai,
730-736, NOV. 1980. China. He received the B.S. and M.S. degrees in
J . Y. S . Luh, M. W. Walker, and R. P. Paul, “On-line computational electronics from Tianjing University in 1967 and
scheme for mechanical manipulators,” J . Dyn. Syst. Measure. from the University of Science and Technology of
Contr., vol. 102, no. 2, pp. 69-76, June 1980. China in 1981. respectively, and the Ph.D. degree
J . Y. S . Luh and C . S . Lin, “Automatic generation of dynamic in electrical engineering from Purdue University,
equations for mechanical manipulators,” in Proc. I981 Joint Auto- West Lafayette, IN. in 1985.
matic Control Conf. (Charlottesville, VA, June 17-19, 1981), pp. From 1968 to 1978, he worked as an Engineer on
TA-2DlI-5, new model instrument design for the Electronic
r51 J. Y. S . Luh and Y . L. Gu, ‘‘Lagrangian formulations of robot Instruments Company in Northern China. Since
dynamics with dual-number transformation for computational simplifi- 1981, he has been a Teaching Assistant at the
cation,” in Proc. 1984 Conf. Information Sciences and Systems Department of Physics and a Research Assistant at the School of Electrical
(Princeton, NJ, Mar. 14-16, 1984), pp. 680-684. Engineering, Purdue University. Currently, he is an Assistant Professor of the
T . R. Kane and D. A. Levinson, “The use of Kane’s dynamical School of Engineering and Computer Science at Oakland University.
equations in robotics,” Int. J . Robotics Res., vol. 2, no. 3, pp. 3-21, Rochester, MI. His current interests are in the areas of robot kinematics.
Fall 1983. dynamics, optimal and adaptive control, robotic vision, syntactic pattern
M. Thomas and D. Tesar, “Dynamic modeling of serial manipulator recognition, and intelligent control in robotic systems.
arms,’’ ASME Trans. J . Dyn. Syst. Meas. and Contr., vol. 104, pp.
218-228, Sept. 1982.
Y. L. Gu and N. K. Loh, “Dynamic model for industrial robots based Nan K. Loh (M’70) was born in Malaysia. He
on a compact Lagrangian formulation,” in Proc. 24th IEEE Conf. on received the B.Sc. degree in electrical engineering
Decision and Control (Florida, Dec. 1985), pp. 1497-1501. from the National Taiwan University, Tainan,
A. K. Bejczy, T. J. Tarn, and Y. L. Chen, “Robot arm dynamic Taiwan, in 1961, and the M.Sc. and Ph.D. degrees
control by computer,” in Proc. I985 IEEE Int. Conf. on Robotics in electrical engineering from the University of
and Automation (St. Louis, MO, Mar. 25-28, 1985), pp. 960-970. Waterloo, Waterloo, Ont., Canada, in 1964 and
Y. L. Chen, “Nonlinear feedback and computer control of robot 1968, respectively.
arms,” Ph.D dissertation, Washington Univ., St. Louis, MO, Dec. He taught at the University of Iowa. Iowa City,
1984. from 1968 to 1978, and joined the School of
N. Coleman, N. K. Loh, and Y. L. Gu, “Optimal control with exact Engineering and Computer Science at Oakland
linearization of robotic manipulators,’’ in Proc. 1986 American University, Rochester, MI, in 1978. He is the
Control Conf. (Seatle, WA, June 18-20, 1986), pp. 852-857. Associate Dean for Graduate Studies and Research there as well as Director of
J. Y.S. Luh, M. Walker, and R. Paul, “Resolved-acceleration control the Center for Robotics and Advanced Automation, which he co-founded in
of mechanical manipulators,” IEEE Trans. Automat. Contr., vol. 1981. In 1984 he was appointed the John F. Dodge Professor of Engineering.
AC-25, pp. 468-474, 1980. His areas of research interest include computer controls. computer vision.
0. Khatib, “Dynamic control of manipulators in operational space,” robotics, digital signal processing. time series analysis, and estimation theory.
presented at the 6th IFTOMM Congress on Theory of Machines and He has authored and co-authored over 140 technical papers and reports. He is
Mechanisms, New Delhi, India, Dec. 15-20, 1983. an active engneering consultant and has served as consulting scientist for
u41 ~- , “The operational space formulation in robot manipulator numerous national and international organizations and corporations. The
control,” presented at the 5th ISIR, Tokyo, Japan, Sept. 11-13, 1985. consulting activities include, among others, industrial development, engineer-
1151 Y. L. Gu and N. K. Loh, “Control system modeling for robot ing design, and high technology. He also serves o n the editorial boards of
manipulators by use of a canonical transformation,” in Proc. 1987 several technical journals.
IEEE Int. Conf.on Robotics and Automation (Raleigh, NC, Mar. Dr. Loh is a member of the Board of Directors of the Society for Machine
30-Apr. 3, 1987), pp. 484-489. Intelligence (SMI).

Authorized licensed use limited to: UNIVERSIDAD AUTONOMA METROPOLITANA. Downloaded on January 03,2025 at 18:24:06 UTC from IEEE Xplore. Restrictions apply.