Modelling and Control of a Modular, Redundant

Robot Manipulator
Sylvie CHARENTUS and Marc RENAUD
Laboratoire d'Automatique et d'Analyse des Syst~mes (LAAS)
7, avenue du Colonel Roche 31077 TOULOUSE C6dex (FRANCE)

Abstract

A new type of robot manipulator which possesses rigidity capabilities and a large
workspace is presented. This robot is composed of several modules which consist
of elementary manipulators with the parallel structure of the "Stewart Platform".
However, the structure of our new robot manipulator raises modelling difficulties.
First, one parallel module is studied and a new method of computation of the direct
kinematic model is proposed. Then, the computation of the complete direct and
inverse kinematic model of our robot manipulator is presented.

1 Introduction
Robot manipulators with an open kinematic chain structure [1] offer the advantage
of usually sweeping a wide workspaee but are intrinsically inadequate for positioning and
orientating accurately their end effector; indeed, position and orientation errors at the
different joints increase from the base outward to the end effector. Also, they suffer a
relative lack of rigidity which leads to a small useful mass ratio (weight carried by the end
effector over robot mass in the order of 1/35 for electric robots [2]). To overcome these
difficulties, various robot manipulators using a closed kinematic chain structure [1] have
been proposed or are being developed [3] [4]. Among these are the parallel manipulators
that employ segments to connect two links, one of which is fixed and serves as a base and
the other is mobile and used as end effector. These robots are instrinsically adequate for
positioning and orientating the mobile link relative to the fixed one because the different
position and orientation errors at the level of the different joints are not cumulative. In
addition, they are sufficiently rigid, leading to a useful mass ratio that sometimes exceed
1 (90 in the case of INRIA's prototype [2]). On the other hand, these robots only sweep
a limited workspace.
This paper presents a new, globaily-closed, kinematic chain structure. It combines
the ability to sweep a wide workspace of an open chain structure with the accuracy and
rigidity capabilities of parallel structures. In particular, this structure associates in series
n identical parallel structures. However, this new structure raises difficulties in modelling,
the latter being an essential step in control synthesis. The difficulties encountered in com-
puting both the inverse geometric model (often called inverse kinematic model) of open
 509

chains and the direct geometric model (often called direct kinematic model) of parallel
structures arise at the same time. To overcome these drawbacks, the selected parallel
structure is first investigated and a new method of computation of the direct geometric
model of this structure is proposed. Then, the computation of the various kinematic mod-
els required for the control of the overall structure is surveyed and subsequently applied
to a particular robot manipulator possessing this structure and comprising a series of four
Stewart platforms. This robot, called robot LX, has been designed and realized by the
LOGABEX company.

2 Study of parallel robot manipulators

2.1 General

Parallel structures have been the focus of much attention on the part of mathemati-
cians like Euler and Cauchy who determined the minimum number of segments ensuring
rigidity as well as the maximum number required to compute the forces exerted on the
segments as a function of the force and torque applied onto the mobile link relative to
the fixed one (in this case the structure is said to be isostatic). Thus, six segments are
needed to obtain a rigid, isostatic structure with six degrees of freedom. Numerous au-
thors attribute the first realization of such a structure to Stewart who devised, in 1965, an
aircraft simulator employing three complex segments fitted each with two electrical linear
jacks. In the proposed structure, the two links are called platforms and the structure
itself is sometimes referred to as Stewart's platform. Other constructions using six linear
actuators could be cited in which the six linear actuator end points are connected in pairs
at the corner of an equilateral triangle on at least one platform [5] [6] or in which the
linear actuator end points are connected such that there are six connection points on the
two platforms [7] [8]. Inouc [9] has utilized three sets of pantographs in licu of the six
linear actuators.
The advantage of this construction lies in the accuracy and rigidity capabilities. In
addition, it is composed of simple modular elements based on the actuator and plat-
form size and on the type of actuators selected, that is, electric, hydraulic or pneumatic.
Combined with an effort sensor and a hybrid force-position control, it is well suited for
assembly operations and may be used as a wrist on a robot [10] [7]. On the other hand,
this mechanism is rather difficult to model as detailed in the following.

2.2 Modelling

Modelling difficulties of this type of structure are the inverse of those encountered with
open structures. Indeed, it is the analytical computation of the direct geometric model
which allows achievement of the location (position and orientation) of the mobile platform
as a function of the actuator lengths, which is difficult whereas the analytical computation
of the inverse geometric model which allows expression of the actuator lengths as a function
of the mobile platform location, is straightforward.
 510

2.2.1 Computation of t h e inverse geometric m o d e l

An orthogonal frame is associated with each platform: frame T/o(O0, x0, yo, Zo) is linked
to the lower platform and used as reference and frame TQ(O1, x l , y l , zl) is linked to the
upper platform as shown in Fig.(1). Points O0 and O1 are the circle centres wherein the
linear actuator end points are situated. Points Ai (/3/respectively) are linked to the lower
platform (upper respectively).

B4
B5 b B3

A6
A1 A2

Figure 1: Placement of frames on the platforms

The calculation of the six linear actuator lengths (li, i = 1 ~o 6) is easily obtained from
the upper platform location (position of point O1 and orientation of frame 7E1 relative to
frame T~):
Since:
AiBi = - O o A i + 0~0~ + 01Bi for i = 1 to 6
it is also possible to write for the components:

AiBqo) = -OoAqo) + 0o01(o) + Rol OIBi(1) for i = 1 to 6

where Rm is the orientation transformation matrix from frame TOo to frame T~I and
subscripts (o) or {1) denote the cartesian components in frames 7~o or 7E1.
Then:
nil AiBq0)H
If X represents six independent components defining the upper platform location (ob-
tained from 0o0~(o) and Ro~) and if L = (11...16), the inverse geometric model is repre-
sented by the transformation L=f(X) which is unique.
 511

2.2.2 C o m p u t a t i o n of t h e d i r e c t g e o m e t r i c m o d e l
This is a complex computation. Several researchers have put forward an iterative
solution involving a Jacobian matrix [8],[2]. In this paper 1, we present a solution which
reduces (at least in the particular case where the linear actuator end points are connected
in pairs at the corner of an equilateral triangle on at least one platform), the computation
of the direct geometric model to that of the roots of an n order polynomial equation. This
solution leads to the value n = 8 (asserted by Merlet [11] but has not been proved yet)
and improves our preceding result leading to the value n = 16 [12]. Also, this solution
rectifies some wrong assertions recently published in the literature [13]. Finally it may
be noticed that the literal expression of the 8 order polynomial equation is explicitly
obtained. Bcfore presenting this solution, the iterative method using a Jacobian matrix
is recalled.

I t e r a t i v e m e t h o d using a J a c o b i a n m a t r i x . The direct geometric modelis computed

from the explicit inverse geometric model L=f(X) presented in the preceding paragraph.
A classical Newton-Raphson algorithm can be used, the Jacobian matrix being the
partial derivative matrix of the six linear actuator lengths relative to the six independent
components defining the location of the upper platform : J ( X ) = ~aLv -( /X'~ - From an initial
location arbitrarily chosen, the current location is modified at each iteration step in order
to obtain that location which corresponds to the actuator lengths.
The method requires three or four iterations in the case where it converges and there-
fore involves a computation time which is a function of the number of iterations. How-
ever, this may be a drawback for real-time control. A refinement consists of comput-
ing only one constant Jacobian matrix corresponding to the mean location, i.e., when
0oO1(0) = (0, 0, zm~o,) and Rol = I (identity matrix). This increases the number of ite-
rations but speeds up computation. However it can only be used if the final location is
more or less similar to the mean location [8].
The drawback of this iterative method is that it only yields one solution associated
with the initialization of the procedure, (i.e., the choice of the initial location). Usually,
it does not hinder the control of these manipulators because designers intentionally limit
the workspace so as to ensure method convergence [14].

M e t h o d p r o p o s e d . The new solution is proposed in the case of a manipulator where

the six linear actuator end points are connected in pairs at the corner of an equilateral
triangle on at least one platform. However, we only deal with the simplified case depicted
in Fig. (2) where there are three linear actuator end points on the two platforms [12].
In fact the problem is to find the cooordinates of points Ho(xo, Yo, Zo), Hl(x~, y,, z,),
H2(x2, y2, z2) in frame T~, knowing the lengths Ii (i=l to 6).
Six relationships are associated with the actuator lengths:

t,=llB0n211 , t=-IIB, n~ll , t ~ - I I B , Holl

/4=IIB~Holl , t s = l l B ~ n , II , t~=IIBoH111
'The results presented in this part differ from those revealed in Montreal during the symposium and
have been considerably improved.
 512

x0 x2
y ~ t I-I2 z2Y2

14 11

B0
Figure 2: Simplifiedmodel

that is,
dB 2
z~ = +d + (y2 - --~, + d (i)

t~ = (=o + --
~s) ~ +(yo+~--~,'+B~ ~ +Zo~ (3)

l.~ = (~o - )~ + (yo + ~-~

d. ) + ~o~ (~)
l~ (=~ riB) ~ dB ~
= --- +(+.+~-~, +d (5)

t~ = =~ + (y, - --~j + z~ (o)

and three relations are asssociated with the geometryof the equilateral triangle Ho, HI, H2:
II HoH~I1=11H1H~ II=ll H2Ho II= dH
that is,
d~=(=2-~,)~+(y~-y,)~+(~-~,) ~ (7)
 513

d~ = (x, - =o)~ + (v, - ~o)2 + (z, - ~)~ (8)

= (~o - ~ ) ~ + (~o - ~ ) ~ + (~o - ~=)~ (9)

Eqs (1) and (2) are the cartesian equations of a circle in space whose axis is BOB1;the
radius c2 being given by:
d(i, + l~)~ - ,~. ~/,~. - (l, - l~)~
c2 = 2 dB

It is more convenient to represent this circle by three parametric equations using

for example, as parameter, the angle 02 between vector zo and plane BoB1H2 (Fig.3)
(Oo• ] - ~, +~]):

x2 = - -~-c2 sin 02 - ~(a2 + as) (lO)

c2 . v~ d.)
y2 = ~ - s m 02 - -T-(a2 - -- (11)
Z2 = C2 C O S 02 (12)
with
a2 = da

H2

12
n

02
go
BI

dB

B2

B0

Figure 3: Definition of angle 02

Eqs (3) and (4) are the cartesian equations of a circle in space whose axis is B1B2; the
radius Co being given by:
 514

~/(t~ + ~,)~ ' 4 ~/,e, - (t3 - l,) 2

Co = 2 ds

This circle can be represented by three parametric equations using for example, as
parameter, the angle Oo between vector zo and plane BxB2Ho (00 E ] - ~r, +r]):

X0~ T
ao (13)
dB
~o = - c o sin Oo - 2 ¢ 5 (14)

Zo = co cos Oo (15)
with

ao = dB

Eqs (5) and (6) are the cartesian equations of a circle in space whose axis is B2Bo; the
radius cl being given by:

~/(ts + t0) 2 - ~. ~/~ - (~s - to) ~

cl = 2 d8

This circle can be represented by three parametric equations using for example, as
parameter, the angle 01 between vector z__qand plane B2BoH1 (01 E ] -- a', +Tr]):

• , = -7-c~ sin o~ + (d~ - al) (16)

Yl = ~-sin01 +-T-(a1 + ) (17)

Z1 ---~e I COS 01 (18)

with

al-- dB

Notice that if angular sensors transmitting the value of angles 00, O,, 0~ axe used, the
computation of the direct geometric model can be effected analytically. This solution has
been retained by Inoue [9] who proposed a platform differing from that of Stewart. If the
values of 00, 01,02 are not known, the computation is continued as follows:
The values of coordinates x~,yi, zi(i = 0,1, 2) obtained in Eqs (10) to (18) are substi-
tuted into the 3 Eqs (7), (8) and (9). A system of 3 coupled trigonometric equations with
3 unknowns 0o, 01, 02 is obtained:

sin 01 sin 02 - 2 cos 01 cos 02 + q2 sin 01 + Pl sin 02 = R0 (19)

sis 02 sin 00 - 2 cos 02 cos 00 + q0 sin 02 + P2 sin 00 = R1 (20)
sin 0o sin 01 - 2 cos 00 cos 01 + ql sin 00 + p0 sin 01 = R2 (21)
 515

with
v~(as - a,) dg(aB + ~)
Pi~" 2el , qi= 2ca for i = 0,1,2
and
17o = ~c2 (d2H + d2n (l~ + 312e+ 31~ + q) (lg - l])(l~ - lg))
c W" - 4 - 4~

1 d2v (l~ + 31~ + 31] + l]) (l~ - q)(i~a - 1~))

Rt= c2co(~q 4 4 -- 4d~
R2= 1 .(~q ~ (t]+3l,~+31~+l]) _ (g-t~)(l~-t]))
cocl 4 4 4d~
To solve the system of Eqs (19), (20) and (21), the symmetry is suppressed and we
proceed in several steps:
* From Eq (20), the value of cos 02 can be express as:

cos 02 = sin 02(sin 0o + qo) + p2 sin 0o - R ,

(22)
2 cos 02
and from the relationship cos2 0r + sin s 02 = 1, we deduce:

NI sin 2 02 + N2 sin 02 + Na = 0 (23)

in which Nx, N2, N3 are 2nd order polynomials in sin 0o only.

• The two other Eqs (19) and (21) can be written in the form of a 2nd order linear
equations system with 2 unknowns cos 01 and sin 01:

( - 2 cos 0o) cos 01 + (sin 0o + P0) sin 01 = - q l sin 00 + R2 (24)

( - 2 cos 0o) cos 0t + (sin 02 + q2) sin 0~ = - p t sin 02 + Ro (25)
whose solution is given by:

COS 01 = (sin 02 + q2)(R2 - qt sin 0o) - (sin 0o + p 0 ) ( ~ - px sin 02)

(26)
2((sin 0o + ~ ) cos 02 - (sin 02 + q2) cos Oo)
and
sin01 = (R2 - q, sin 0o) cos 02 - (Ro - p, sin 02) cos 0o
(27)
(sin Oo + vo) cos 02 - (sin O~ + q2) cos Oo

and from the relationship cos2 02 + sin s 0r = 1, we deduce:

Gt sin 2 02 + G2 cos 2 02 + Gs sin 02 cos 02 + G4 sin 02 + G5 cos 02 + Gs = 0 (28)

in which G1, G2, G4, G6 are 2nd order polynomials in sin 0o only, and G3 = cos OoG'3,
G5 = cos OoG's with G[, G's 1st order polynomials in sin 00 only.
 516

Using once again cos2 02 + sin 2 02 = 1, Eq (28) can be witten in the form:
(GI - G2) sin 2 0~ + G4 sin 82 + G2 + Ge + (G3 sin 0o + Gs) cos 02 = 0 (29)

and using Eq (22), Eq (29) can be written in the form:

cos Oo(M1sin 2 02 + M2 sin 02 + Ms) = 0 (30)

or since cos O0 is not always a solution, in the form:
M1 sin 20z + M2 sin #2 + 3/3 ~- 0 (30
with
M1 = 2(G1 - G2) + (sin0o + qo)G~
M2 = 2G4 + (sin 00 + qo)G's+ (p2 sin00 - R,)G~3
M3 = 2((72 + Gs) + (p2 sin 0o - R1)G~
M1, M~, Mz are 2nd order polynomials in sin 00 only.
Now, it sufllcez to show that there exists a common root to the Eqs (23) and (31).
This can be accomplished by equating the resultant of these two equations to zero:

M1 M2 M3 0
A(sin00)= 0 M1 M2 M3
N1N2 N3 0
0 NI N2 N3
= (M1N3 - M3NI) 2 - (MlJV2 - M2NI)(M21V3 - M3N2) = 0

Since M~ and N~ (i = 1, 2, 3) are 2nd order polynomials in sin 0o only, A(sin 0o) is
an 8 order polynomial in sin 3o only.
On the one hand, this demonstrates that the maximum number of real solutions
sin$o in the interval [-1, 1] is 8. On the other hand, it is shown in the following
that this maximum number can be obtained.
Since two values of cos 0o can be associated with each value of sin 0o (cos 0o =
V/1 - sin 2 0o and cos$o = -~/1 - sin 2 $o), the maximum number of reM solutions 00
in the interval ] - ~r, zc] is 16.
• At each value of 0o corresponds a unique value of 0] and 02 in the interval ] - ~r, zr]:

- sin02 = M~N,-M1N~
M I Nrt - M2 N I
and cos02 given by Eq (22).
- cos 0~ given by Eq (26) and sin 01 given by Eq (27).

Then the maximum number of reachable locations is 16 and the example shows here-
after that this maximum number can be obtained. Notice that the two locations corre-
sponding to the two opposite values of cos #0 are symmetrical with respect to the lower
platform, as shown in [16], and such that the coordinate z0 is either positive or negative.
Balrstow's numerical method has been used to compute the roots of this polynomial
equation.
 517

Example:
C o m p u t a t i o n of t h e p l a t f o r m locations based on t h e six a c t u a t o r lengths a n d t h e
m o d u l e geometry.
Let:
11 = 23.0000 , 12 = 23.5000 , /3--24.5000

14 = 24.0000 , 15 = 22.5000 , Is = 22.0000

dH-----6V/3 , ds=13v~

Computation of constants:

aO ffi 1.0770 al = 0.9882 a2 = -1.0326

cO = 2 1 . 4 7 2 9 cl ffi 1 9 . 1 8 6 7 c2 = 20.3374
pO = 0.8647 pl = 0.9717 p2 = 1.0028
qO = 0.9516 ql = 1.0609 q2 ffi 0.9149
RO = - 2 . 0 2 2 9 R1 = - 2 . 0 7 3 7 R2 = - 2 . 0 5 8 9

Computation of the 8 order polynomial coefficients:

degree 8 : 4490.4291
degree 7 : 13549.0803
degree 6 : 16123.8510
degree 5 : 9416.4307
degree 4 : 2650.0159
degree 3 : 240.1017
degree 2 : -32.3998
degree 1 : -6.9443
degree 0 : -0.2695

Computation of the polynomial equation roots using Bairstow's numerical method:

real part imaginary part

-0.1921 0.0000
-0.2089 0.0000
-0.0573 0.0000
-0.6034 0.0000
-0.6226 0.0000
-0.7243 0.0000
-0.7388 0.0000
0.1300 0.0000

T h e 8 solutions are therefore:

2This example is derived from that of3.P. Merlet.

 518

sinOo GO yo zo xl Yl zl x2 Y2 z2
-011921 0J5385 -2.3756 21.0731 -6.4733 -3.1668 13.4436 -7.0558 4.6699 20.2441
-0.2089 0.5385 -2.0139 20.9990 7.8413 5.0977 18.9754 7.6732 -3.8340 13.6653
-0.0573 0.5385 -5.2707 21.4377 5.3519 3.6604 19.1867 -4.9753 3.4687 20.3322
-0.6034 0.5385 6.4560 17.1239 -4.8294 -2.2177 15.1360 5.5217 -2.5918 15.9815
-0.6226 0.5385 6.8692 16.8033 -4.5683 -2.0670 15.3661 -9.3734 6.0079 19.8053
-0.7243 0.5385 9.0525 14.8056 9.8485 6.2566 18.4806 2.0058 -0.5619 18.4676
-0.7388 0.5385 9.3642 14.4711 9.4822 6.0451 18.5935 1.0387 -0.0036 18.9428
0.1300 0.5385 -9.2910 21.2907 -4.0665 -1.7773 15.7827 6.2410 -3.0071 15.2912

These 8 locations are illustrated in Fig. (4).

Figure 4 : 8 locations corresponding to the same combination of actuator lengths (with Zo

positive)

C o n c l u d i n g r e m a r k o n t h e d i r e c t g e o m e t r i c m o d e l c o m p u t a t i o n : The approach
proposed presents a great advantage: it yields all the solutions of the direct model and
then deletes the convergence problems associated with iterative procedures.

2.2.3 Singularity study

The concept of parallel manipulator singularities is not so precisely defined in the

literature. In fact, for this type of manipulator, singular locations must be defined in the
same way as singular configurations are defined in the case of series manipulator.
 519

A parallel manipulator possesses a singular location when the mobile platform is no

longer rigid, that is, when it is unstable for a given combination of actuator lengths. This
is what occurs, for example, when the upper platform is rotated 90 degrees about the
z axes, both platforms being parallel; a slight screwing motion can then be performed
through an external force about the z axis on the upper platform without changing the
actuators lengths [5].
These singular locations can be detected by either:
• analytically or numerically cancelling the determinant of the Jacobian matrix de-
fined previously or of the matrix of P15ker's coordinates for the six linear actuator
axis [5]. The practical implementation cannot be achieved in the first case [2], it
is certainly easier in the second case using Pl~ker's coordinates in an appropriate
frame [15].
• geometrically using Grassmann's line geometry. The rank [16] of each linear actuator
axis subset must be computed. Degeneracy cases are obtained when the rank of this
subset is less than the number of axes it consists of. These cases correspond to the
parallel manipulator singular locations [17].
Then, the singular configurations are obtained from the singular locations using the
direct geometric model.

2.3 Workspace and control

2.3.1 W o r k s p a c e characterization
The workspace is limited by several constraints:
• mechanical constraints due to linear actuator travel, joint rotation and size of the
different manipulator parts;
• geometric constraints due to linear actuator end point positions on the platforms;
• geometric constraints due to singularitie avoidances.
The geometric characterization of the workspace generated by this type of parallel
manipulator is not easily obtained. The concept of aspect introduced by Borrel [18]
for open structures, can be employed, providing it is generalized to closed structures.
Indeed, the location workspace would then be obtained by the union of aspects. In
this particular case, an aspect is a connected open set of the location workspace
in which the sign of the Jacobian matrix determinant is constant. The aspects
are therefore separated by a boundary which constitutes a singularity locus (de-
terminant of the Jacobian matrix equal to zero). Thus, determining the number of
aspects should permit to obtain the maximal number of direct geometric models. In
other words, a singularity locus in the Iocation space separates two distinct locations
of the upper platform ( one in each aspect) that correspond to the same actuator
lengths [19].
The avoidance of singularities could then be solved by limiting the location workspace
to that of a single aspect.
 520

2.3.2 Control
It may seem simple, in the first place, to carry out the control of this type of paral-
lel robot manipulator since the inverse geometric model that allows computation of the
actuator lengths as a function of the upper platform location is explicit. However, the
analysis of this structure must be continued to highlight the existence of singularities
that are awkward for the control. Thus, if the trajectory to be followed in the location
workspace lies exactly in the same aspect, then the singularity problem will not exist at
the level of control.

3 A p p l i c a t i o n of parallel structures: r o b o t m a n i p u -
lator c o m p o s e d of n parallel s t r u c t u r e s c o n n e c t e d
in series
3.1 Presentation
A parallel robot manipulator (referred to in the following as a module) possesses a
small workspace. But if several modules are stacked in series, the manipulator will have a
greater workspace and will maintain the major characteristics of a module, that is, rigidity,
ease of construction and ability to lift heavy loads. These manipulators consisting of n
parallel structures connected in series are then highly redundant since they possess 6n
degrees of mobility ( six per modules), this number exceeding the six degrees of freedom
of the end effector as soon as at least two modules are assembled. This redundancy
complicates control synthesis but nonetheless enables the robot, on the one hand, to
avoid the possible obstacles in its workspace and, on the other, to reconfigure itself for
continuing the task in case of the actuator becoming blocked.
These robot manipulators may be considered from two viewpoints. If it is assumed
that the 6n generalized coordinates are the 6n actuator lengths, they then possess a closed
chain structure and computation of their direct geometric model is highly complex. On
the other hand, if the 6n generalized coordinates are assumed to be the 6n location
coordinates of the n upper platforms of each module, then~ they have an open chain
structure - each module appearing as a complex joint - and computation of their direct
geometric model is straightforward. This is why this second viewpoint has been retained
in the following paragraph.

3.2 Modelling
In the following:
• operational coordinates refer to location coordinates of the end effector in frame T~0
linked to the robot base. They are represented by vector X E R 6.

• generalized coordinates stand for the location coordinates of the upper platform
of each module. These generalized coordinates define the overall configuration of
the robot. They are represented by vector Q = L~l,q2,...,q,] E / ~ n where the 6n
generalized coordinates qlj (j = 1 to 6) are the components of the n generalized
 521

vectors 5~(i = 1 to n) and where n is the number of modules composing the robot.
An ortho$onal frame TQ is associated with each upper platform of each module i.

3.2.1 C o m p u t a t i o n of t h e direct geometric m o d e l

This model consists of computing the location of the robot end effector as a function of
the locations of the upper platforms of each module. This computation is straightforward
and can be effected by introducing the notion of homogeneous transformation matrix (one
per module) [15].
The location of the end effector ( regarded here as the upper platform of the last
module) in frame T~ is given by the homogeneous matrix T0,n:

To,. = ...

where ~_l,i(ql) , is the homogeneous transformation matrix from frame TQ_I to frame
7~i, function of the generalized coordinates qij.
X can be obtained from T0,. and this model is classically denoted by X= f(Q).

3.2.2 C o m p u t a t i o n of t h e inverse geometric model

This model consists of computing the locations of the upper platforms of each module
as a function of the end effector location. Of course by reason of redundancy there exists
an infinity of solutions. Given the complexity of the problem due to the number of
mobilities involved, the analytical computation seems difficult. We therefore propose an
iterative numerical method relying on the computation of a Jacobian matrix.
The iterative algorithm is given in Fig. (5).
Remarks:

• The Jacobian matrix of dimension 6x6n is derived from the direct geometric model
X=f(Q) in the form:

OX
J(Q) = --~(Q)

• Among the set of generalized inverse matrices of J, we selected the pseudo-inverse

matrix [20] given by:

Ji.. = j r ( j j,)-i

Matrix (jdt) is square, of dimension 6 and it must be regular.

• At each iteration k, one has: QI~ = Q~-I + OQk.
The generalized coordinates must satisfy the mechanical constraints of each module,
that is, they must belong to the workspace of a module.
Prior to dealing with robot control, the diagram depicted in Fig. (6) summarizes the
modelling difficulties for such a robot.
 522

~.~.~__ opetstltmal cmordinates Xd'~

----.._.. / "
Computation of the operational coordlnates X=f(Q)
(robot direct geometric model)

1
Robot configuration{2

1
Computation of the Jacob/an matmt J

IComi~,~m.tbm of the generalized coordlnateJt I

Q+Iid

Possible modification of the generalized ~ d l n a m ~ ]

ff they are outside the module worhlmCe
I
1
Computation of the ~4n~tion~ coordinates Xff(Q)
(tabor direct geometric model)

[
Figure 5: Iterative algorithm for the computation of the robot inverse model

3.3 Workspace and control

3.3.1 Workspace
Neither the analytical expression of the location workspace of a module nor, conse-
quently, that of the robot are known. The latter is of course a function of the number of
modules and of the workspace of each module.

3.3.2 Control
To carry out position control, the robot inverse model together with the inverse model
of each module must be employed.
From Fig. (6) robot singularities may exist at different levels. In the aforementioned
study of the parallel structures, we have highhghted singularities associated with the
 523

MODULE DIRECT MODEL ROBOT DIRECT MODEL

(difficul0 Co~y)
Actuator lengths . Locationofplafform I
of module 1 relative to the base

Actuator lengths Location of platform 2

of module 2 * relative to platform 1
End cffe~tor location

Actuator lengths Location of platform n

ofmodulen J' relative to platform n-1

MODULE INVERSE MODEL ROBOT INVERSE MODEL

(easy) (difficult)
Actuator lengths ~ Locationofplafform l .t
of module 1 relative to the base

Actuator l~gths Location of platform 2

of module 2 ~ mlativo to platform 1
E~d ¢ffcctor location

Actuator lengths ~ Location of platform n :

of module n relative to platform n-I

Figure 6: Different models for a robot manipulator composed of n parallel structures

geometry of the structure itself. These naturally reappear during assembly in series of
several modules; the best solution is to limit the workspace of each module in order to
reject these singularities outside the domain. But other singularities of a kinematic nature
may occur as a result of the use of the inverse matrix.

3.4 Application: LOGABEX's robot LX

3.5 Presentation of the robot LX

LOGABEX's robot manipulator LX is depicted in Fig. (7). It consists of 4 identical
modules fitted with 24 linear electrical actuators able to sustain each a 100 daN force.
They are controlled by a d.c. motor, the output velocity of the actuator rod being 6
mm/s. This 120 kg robot can displace a weight of about 50 kg over the total workspace
or a heavier weight in a more restricted area.

3.6 Simulation
Fig. (8) represents the cross section of the position workspace of a module.
One configuration of the robot LX is depicted in Fig (9). The program computes a
 524

rTTT]-~-v-rTT_V~.~_$TTTVT-T-T-T-r TT]TT-I
LI J I I I I I I . ~ ~ . ~ . J _ L I I II I ! !! !!
t l I I I II.l~f'/'Xv~~
ltl : 'I',IIIII
I I I 1_IL--'~[r~-~pTi I t I I t f-'I"r.,I~le~4k~bLl t I 1 1

-------~1 I I I I I I I"1_'~1t I I

]tl]]]ll ~ ~ '
l lital I~ ~ , . : _L

Figure 7: LOGABEX's robot LX

Figure 8: Position workspace of a module

robot configuration, that is, the generalized coordinates and the actuator lengths from
the desired operational coordinates and is usually run for n modules.

3.7 Test implementation

The control program is implemented on the robot site. It operates on a multi-task,

real-time, micro-computer INTEL 310.
We have controlled this robot comprising 1, 2, 3 or 4 modules and it may be noticed
that the computation time for the robot inverse model increases with the number of
modules. At this stage, the hardware and software facilities involve computation times
that are inadequate for real-time control.
 525

tt

Figure 9: One configuration of the robot LX

4 Conclusion
In this paper, the major characteristics of a parallel robot manipulator have been
highlighted, i.e., rigidity, ease of construction and ability to lift heavy loads. In particular,
LOGABEX's modular, redundant robot manipulator LX which constitutes an original and
patented application, has been presented.
The modelling of Stewart's platform as well as that of the robot LX, which is needed for
control synthesis, has also been surveyed. Our contribution has primarily been concerned
with a new method of computation of a module's direct geometric mode[ which furnishes
all the solutions and the computation of the robot inverse model involving an iterative
procedure utilizing a Jacobian matrix.
Nevertheless, several areas must still be investigated especially the modelling of such
major domains as the characterization of the workspace and that of singularities. As
to the robot, this redundant structure is interesting in ordcr to implement and test new
controls, notably in a clustered environment [21] [22].

5 Acknowledgments
The authors are grateful to Logabex 3 for providing the pictures and a movie of the
robot LX.

3Logabex 3, avenue Didier Daurat 31400 Toulouse FRANCE

 526

References
[1] I. Artobolevski. Th~orie des mdcanismes et des machines. Technique Sovi6tique.
Editions Mir Moscou, 1977.

[2] J.P. Merlet. Parallel manipulators.part i : theory,design,kinematics,dynamics and

control. Rapport de recherche No 646, Institut National de Recherche en Informatique
et en Automatique (I.N.R.I.A.), March 1987.

[3] C. Gosselin. Kinematic analysis, optimization and programming of parallel robotic

manipulators. PhD thesis, McGill University, Montreal, Canada, June 1988.

[4] V. Hayward and R. Kurtz. Modeling of a parallel wrist mechanism with actuator re-
dundancy. Technical report, McGill Research Center for Intelligent Machines, McGill
University, Montr&~l, Canada, January 1989.

[5] E.F. Fichter. A Stewart Platform-based manipulator : general theory aaad practical
construction. The International Journal of Robotics Research, Vol.5(No 2):157-182,
Summer 1986.

[6] H. McCallion and P.D. Truong. The analysis of a six-degree-of-freedom work station
for mechanised assembly. In Proceedings of the Fifth World Congress on Theory
of Machines and Mechanisms, pages 611-616. The American Society of Mechanical
Engineers, July 1979.

J.P. Merlet. Contribution k la formalisation de la commande par retour d'efforts en

robotique. Application h la commande de robots parallNes. Th~se de l'Universit~,
Paris 6, June 1986.

[8] C. Reboulet. Techniques de la robotique. Architectures et commandes. Chapitre 8 :

Moddlisation des robots parall~les. Hermes, 1988.

[9] H. Inoue, Y. Tsusaka, and T. Fukuizumi. Parallel manipulator. In Robotics Research:

The Third International Symposium, O. Faugeras and G. GiraIt (Eds}, MIT Press,
Cambridgej Massachusetts, pages 321-327, October 1985.

[10] C. Reboulet and A. Robert. Hybrid control of a manipulator equipped with an

active compliant wrist. In Robotics Research: The Third International Symposium,
O. Faugeras and G. Giralt (Eds}, MIT Press, Cambridge, Massachusetts, October
1985.

[11] J.P. Merlet. Force-feedback control of parallel manipulators. In IEEE International

Conference on Robotics and Automation, Philadelphia (USA), pages 1484-1489, April
1988.

[12] S. Charentus, C. Diaz, and M. Renaud. Modular serial parallel redundant robot. In
Preprlnts IMA CS on System Modelling and Simulation, Getraro Italy, pages 57-63,
September 1988.
 527

[13] P. Nanua and K.J. Waldron. Direct kinematic solution of a Stewart Platform.
In IEEE International Conference on Robotics and Automation, Scottsdale (USA),
pages 431--437, May 1989.
[14] A. I~bert. Commande hybride position-force. Mise en oeuvre et exp&imentation
sur un micro-manipulateur parall~le. Th~se de docteur ing~nieur, Ecole Nationale
Sup6rieure de l'Adronantique et de l'Espace, December 1986.
[15] B. Gorla and M. Renaud. ModUles des robots-manipulateurs. Application ~ leur
eommande. CEPADUES Ed. Toulouse. France., 1984.
[16] K.H. Hunt. Kinematic geometry of mechanisms. Oxford engineering science series.
Clarendon Press, Oxford, 1978.
[17] J.P. Merlet. Parallel manipulators.Part 2 : Theory.Singular configurations and Grass-
mann geometry. Rapport de recherche No 791, Institut National de Recherche en
Informatique et en Automatique (I.N.R.I.A.), February 1988.
[18] P. Borrel. Contribution b. la mod61isation g&~m~trique des robots manipulateurs.
Application b, la conception assistde par ordinateur. Th~se d'dtat, Universit~ des
Sciences et Techniques du Languedoc, Montpellier, July 1986.
[19] T. Berthomieu. Etude d'un micro-manipulateur parall~le et de son couplage avec
un robot porteur. Th~se de docteur ingSnieur, Ecole Nationale Sup6rieure de
l'A6ronautique et de l'Espace, January 1989.
[20] C.A. Klein and C.H. Huang. Review of pseudoinverse control for use with kinemati-
cally redundant manipulators. IEEE Transactions on Systems, Man, and Cybernet-
ics, Vol.SMC-13(No.3):245-250, march/april 1983.
[21] B. Espiau. An overview of local environment sensing in robotics applications. In
NATO advanced Research Workshop: Sensors and Sensory Systems for Advanced
Robotics, Maratea (Italy), April 1986.
[22] T. Yoshikawa. Manipulability and redundancy control of robotic mechanisms. In
IEEE International Conference on Robotics and Automation, St Louis (USA), pages
1004-1009, 1985.