You are on page 1of 10

ROBOTIC MANIPULATORS AND THE PRODUCT OF EXPONENTIALS FORMULA ~"

R. W. Brockett
Division of Applied Sciences
Harvard University
Cambridge, Massachusetts 02138

ABSTRACT. The manipulation of rigid bodies by manipulators which are motor driven
kinematic chains is a fundamental aspect of robotics. In this paper, we discuss
the kinematics of such processes and discuss the classification of kinematic
chains using ideas from algebra and group theory. Earlier work on the role of Lie
groups in mechanisms is contained in Herv~ [ 7], but the role of Lie algebras is
not considered by this author. More relevant (but less group theoretic) is the
extensive case-by-case analysis found in Pieper's thesis [ 9]. In fact, Pieper's
work suggests an interesting and rather general problem in Galois theory which is
directly related to manipulation. Also of interest is the well-known Baker-
Campbell-Hausdorff formula for the derivative of a product of exponentials since
such products are of fundamental importance in the study of kinematic programming.

This work was supported in part by the U. S. Army Research Office under Grant
No. DAAG29-79-C-0147, Air Force Grant No. AFOSR-81-7401, the Office of Naval
Research under JSEP Contract No. N00014-75-C-0648, and the National Science
Foundation under Grant No. ECS-81-21428.
121

i. THE EUCLIDEAN GROUP

The manipulation of rigid bodies, whether it be done by animate or inanimate


objects, can be described by a curve in the group of rigid motions in three
dimensional euclidean space. This means, of course, that in a mathematical dis-
cussion of manipulation the Lie group ]E(3),consisting of all euclidean motions,
is bound to play a fundamental role. We begin, therefore, with a brief discussion
of some of its properties.

We use the notation ]E 3 to denote ordinary cartesian 3-space with the standard
inner product (x,y) = ZxiY i. An affine transformation

y = Ax + b

is said to define a euclidean transformation if A is an orthogonal matrix. If


y = AlX + b I and z = A2Y + b2, then z = A2AIX + A2b I + b 2. A comparison of this,

with the identity

establishes that the euclidean transformations, thought of as a group under


composition and the set of matrices of the form

thought of as a group under multiplication, are isomorphic. We denote the three


dimensional euclidean group by IE(3) and refer to the above representation as the
standard matrix representation.
Of course, the set of 3 by 3 orthogonal matrices is a three dimensional group
and so ~E(3) is 6-dimensional. Its composition law is continuous in the obvious
sense, and so IE(3) can be regarded as a Lie group. Its Lie algebra has the
representation

The Lie group ]E(3) is neither simple nor solvable, but is the semidirect product
of the simple Lie group of orthogonal transformations and the abelian Lie group
consisting of all translations.
The Killing form defines a natural Riemannian metric on the orthogonal group,

and, of course, on IE 3 there is a Riemannian structure as well. There is,


moreover, a canonical choice of Riemannian structure on ]E(3) and, in particular,
a natural volume measure on ]E(3). As a fifial remark, given any matrix of the form
122

there exists

such that expN = M. That is, the exponential map is onto ~ ( 3 ) . Since we can
obviously write

we obtain Euler's theorem stating that any rigid motions can be thought of as a
translation followed by a rotation about a line passing through a preassigned
fixed point. Chasles stated that any rigid motion can be achieved by rotation and
translation which commute. A rotation about the origin

and a translation

commute if, and only if, Ab=b. Thus, Chasles theorem says that we can express any
euclidean transformation as shift of origin followed by a commuting rotation and
translation; i.e.,

[:
with Ad=d and <c,d> = 0. (This last condition is necessary in order to make c and
d unique.) Since the range space and null space of a skew symmetric matrix are
orthogonal, the identity

implies that every motion of the form expl ~ ~I@ can be thought of as screw motion

with respect to some choice of origin [ 7 ]. We may also remark that ~ ( 3 ) is an


algebraic group in the usual sense [ill. This fact underlies some work of Pieper
[ 9 ] which we will come to later.
123

2. KINEMATIC CHAINS

Consider now robotic manipulators which consist of rigid bodies jointed at


single degree of freedom joints (see Figure i). If we fix a right-handed triad of

er %

~ : A Kinematic Chain

orthogonal vectors at the tip of each member of the chain, it is not too difficult
to see that the euclidean transformation, which describes the position and orien-
tation of the (i+l) st triad in terms of the i th, is

exp @i

~hus,
• the triad fixed at the free end is related to that at the base by the product

T(81" 2''''er)=Ml(exPOl[2 r OJ

where the M. is the element of the euclidean group which mapS the coordinate
i
system at the end where it joins the (i-l) st element to the coordinate system
where it joins the (i+l)St°
It is easy to see that P(expM)P -I = exp(PMp-l), and, thus, we can use the

identity MexpN = (expMNM-I)M repeatedly to write

HI@I H2@2 Hr 8r
01,02 .... %) = Me e e

This product of exponentials formula not only applies to chains of the form shown
in Figure i, but also applies to mechanisms containing prismatic joints (see
Figure 2) if we allow the product to include generators of the translation
elements, i.e., factors of the form
124

~ure 2: A Prismatic Joint

It is to be emphasized that any such kinematic chain defines an M and the


H!,H2,...H r with the H i being elements of the Lie algebra of IE(3) and the M an

element of the Lie group IE(3). Once a choice for the coordinate systems at the
beginning and end of the chain are made and a definition of which angles correspond
to O. = 0 is selected, those matrices characterize and are characterized by the
l
chain.
In view of the fact that ]E (3) is an algebraic group, it is possible to re-
express matters in such a way as to eliminate transcendental functions. Thus, we
can algebraically parameterize the one parameter subgroup as

-i 0 0 - x 0 x2+y 2 = i
exp e = ;
0 0 0 1
0 0 0 0

and, in a similar way, describe the general one parameter subgroup (= screw motion)
in purely algebraic terms. A basic problem in robotics is that of solving the
equation in ~ (3)
T = eXpHl81 ... expH6O 6

for the unknowns ei,82,...86, An interesting question arises in this regard. How

does the degree of this set of equations depend on the fixed generators
HI,H2,...H 6. This problem is studied in the thesis of Pieper [9 ], Chapter 3. An
even more relevant question is that of determining how the Galois group of these
equations depend on the choice of HI,H2,...H 6. This latter circle of ideas seems

to be vaguely related to our study [i0].


125

3. THE DIFFERENTIATION FORMULA

The equations of motion for the rigid bodies, which make up a robotic manipu-
lator, can he deduced from classical physics. The equations of motion for the
motors, which move the joints, must combine with these to get the overall
description. However, there is a purely kinematical problem which is both con-
ceptually and practically significant and which may be considered independently.
The expression T(81,e2,...er) of the previous section defines a mapping of

M r into ~(3). The user of the robot would like to specify a path in E(3),
whereas the motors of the robot drive the joint angles 81,82,...e r. This means

that the robot should provide computational equipment to invert the mapping

T:IRr + ~(3). If a path is specified in IE(3), then finding a path in IRr which
generates it can be thought of as finding a "program," i.e., a sequence of moves
for the joint angles.
The paths in ~(3), which are required in robotic manipulation, may be very
complex; however, it seems that it is quite useful to understand how to solve the
above kinematical programming problem for curves in ~(3) which are constructed
from segments of one parameter subgroups in~(3), i.e., segments of the form

Of course, straight line displacements, rotations, screw motions, etc., are all of
this folm. For such motions, the kinematical progran~ing problem is that of
solving

<I I>
exp
0
~(t) M = Mexp H l O l ( t ) exp H202(t ) . . . exp H r O r ( t )

for Ol,e2...0 r.

Since it is not enough to simply solve this for one value of ~(t), it is
natural to use implicit function theorem ideas to find solutions in the neighbor-
hood of a given solution and in that way generate the joint angle program which
corresponds to a given choice of 4('). The obvious way to do this is to ask that
be differentiable and to generate the 6's by solving the differential equation

~(t)[~ ~] exp[~ ~ ] ~ ( t ) = ~t[exp Hl01(t)exp H202(t ) .,. exp H m 8 r (t)].

Of course~ the derivative of the product of exponentials is somewhat messy to


express, depending upon the commutation relations on the H i . It is useful for the
purposes of kinematic programming to express the derivative with T(81,@2,...Sr)

factored out on the left. This yields


126

-H 0 +H @
T(t) = T(t) (Hr@ r + e r rHr_l~r_l e r r ) .

Such formulas have been found to be quite useful in various aspects of applied
mathematics. In this case, they yield a differential equation for the O's.

-HrO +H @ -H 0 -H2O 2 +H202 +H 8


+ e rHr_l e r r~ + .e r r ~ r r
Hr~r r-i . . . . . e ie ...e

Since ~ (3) is six-dimensional, this represents six differential equations.


If r is less than six, we do not expect to be able to solve this equation; if it
exceeds six, we would expect to have non-uniqueness. We begin with a discussion
of the ease r=6. In order to reduce the abo~e equation to something of the form

ii
i i 2 ~ = G(e)

1 61
@2

e6

H -H
we need to simplify expressions such as e r rH ie r r The main tool here is
r-
the Baker-Campbell-Hausdorff formula

eAtBe -At = B + [A,B] + [A,[AIB]]/2: + ...


d~f
exp adAt(B)

where [A,B] = AB = BA. As was emphasized by Wei and Norman [4], repeated use of
this formula allows one to find the matrix G(@).
If the matrices HI,H2,...H 6 are independent elements of the Lie algebra of
IE(3), then G(0) will be invertible for 0=0 and some neighborhood about 0=0: in
fact, the set of @ for which G(8) is not invertible will have measure zero.
However, there will, for topological reasons, always be points such that G(0) is
singular. Near these points, moderate values of ~ may call for very large values
of @, values which cannot be achieved. Thus, the singularities of this map are
quite significant.
Turning now to the case where r exceeds 6, we have the problem of solving the
above equation, ~ = G(@)0, with G(8) being rectangular. In view of the fact that
large values for 8 may be difficult to achieve, it is not surprising that there
have been experiments which use the Penrose inverse G(0) # to get a solution, thus
selecting the solution 0 of minimum length. Despite the attractiveness of this
approach, it can lead to problems. For example, typically robotic devices repeat
the same task time after time. That is to say it needs to repeatedly traverse a
127

closed path in ~ ( 3 ) . If we set 8 = G(B)#e~ along pieces of this closed path,

does a closed path in ~ (3) generate a closed path in IRr? Geometrically speaking,
this is a question about the integrability of the distribution in IR r defined by
Range G(0) #. If this distribution is integrable, then the Penrose inverse approach
will map closed curves in ]E(3) into closed curves in IR r, otherwise it will not.
One can see from the literature [6] that the generalized inverse techniques need
not map closed curves into closed curves.
128

4. CLASSIFICATION OF ROBOT KINEM~%TICS

We may further interpret the product of exponentials formula geometrically.


If we have a r degree of freedom chain and hence have

HI@ I H @
T(t) = Me e r r

how many parameters does it take to give an intrinsic characterization for the
chain? In view of the arbitrary choice of coordinates with which we describe T,
we see that M is not intrinsic and can be ignored in a parameter count. From the
fact that we can replace 8 i by ~iOi + ~i' it appears that an r-link chain should

be described by (r-2)'dim ~ (3) = 4r parameters. In fact, this can be verified by


a direct geometrical analysis and is a well-known result. Thus, there is a 24
parameter family of 6 degree of freedom manipulators.
There are, however, a variety of special circumstances which account for the
majority of robotic designs; again, we refer the reader to Pieper [9]. One can
describe these special cases in terms of geometrical configurations in 3-space or
in terms of the properties of the matrices HI...H 6 which appear in the above
formulas. For example, if we have a string of three successive H's which generate
a Lie algebra isomorphic to so(3), then we see that we have rotations about three
axes which intersect at a eom~on point. This is called a spherical joint by kine-
m~ticians; from the point of view of the product formula, it would~ perhaps, most
naturally be called an so(3)-joint.
A second case of considerable interest corresponds to the situation in which
three successive H's have no rotational part. This occurs in the so-called x-y-z-
robots and results in a considerable simplification of the kinematic programming
equations.
In the case r=6, one could attempt to find a suitably general and completely
intrinsic classification of robotic geometries based on the structure constants
Yijk of the Lie algebra of the H's; i.e., the constants such that

HiHj-HjH i = k~ I TiJ k ~ "

A second possibility would be to base it on the Galois group of the kinematic


equations -- a logical continuation of the work of Pieper.
129

REFERENCES

[i] Richard Paul, Robot Manipulat0rs: Mathematics~ Programming and Contro_!l.


Cambridge~ Mass.: MIT Press, 1981.

[2] Michael Brady, et al., Robot Motion: planning and Control. Cambridge,
Mass.: MIT Press, 1982.

[3] K. T. Chen, "Decomposition of Differential Equations," Math. Ann., Vol. 146


(1962) 263-278.

[4] J. Wei and E. Norman, "On the Global Representation of the Solutions of
Linear Differential Equations as the Product of Exponentials."
Proceedinss of the America nMathematieal Society, 1964.

[5] R. W. Brockett, "Nonlinear Systems and Nonlinear Estimation Theory," in


Stochastic Systems (M. Hazewinkel and J. C. Willems, eds.). Dordrecht,
The Netherlands: Reidel Publishing Co., 1981, pp. 441-477.

[6] C. A. Klein and C. -H. Huang, "Review of Pseudoinverse Control for Use with
Kinematieally Redundant Manipulators," IEEE Trans. on S~stems~ Man~ and
Cybernetics, Vol. SMC-13 (1983) 245-250,

[7] J. M. Herv~, "Analyse strueturelle des m~canismes par groupes de desplace-


ments," Mechanisms and Machine Theory, Vol. 13 (1978) 437-450.

[8] K. H. Hunt, Kinematic Geometry of Mechanisms. Oxford, England: Oxford


University Press, 1978.

[9] D. L. Pieper, The Kinematics %f Manipulation Under Computer Control. Stanford


Artificial Intelligence Laboratory, Stanford University, AIM 72, 1968.

[i0] R. W. Brockett, "Linear Feedback Systems and the Groups of Lie and Galois,"
J. of Linear Algebra and Its Applicati0n_s, Vol. 50 (1983) 45-60.

[ii] J. E. Humphreys, Linear Algebraic GroupS. New York: Springer-Verlag, 1975.

You might also like