By Dimiter Zlatanov, Ilian Bonev, and Clément Gosselin
Département de génie mécanique, Université Laval
Québec, QC, Canada
If you wish to cite the current document, please cite the following paper instead: 
Zlatanov, D., Bonev, I.A., and Gosselin, C.M., "Constraint Singularities as CSpace Singularities," 8th International Symposium on Advances in Robot Kinematics (ARK 2002), Caldes de Malavella, Spain, June 24–28, 2002.

This report continues our examination of the phenomenon of constraint singularity.
We focus our attention on the fact that constraint singularities always represent singular
points of the configuration space of the kinematic chain. As such they separate distinct
regions of the configuration space and may allow transitions between dramatically different
operation modes. All this is exemplified by a fascinating multioperational
parallel mechanism that can undergo a variety of transformations when passing through singular configurations.
^{ }Prototype of the 3URU DoubleY MultiOperational parallel mechanism
1. Introduction
In our previous reports (, )
we defined the concept of constraint singularity. It occurs in a configuration of a parallel mechanism
with n < 6 DOFs, where both the mechanism as a whole and the moving platform have at least n + 1 DOFs.
By that definition, a constraint singularity is always an IIMtype singularity [Zlatanov et al., 1994], i.e., a configuration with increased
instantaneous mobility (IIM) of the mechanism.
IIM singularities are in fact the singular points of the configuration space of the mechanism. In other words, no
neighbourhood of such a configuration in the configuration space is a smooth manifold. The existence of such points
is usually associated with a phenomenon that can be loosely described as "branching" or "selfintersection"
of the configuration space. Thus, a motion starting from an IIM configuration can end up in any one of the two (or more)
"sheets" of the configuration space that meet at the singularity. The behaviours of the mechanism in the neighbouring
nonsingular regions can be quite different. When the configuration space is also a constraint singularity the different
adjoining regions of the configuration space can be expected to sometimes have different possible motions of the
moving platform and, therefore, result in distinct modes of operation of the parallel manipulator.
In the present paper we will discuss in some depth the phenomena arising in the neighbourhood of constraint singularities.
This will be done using the example of a 3DOF parallel robot that can undergo amazing transformations when passing through
a constraint singularity and going from one 3dimensional region of its configuration space to the next.
2. The Configuration Space and the Singularities of a General ClosedLoop Mechanism
The configuration space (or the Cspace) of a mechanism is simply the set of all possible configurations. One can
think of it as the set of all feasible values of the joint parameters. For a serial kinematic chain all joint values are
feasible (within the joint limits). However, if the chain contains closed loops, the possible joint values are only those
that satisfy the loop closure equations. In other words, the configuration space is only a subset of the joint space. In
general, this subset does not have to be a smooth manifold and may have singular points.
The configuration space completely describes a kinematic chain. However, it is completely independent of both the choice
of the input parameters (the actuators) and the output parameters (the space of desired endeffector poses). Hence, to
describe a closedloop mechanism completely, one needs to specify two mappings. The input map must give the
values of the input parameters for every configuration, while the output map defines the values of the endeffector
pose for each point in the configuration space. At a nonsingular configuration, both mappings are smooth and nonsingular.
A mechanism singularity occurs if either there is a configuration space singularity, or else if one of the two mappings
becomes singular. Note that, for a general nonredundant closed loop mechanism, we can speak of an "inputoutput map"
only in the neighbourhood of a nonsingular point in the Cspace.
This most general definition of singularity was the basis of the approach to singularity analysis described in [Zlatanov, 1998].
It naturally leads to the recognition of three fundamental kinds of singularity: configurationspace, input and output
singularities. Further study of the degeneracies of the associated tangent mappings leads to more detailed and comprehensive
classifications [Zlatanov et al., 1994].
Using a similar approach, the three basic kinds of singularity were described in [Park and Kim, 1999]. A word of caution is needed here.
The Cspace, input, and output singularities should not be confused with the three popular singularity types introduced
in [Gosselin and Angeles, 1990], although for some cases they may be the same.
Herein, we will be interested mostly in configuration space singularities and to a smaller extent in output singularities.
We will be concerned mainly with mechanism properties that are determined by the geometry of the kinematic chain and the
choice of the endeffector, but do not depend on the choice of actuators.
3. Constraint and CSpace Singularities as Branching Points
Possibly the simplest interesting example of a Cspace singularity is a flattened fourbar linkage. When all four
links have different lengths, the configuration space is homeomorphic to the figure eight, i.e., from a topological point of
view, it may be described as two circles joined by a single common point. That point is the Cspace singularity and
in its vicinity the configuration space is not a smooth manifold. Yet, one can still define a meaningful "tangent space"
at this point, only this will be a linear space with two rather than one dimensions. The increased dimension reflects the fact
that the configuration is an IIM singularity and the fourbar has a transitory mobility of two. Thus, the configuration space
of the flat fourbar has two onedimensional parts, and the mechanism can switch from one to the other only by passing through
the IIM configuration. This is why Hunt [Hunt, 1978] refers to such singularities as uncertainty configurations.
Figure 1:^{ }Flattened fourbar mechanism in (a) a nonsingular configuration and (b) a constraint singularity.
For a fourbar mechanism (Fig. 1), the two regions of the configuration space can be easily differentiated
by the orientation of the left rotor (the usual input link). In one region the moving end of this link will be above the line
of the base, in the other it will be bellow. Since in the shown chain the link is a rocker, the only transition point is the
singularity in Fig. 1(b).
As it was pointed out in the , the situation in Fig. 1(b) can be
viewed as a constraint singularity if the output link is the coupler. The constraint singularity divides the Cspace
into two nonintersecting nonsingular regions, each being an open and connected set as well as a smooth onedimensional
manifold. The two regions correspond to two different sets of poses of the coupler. However, in every nonsingular point of the
Cspace, regardless of the region, the instantaneous motion will be of the same type: a rotation about an instantaneous
rotation axis. In other words the screw system of the output twists is always the same.
Things become different when all four links have the same length (Fig. 2). In this case, there are three Cspace
singularities, which are also constraint singularities since in each one the coupler has two instantaneous DOFs. The three
points divide the configuration space into six nonsingular regions. At each Cspace singularity four regions meet.
In fact, the Cspace is composed of three smooth closed curves in the joint space. Each curve contains two
Csingular points, which divide it into two regions. Every two of the three curves have a common Csingular point
and at each of the Csingularities exactly two curves meet. In the case of a fourbar the Cspace can be easily
illustrated since it has a faithful projection on a twodimensional torus. Indeed, every configuration can be determined in a
unique way by the value of the angles of the two base joints. Fig. 3 shows this projection. The torus is the space of all
values of the two rotor angles, the (green, red and blue) curves on the torus form the feasible combinations of the two angles for
which there exists a (unique) configuration.
Figure 2:^{ }Fourbar mechanism with equallength links illustrating the notion of constraint singularity.
Figure 3: Configuration space of the above fourbar mechanism.
Unlike the more general flat fourbar in Fig. 1, here two distinct regions of the configuration space may allow completely
different coupler motions. While on (the green) one of the three curves (i.e., in two of the Cspace regions) the coupler remains
parallel to the base, on each of the other two curves (four Cspace regions) the coupler rotates about a point fixed
in the base. At two of the three constraint singularities the mechanism has the options of entering two translational and two
rotational regions, while at the third Cspace singularity the fourbar can choose among four rotational regions. We
could say that the mechanism can be used both as a translational and as a rotational device and the transition between the
two modes of operation may occur when the device goes through a constraint singularity.
It turns out that such "multifunctionality" can be observed in much more complex devices, and the transformations between
different modes are quite fascinating.
4. The 3DOF 3URU DoubleY MultiOperational (DYMO) Parallel Mechanism
Consider the mechanism shown in Fig. 4. It is a parallel device with three 5R or, more precisely, URU legs. The
centres of the three Ujoints at the base form an equilateral triangle. The same applies for the three Ujoints
at the platform. The three Rjoint axes fixed in the base meet in the centre of the base triangle (in a "Y" shape) and the same is
true for the platform Rjoints and the moving triangle. The three intermediate R joints in each leg are all
parallel. A plastic model of the mechanism was built to facilitate its analysis.
Figure 4: Wireframe model of the 3URU DYMO parallel mechanism in its nexus configuration.
If our readers find this design vaguely familiar, they are not mistaken. In essence, it is the same as the design of the
malfunctioning built at Seoul National University (SNU), the
manipulator at the origin of our exploration of constraint singularities. There have been some modifications. Firstly,
the prismatic joints of the SNU robot were replaced by R joints, a kinematically nonessential change of little
practical importance. Secondly, the new design allows the platform triangle to be lowered in and through the base plane,
a change that has no kinematic value whatsoever, but is quite important in view of the new tricks that this old design
will be asked to perform. In view of this slightly modified design and the newly discovered properties, we will refer to
the new manipulator as the 3URU DoubleY MultiOperational (DYMO) parallel mechanism.
We will assume that the two parts of each leg (i.e., the two links in the leg chain with nonzero length) are close but
different in length. We will also assume that they are quite long, and we will not concern ourselves too much with the
configurations where a leg becomes fully stretched or folded. We will also ignore the effects of link interference.
Our interest is in the structure of the configuration space and the nature of the endeffector motion rather than the
workspace limits.
Since there are 15 joints, all revolute, the configuration space of the described chain is a threedimensional subset
(with singular points) in a 15dimensional torus. However, to avoid unnecessary (and possibly overwhelming) complexity,
we will not distinguish between configurations with the same platform pose and identical orientation of the 4th Rjoint
axes. Given a generic pose of the platform there are 4 different solutions for
the posture of the corresponding leg, and hence, there will be 64 different
configurations of the mechanism as a whole. The distinction between these will not be essential for our purposes.
Hence, we can assume that the ambient manifold, where our Cspace is embedded, is the Cartesian product of
SE(3) and three copies of ,
the 1dimensional projective space. (Note that we do not distinguish between angles in the platform
Rjoints that differ by 180°, so the configuration space of the joint is not a circle but the projective line.)
We will now proceed to study the configuration space of the kinematic chain in a way similar to the analysis of the folding
fourbar in Section 3. The goal is to identify the Cspace singularities and the nonsingular Cspace regions.
Moreover, we want to know whether the Cspace singularities are also constraint singularities and what endeffector
motions are possible in each nonsingular region. Furthermore, it is of interest to study the possibilities and forms of
transition between the Cspace regions. It turns out that there are several distinct modes of operation of this
parallel chain with completely different motion capabilities of the platform.
5. A Translational Manipulator – or Two
Since the original design was conceived as a positioning mechanism it is no surprise that the mechanism can be translational
in part of its configuration space (Fig. 5). Translational mechanisms of this type (with two Ujoints in each leg)
were proposed by Tsai [1996]. Their singularities, including constraint (in this case, rotational) singularities were
studied by Di Gregorio and ParentiCastelli [1999].


Figure 5: Still from the animation illustrating the translation mode. 
Let the base and platform have the same orientation, and neither of the legs be singular, i.e., the five Rjoints
in each legs are linearly independent. Then, the constraint imposed by each leg on the platform, i.e., the wrench reciprocal
to all the leg Rjoints, is a pure moment, perpendicular to all five of the Rjoints. The three constraints
will remain linearly independent unless one of two things happens: (a) the platform centre is right above the base centre;
or (b) the platform triangle is lowered in the base plane. The intersection of these two singularity sets should be considered
separately as case (c) where the platform centre is at the base centre (as in Fig. 4). We can conclude that once the mechanism
is assembled with zero platform orientation it can move only as a positioning device until it reaches one of the configurations
defined by conditions (a) or (b).
Case (a) corresponds to the constraint singularity described in .
This is an IIM configuration where both the kinematic chain as a whole and the moving platform have five instantaneous DOFs.
Some of the allowed twists are all translations, as well as all rotations with horizontal axes.
When the platform is in the base plane, case (b), the three constraint moments become coplanar and the platform acquires the
capability to rotate about any vertical axis. The mechanism as well as the platform will generally have 4 DOFs. If the platform
centre is lowered in a baseplane point lying on one axis of a base Rjoint, then the mechanism will still have 4 DOFs
but the platform will lose one freedom and will have only three instantaneous DOFs. Indeed, when the platform centre is on the base "Y"
(formed by the base R axes), but not in the base centre, there is a single continuum of configurations with a fixed platform because
of the spin of one of the legs. However, only one of those configurations, where all three legs lie in the base plane, is a
Cspace singularity. In this case the mechanism has 4 DOFs but the platform only three, two horizontal translations and
a vertical rotation.
Finally, it is important to note that case (c), when the platform and base centres coincide, corresponds to infinitely many
translationmode configurations. The reason is that each of the legs can spin along its axis while base and platform coincide.
However, not all of those configurations, forming a 3dimensional torus, are adjacent to translationmode regions. These points
form a 2dimensional surface on the 3dimensional torus. This situation is further addressed in Section 8.
Note that there are Cspace singularities, of cases (b) and (c), for all positions of the platform centre in the base plane.
Yet, these configurations do not divide the set of nonsingular zeroorientation translationmode configurations into two.
There exist paths that avoid Cspace singularities and join any two configurations on both sides of the base plane. For such paths,
the platform centre passes through the base plane in a point along the base "Y" (but not the base centre). As we pointed out above,
such configurations are not Cspace singularities as long as the single singular leg is not in the base plane. Moreover,
note that each such configuration has a neighbourhood of nonsingular Cspace points. Hence, the nonsingular Cpoints with
zero platform orientation form a connected open set.
It is now time to observe that everything written since the beginning of the present section is valid if "zero orientation" is replaced
with a "180degreeyaw orientation". Hence, there are two translationmode regions. Clearly, the two translation modes (with different
orientations) cannot have common boundary points but, as we will see later, this does not mean that there are no configuration space
paths between the two.
6. An Orientational Manipulator
In our previous paper (see Fig. 1 in ) we used the example of a parallel wrist
with three UPU legs first proposed in [Karouia and Hervé, 2000]. The constraint singularities (in this case, translation singularities)
were first reported by Di Gregorio [2001]. The linkgeometry condition is that the base R joints intersect in a
common point and so do the platform R joints. This condition is satisfied by the DYMO chain as well and therefore it has an
orientation mode of operation (Fig. 6).


Figure 6: Still from the animation illustrating the orientation mode. 
Assume that the base and platform centres coincide and no leg is singular (i.e., no platform R joint is aligned with the
corresponding base R joint), as is the case in Fig. 6. Then, the unique constraint wrench of each leg is a force through
the common centre and parallel to the intermediate R joints of the leg. When the three forces are linearly independent,
the system of the platform twists will be the 3system of all rotations about the centre. When the three constraint forces become
coplanar the platform acquires the freedom to translate in a direction orthogonal to the plane of the constraints. Such a
configuration will then be a Cspace and a constraint singularity since the platform will have at least 4 DOFs (if all
three forces are parallel, which can happen only if the platform is horizontal, the platform acquires two horizontal
translations for a total of 5 DOFs).
We will now consider the configurations where a leg becomes singular and its base and platform Rjoints align.
This adds one more leg constraint: a pure moment perpendicular to the plane of the U joints. This constraint reduces by
one the freedoms of the platform, and then the endeffector may remain with 3 DOFs, even when the constraint forces are coplanar.
However, the configuration is still an IIM singularity, because now there is an extra mechanism freedom not involving the platform.
Indeed, the singular leg can freely spin about the aligned revolute axes even when the platform is fixed.
Either exactly one or all three legs can become singular when the centres coincide. In the first case, the platform always has
exactly two rotational freedoms. If the constraint forces are coplanar (which can be arranged for any of the considered platform
orientations by spinning the singular leg) then the platform can translate as well. The platform has three DOFs and
the mechanism four.
When all three legs are aligned (i.e., the platform has zero or 180degreeyaw orientation), each leg imposes two constraints
– a force and a moment. The configuration is adjacent to an orientationmode region only if some rotations are allowed, i.e.
if the three moment vectors are coplanar. If this is the case, then the mechanism as a whole will have at least 4 DOFs because
of the spinning of the three legs, hence we have an IIM (Cspace) singularity.
We can conclude that the Cspace singularities adjacent to the orientation mode are exactly all those configurations
(with coinciding centres) where the lines defined by the intermediate R joints are parallel to the same plane. It can
be shown that these configurations all have an orientation of the platform with zero yaw, i.e., they are obtained from the zero
orientation by a single (finite) rotation with a horizontal axis (a tilt in the vertical plane of some azimuth).
All such orientations correspond to Cspace singularities.
However, if the axis of the finite (nonzero) tilt is one of the base Rjoint axes, then there are infinitely many configurations
(due to leg spinning) and not all of them lead to coplanar constraint forces and Cspace singularities. Only one of the
orientations of the spinning leg (for every halfturn) corresponds to a constraint singularity, while for any other the chain is in a
nonsingular Cspace point.
The problem of the identification and description of the platform orientations
corresponding to constraint singularities in the orientation mode was solved by using the socalled modified Euler angles
as introduced in [Bonev and Ryu, 1999].
Figure 7:^{ } Modified Euler angles
This formulation allowed for particularly simple final expressions and also made it easy to establish
that the IIM singularities divide the set of nonsingular orientations of the platform into two connected components
(with positive and negative yaw angle, respectively).
However, this does not mean that the surface of Cspace singularities also divides the set of nonsingular configurations
in the orientation mode. The situation is a close analogy to the translation mode and the role played by the configurations with
platform centre in the base plane. In orientation mode as in translation mode, there are nonsingular Cspace paths that can pass
through the singularity surface. These paths cross from positive to negative yaw through any configuration where the tilt
axis is one of the base Rjoints. Moreover, as was the case with each of the two translation modes, the nonsingular orientaitonmode Cspace points
form a connected open set.
The existence of the orientation mode answers one transition question from the previous section. Assuming that the leg length is
sufficient, it is possible to move continuously between two translationmode configurations with opposite yaw angles. The path can
go through the orientation mode, where it can enter with zero orientation and exit with a horizontal platform and a yaw of 180°.
We will see later that this is not the only route between the two translation modes.
7. Two Mechanisms for Planar Motion
When analyzing the translation mode in Section 5 it was mentioned that the platform can be lowered into the base where it will
acquire as a fourth DOF a verticalaxis rotation. In Section 6, it was mentioned that if all three constraint forces are
vertical, while the mechanism is in orientation mode, then the platform acquires two horizontal translations.
In both cases the system of the possible platform twists includes as a subsystem the planar motions. It is not difficult to see
that the ability for planar motion is not transitory and that finite planar motions can originate from the described configurations.
Indeed one glance at the mechanism should be sufficient to immediately recognize the features of a 3RRR planar parallel
manipulator (Fig. 8).


Figure 8: Still from the animation illustrating the planar mode. 
Let the platform be in the base plane (with arbitrary orientation). If no base and platform Rjoint pair is aligned (and
there can be no leg spinning) then all intermediate Rjoints must have vertical axes or, in other words, all legs lie in
the base plane as well as the platform. If no leg is singular, the unique leg constraint will always be either a vertical force
or a horizontal moment. Therefore, the planar motions will always be allowed and this will be a planar mode of operation of the
mechanism.
It must be observed that the above is true not only when the platform is in the base plane "face up" but also
"face down", although then the 3RRR planar parallel mechanism will be less recognizable since the base and platform
attachment points will be in reverse circular order. Also, the leg length must be larger than in the shown model and there will be
link interference.
So there are two planar modes, which obviously do not have common adjacent configurations, yet it is still possible to go with a
continuous motion from one to the other via the orientation mode. We will see later that there is another route as well.
We now identify the Cspace singularities adjacent to the planar modes. The picture in the "facedown" is quite different from
"face up". This is unlike the situation with the two translation modes where there was complete symmetry and all singularities
for the two fixed orientations were essentially the same.
In any mode of operation, when a leg is nonsingular the unique leg constraint is either: a force parallel to the intermediate
Rjoint axes and passing through the intersection of the base and platform Rjoints; or, if the base and platform
Rjoints are parallel, a moment perpendicular to the Ujoints plane.
In planar mode, a constraint singularity will occur when either: (a) the three vertical constraint forces are applied on a common
line in the base plane; (b) the three constraints are horizontal moments; (c) one of the constraints is a moment and the other
two are forces whose plane is perpendicular to the moment. (Note that if two constraints of two legs are moments then the third
leg must have parallel base and platform joints).
First we consider "faceup". Here only case (b) is possible. Hence the Cspace singularities must be among the configurations
with zero or 180° orientation of the platform. When no legs have aligned 1st and 5th joints, the platform has 4 DOFs, adding
the vertical translation to the planar freedoms (Fig. 9). When for one leg the platform vertex is on the base Rjoint axis,
the mechanism has four DOFs but the platform has only the three planar freedoms. When all three legs have their first and last
joints on the same line, i.e., when the platform is at the centre with an orientation of 0° or 180°, the mechanism has
6 DOFs, but the platform can still move only with the three planar DOFs. All of those singular Cspace points are boundary
points to a translation region. The central configuration is adjacent to the orientation mode as well. The Cspace
singularities divide the "faceup" planar mode into two regions (with positive or negative orientation angle).


Figure 9: Still from the animation illustrating the transition between the planar and the translation mode. 
If the platform is facedown, only cases (a) and (c) can occur. It can be shown that these singularities correspond to the
following platform planar poses. For every orientation of the platform the singular locations of the platform centre are
given by a pair of perpendicular lines through the origin. The orientation of this pair of lines changes with the orientation
of the platform, only twice slower. The singularity surface formed by these rotating lines divides the facedown planar mode
into four component regions.
The singularities in type (c) are those obtained when the platform orientation is such that one of the platform joints is either
along or orthogonal to the base joint in the same leg.
In both the (a) and (c) singularities the platform acquires a nonplanar rotation about a horizontal axis. This axis is along the line
defined by the application points of the constraint forces. In case (a) the mechanism has 4 DOFs and so does the platform, in
case (c) the platform has still 4 DOFs but the mechanism has 5.
We should note that when the platform centre is not at the base centre, the constraint singularities of the face down mode do not
border on either the translational or the orientation mode. They allow the mechanism to escape into another mode of operation
which will be described later.
8. Two Lockup Poses Each With Three DOFs
Let us consider the case when the base and platform centres coincide and the platform orientation is either zero (Fig. 10) or
with a yaw of 180° and the platform is horizontal and face up (zero azimuth, zero tilt).


Figure 10: Still from the animation illustrating the lockup mode. 
Since all legs can spin independently, there is a 3torus of such configuration for each of the two platform orientation. For
most points of the torus the mechanism has 3 DOFs but the platform cannot move, in other words we have a triple infinity of
deadpoint configurations where the platform is "stuck".
It is convenient to view this as a separate "mode of operation". In this mode, there are Cspace singularities where the
platform acquires a freedom and can escape the lockup pose.
Each leg imposes two constraints, a force and a moment. There are two types of IIM singularities: (a) when the three moments are
coplanar and (b) when the forces are coplanar.
In case (a) the platform can escape with a rotation about the common centre and enter the orientation mode, while in case (b)
the platform can translate and enter the corresponding translation mode.
There are singularities which satisfy both conditions (a) and (b). For example, when all three legs lie in the base plane, the
platform has 3 DOFs and it can escape with any planar motion.
Another example is the nexus configuration with coinciding base and platform (Fig. 4) and horizontal Ujoint planes. Now all
legs are singular and each imposes two constraints (a force and a moment). However, the moment, in the vertical direction, is
the same for all legs, while all three forces are horizontal. The resulting constraint system consists of all forces in the
base plane and the vertical moment. This is a selfreciprocal 3system and hence the system of the platform freedoms is spanned by two
nonparallel rotations in the base plane and a vertical translation. This configuration is on the border of both the translation
and orientation regions. It is also adjacent to the mode described in the following section.
9. A New Parallel Manipulator With Three Mixed Freedoms
At several occasions in the preceding discussion we described singular Cspace points from which the kinematic
chain can enter a new mode of operation where the platform is neither parallel to the base nor with a fixed point. This is in
fact a fourth type of operational mode with 3 output DOFs (apart from the translational, orientational and planar modes).
In this mode, the mechanism operates as a new (or at least previously unknown to the authors) manipulator with three mixed freedoms.
Consider for example the configurations where the platform centre is right above (or below) the base centre and the platform
is parallel to the base and yawed at an angle of 0° or 180°. From these IIM configurations the mechanism can exit
into the translation modes using three of its five instantaneous freedoms. Can it exit while using the two rotational
freedoms? Or consider the singularity in orientation mode where the platform plane is not parallel to the base and yet
the mechanism acquires an instantaneous translation. Can it exit the orientation mode along this translation? The answer to
both of the above questions is affirmative as can be seen on Fig. 10. Once the possibility is recognized it is easy to find
configurations of the mechanism where the two centres are distinct and the base and platform are not parallel.


Figure 10: Still from the animation illustrating the mixed mode. 
To find the set of all these configurations we can start by looking at the possible instantaneous motions at one such
configuration. Assuming that no leg is singular and that the base and platform intersect, the leg constraint wrench is either:
a force passing through the intersection point of the base and platform Rjoint axes and orthogonal to the
plane of the leg; or, if the latter axes are parallel (they cannot be skew since they are both in the leg plane),
a moment orthogonal to the plane of the leg joints. This implies that if a constraint is a force it must be orthogonal
to the line connecting the base and platform centres since this line lies in all leg planes.
This key observation leads to an important conclusion. It allows us to deduce that an instantaneous translation along
the line joining the centres is always feasible since this direction is orthogonal to all constraint forces.
This purely infinitesimal feature of the motion is instrumental to the description of the global set of platform poses.
Indeed, it implies that from any feasible configuration (with distinct centres) the platform will be able to perform a
finite translation along the shortest path all the way to the origin. Once the origin is reached the mechanism will enter
with a translation into its orientation mode and hence will have the same orientation as in the starting configuration
and all configurations along the way. The point of entry into the orientation mode must allow a translation and hence it
will be a constraint singularity where the three constraint forces are coplanar.
Conversely, all configurations in the mixed mode can be obtained starting from an orientationmode constraint singularity
by translation along the exit vector.
We already described the singular points in the orientation mode. They are obtained with a tilt of the platform about any
horizontal axis (zero yaw). We need to know the direction of the exit vector for each such singularity.
It is the direction orthogonal to the common plane of the constraints and, as it it turns out, it is obtained from the
vertical by exactly half of the tilt of the platform. In other words the exit vector is inclined in the same vertical
plane as the platform normal and it bisects the angle between the platform normal and the central vertical axis.
(This result, just as the description of the orientationmode Csingularities, was obtained with relative ease
due to a symbolic formulation using the modified Euler angles).
Now we can describe the configuration set of the mixed mode. The set is axisymmetric with respect to the central vertical
line. For every azimuth there are configurations with arbitrary tilt between 0° and 180°. For all configurations with
a given tilt angle, the platform centres are along a ray from the origin and at half the tilt angle from the vertical axis.
Thus, along the vertical axis we have configurations with tilt zero; along a ray inclined at 45° are the configurations
with tilt 90°, i.e., with a vertical platform facing away from the central axis; and along a ray in the base plane we
will have a tilt angle of 180°, i.e. the platform is face down and entering the facedown planar mode.
This construction is performed twice, once starting with a horizontal platform with zero orientation and once starting with
a platform yawed at 180° (about the platform normal). In other words for every feasible configuration of the mixed
mode there is one configuration with a platform rotated at 180° about its normal axis. That may sound similar to the
situation with the two translation modes. The difference is that in the mixedmode case there are direct paths between the
two regions. This is not entirely obvious, and we will try to clarify it below.
But first, let us describe completely the constraint and freedom systems at each mixedmode configuration, where the
platform is not parallel to the base. We already discovered one possible platform motion, the translation along
the line joining the centres, which is orthogonal to any constraint force. In addition, now that we have described
the platform poses in the mixed mode, we know that the intersection line of the base and the platform is also
orthogonal to the centrejoining axis. Therefore, since the constraint forces are applied in points on the intersection
line, it follows that all constraint forces lie in one plane orthogonal to the line between the centres and passing
through the intersection line. Furthermore, this plane must always be equidistant to base and platform, since
the plane's normal is inclined at half the angle of the platform normal.
Therefore, the freedom and constraint systems in a mixedmode configuration are given by the same
4th special screw 3system. It is composed of all zeropitch screws in the bisecting plane and the
infinite pitch screws perpendicular to it.
The singular points in the mixedmode part of the configuration space occur along the vertical axis and in the base
plane. Both groups of singularities have already been described. The vertical axis contains points adjacent to the
mixed and the two translation modes. The configurations in the base plane are common with the facedown planar
mode. We recall from Section 6 that the constraint singularities in the face down state correspond to platform
positions along rays from the origin, where every orthogonal cross of rays corresponds to a different orientation angle.
These orthogonal crosses describe how the platform can enter and leave the origin facedown during mixedmode operation.
The platform can end up face down and with an arbitrary orientation angle in the base plane. For every such orientation,
however, it can leave the origin only along trajectories tangent to a unique cross. Starting along one line of the cross,
the platform can be tilted back up until it is face up with zero orientation. Starting along the other line of the cross,
the platform will be tilted back up to a yawof180° orientation. Thus, at any mixedmode configuration, we can
turn the platform around without leaving the mixed mode. We need to tilt down (and translate) until the platform is
face down and at the base centre; next (translate and) tilt up 180° in the orthogonal direction, until we are face up;
and finally, tilt down to the original position, now with the platform rotated halfturn about its normal axis.
This trick provides, as previously promised, new paths connecting the two translation modes, as well as the faceup and facedown
planar modes.
Every mixedmode configuration with the platform centre above the base plane has a double below the base plane. Hence,
there are a total of four mixedmode regions and one can move directly between any two through constraint singularities.
The configuration with the platform face down (horizontal) and at the base centre, and the single singular leg is in the base plane
(a configurations where both the platform and the chain have 5 DOF), is adjacent to all four regions.
10. The Answer to the Original Question
It is now time to recall that the whole issue of constraint singularity was raised by an "upright configuration"
of the mechanism, i.e., a configuration with the platform centre above the base centre.
We already know that this is a constraint singularity, where the mechanism has five instantaneous
freedoms and therefore no choice of three actuators can prevent the platform from moving instantaneously.
The actuated Pjoints in the Seoul robot correspond to the middle (third) Rjoint in the DYMO chain.
When these actuators are locked, as we showed in , there are two
remaining instantaneous freedoms: a pencil of horizontal rotations passing through the point of the central vertical axis
where the lines of the legs (the lines connecting the corresponding base and platform vertices) meet.
However, it was still in some doubt whether there can be a finite motion with locked actuators.
The analysis of the mixed mode can provide a proof that the answer is negative.
Let us assume that there exists a finite continuous motion with locked actuators starting from an upright configuration.
The initial twist of this motion must be in the 2system of allowed freedoms with locked actuators, namely, a horizontal
rotation through the legintersection point. This cannot be done in a translation mode, hence, it must be a motion in the
mixed mode. Therefore, there must exist a continuous motion in the mixed mode which ends in an upright configuration and
for which the twist tangent to the motion in the final pose is a horizontal rotation through some given
point on the central vertical axis.
However, a twist in the mixed mode cannot leave the 4th special 3system associated with each configuration.
It will always remain a linear combination of: the translation along the twocentres line; and two rotations in the
bisecting plane. In the course of the motion the twist will change continuously and so will the three basis twists.
How will these basis twists change when the motion approaches the upright configuration and, therefore,
the tilt angle approaches zero? Clearly the translation will converge to a vertical translation;
and the rotations will remain coplanar and end up in a limit plane. It is easy to see that when the tilt angle approaches zero
the bisecting plane will converge to a horizontal plane equidistant from the base and platform.
Hence, the motion can exist only if the legintersection point is exactly at the midpoint between the platform and base centres.
This can happen only if base and platform triangles are identical in size and we are in the 180degreeyaw
translation mode. Since neither is true for the Seoul configuration, it follows that there is no
"architecture singularity".
The above does not mean that there can be finite motion for identical base and platform at some
twisted configuration, only that firstorder infinitesimal motion analysis at such a configuration cannot
rule it out.
The apparent motion is, therefore, caused by the joint clearances of the mechnanism.
It should be said that the theoretically impossible motion of the platform with locked legs
is very noticeable both in the Seoul prototype and in the DYMO plastic model.
12. Conclusion. Modified definition of Constraint Singularity
In our study of the DYMO architecture we reavealed that the configuration space
is composed of multiple threedimensional connected open sets separated by surfaces of Cspace
singularities. At last count, there were at least 14 such regions, and the analysis of
the lockup mode has not yet been finalized. The 3dimensional regions allow for
five fundamentally different types of motion: translational, rotational, mixed, planar, and no motion at all.
And, most amazingly, a transition between any two regions or modes is possible without dissassembly.
The studied architecture has been of particularly high educational value in
revealing the complexities of the workspace geometry and topology. However, there is no doubt
that similar behaviour is exhibited by many other mechanisms and in particular by parallel
manipulators with reduced freedoms. In fact, as we showed in Section 3, the described phenomena
can be observed, albeit in a much simpler form, in a fourbar mechanism. (Which
possibly confirms the conjecture that a kinematician can learn everything he or she needs to know
from a fourbar).
The transition path joining different modes must always encounter a configuration space singularity.
At each such configuration, a dooorway opens allowing the kinematic chain to enter a
new domain and, sometimes, undergo dramatic transformations.
In most of the examined cases, these singularities were identified as constraint singularities since
the platform had the ability to instantaneously move with more than three degrees of
freedom. Yet, sometimes the platform remained with three DOFs, although the singularity
and the possible transitions did involve the motion of the endeffector.
Possibly the best example is the nexus configuration in Fig. 1, where the mechanism as
a whole has 6 DOFs but the platform has only 3 DOFs, one translation and two horizontal
rotations. On the other hand, this configuration is a gateway between the translation and orientation
mode, and one could describe the transformation as a tradeoff in progress of rotational and translational
freedoms, a process which, in the configuration itself, is only partially completed. Moreover, it
could be argued that here the platform does indeed gain extra freedoms, but the three
singular legs reduce the total to three.
In view of these consideration it may be a good idea to somewhat modify the definition
of a constraint singularity, so that such configurations would be covered as well.
A constraint singularity would then occur when an nDOF parallel mechanism acquires at least
n + 1 DOFs and the endeffector has n  d+ 1 DOFs, where d is the total number
of additional constraints imposed by leg chains that are singular in the configuration.
Clearly, although the basic nature and some principle consequences of the phenomenon
of constraint singularity have been revealed, many open questions remain.
Acknowledgements
We would like to thank PierreLuc Richard for building the model of the 3URU DYMO parallel mechanism
in Pro/ENGINEER and importing it into ADAMS. We also appreciate his efforts and those of Thierry Laliberté
for building the plasticwooden model using rapid prototyping and some manual work. Although the model
seemed to be a complete fiasco at the beginning (due to link flexibility and joint clearance), at the end, it proved to
be essential for understanding the motion of this intricate mechanism.
Replicas of the model can be ordered from the
.
Bibliography

Bonev, I. A., and Ryu, J., 1999, "Orientation Workspace Analysis of 6DOF Parallel Manipulators",
ASME Design Engineering Technical Conferences (DETC'99), Las Vegas, NV, September 1215.


Di Gregorio, R., and ParentiCastelli, V., 1999, "Mobility Analysis of the 3UPU
Parallel Mechanism Assembled for a Pure Translational Motion," Proc. 1999 IEEE/ASME International
Conference on Advanced Intelligent Mechatronics, Atlanta, GA, pp. 520525, September 1923. 

Di Gregorio, R., 2001, "Statics and Singularity Loci of the 3UPU Wrist,"
IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM'01), Como, Italy, July 811. 

Gosselin, C., and Angeles, J., 1990, "Singularity Analysis of ClosedLoop Kinematic Chains,"
IEEE Transactions on Robotics and Automations, Vol. 6, No. 3, pp. 281290. 

Hunt, K. H., 1978, Kinematic Geometry of Mechanisms, Oxford University Press. 

Karouia, M., and Hervé, J. M., 2000, "A ThreeDOF Tripod For Generating
Spherical Rotation," Advances in Robot Kinematics, J. Lenarcic and M. M. Stanisic (eds.), Kluwer Academic Publishers,
pp. 3954020. 

Park, F.C., and Kim, J.W., 1999, "Singularity Analysis of Closed Kinematic Chains,"
ASME Journal of Mechanical Design, Vol. 121, pp. 3238. 

Tsai, L.W., 1996, "Kinematics of a ThreeDOF Platform With Three Extensible Limbs,"
Recent Advances in Robot Kinematics, J. Lenarcic and V. ParentiCastelli (eds.), Kluwer Academic Publishers,
pp. 401410. 

Zlatanov, D., Fenton, R. G., and Benhabib, B., 1994, "Singularity Analysis of
Mechanism and Robots Via a MotionSpace Model of the Instantaneous Kinematics," Proc. IEEE International
Conference on Robotics and Automation, San Diego, CA, pp. 980985. 

Zlatanov, D., 1998, "Generalized Singularity Analysis of Mechanisms,"
Ph.D. thesis, University of Toronto, Canada. 
