Gimbal Dynamics

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

© JASON OSBORNE

DYNAMICS
AND CONTROL he double-gimbal mechanism (DGM) is a multibody mechanical

ON THE TORUS
T device composed of three rigid bodies, namely, a base, an inner
gimbal, and an outer gimbal, interconnected by two revolute
joints. Figure 1 shows a typical DGM, where the cylindrical
base is connected to the outer gimbal by a revolute joint, and
the inner gimbal, which is the disk-shaped payload, is connected to the
outer gimbal by a revolute joint. The DGM is an integral component of
an inertially stabilized platform, which provides motion to maintain line
of sight between a target and a platform payload sensor [1], [2]. Modern,
commercially available gimbals use two direct-drive or gear-driven
JASON OSBORNE, GREGORY HICKS,
motors on orthogonal axes to actuate the joints. Many of these mecha-
and ROBERT FUENTES nisms are constrained to a reduced operational region, while more

Digital Object Identifier 10.1109/MCS.2008.924794

44 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008 1066-033X/08/$25.00©2008IEEE


sophisticated models use a slip ring to allow continuous design problems considered in [1] and [2] are not dis-
rotation about an axis. Angle measurements for each axis cussed here, we present a geometric approach to the
are obtained from either a rotary encoder or a resolver. The DGM that provides a solid foundation for the global
DGM is a fundamental component of pointing and track- mechanical behavior of this device.
ing applications [3] that include missile guidance systems, This article is organized into three main sections, “Kine-
ground-based telescopes, antenna assemblies, laser com- matics and Topology,” “Metric Geometry,” and “Designing
munication systems, and close-in weapon systems (CIWSs)
such as the Phalanx 1B shown in Figure 2.
The design, modeling, and control of a DGM are non-
trivial tasks. Engineering requirements on the speed of
response, pointing accuracy, size, weight, and cost must be
addressed by a suitable design. Physical modeling of a
design can range from simple linear models to extremely
complex nonlinear models that incorporate dynamic
effects such as structural flexibility, stiction, backlash, and
cogging. Control design must take into account not only
the usual performance and robustness considerations but
also the physical constraints and limitations imposed by,
for example, an inertially stabilized platform.
A proportional-integral-derivative (PID) control strate-
gy for the DGM is used in many pointing and tracking
applications. In particular, when the tracking application
requires only a restricted range of motion near a nominal
orientation, a local, linear control design may be sufficient
to achieve a tracking objective. However, for global, large-
angle controlled motions of the DGM, PID control design FIGURE 1 The double-gimbal mechanism (DGM). The DGM is a
can result in undesirable inner and outer gimbal coupled multibody mechanical device composed of three rigid bodies and
motions. These undesirable motions suggest that PID con- two revolute joints. A typical DGM contains a cylindrical base con-
trol is ineffective beyond its local region of applicability. nected to the outer gimbal by a revolute joint as well as an inner
gimbal, the circular mirror, which is connected to the outer gimbal by
As addressed in this article, the solution to this global con- a revolute joint. Global analysis of the DGM reveals features that
trol problem requires an understanding of the global are relevant to control issues, such as gimbal lock and limitations of
behavior of the DGM. The DGM exhibits a richness of linear control design.
global topological and non-Euclid-
ean geometric properties that pro-
vide crucial information about its
global controlled dynamics [4]. A
global description of the DGM
thus requires a departure from the
Euclidean space commonly used
in engineering.
The goal of this article is to
provide an accessible introduc-
tion to relevant topological and
geometric topics, which are wide-
ly discussed in the physics litera-
ture [5], [6] but less so in applied
engineering texts. Our focus is the
idealized DGM, which we use as
the primary example to illustrate
general topological and geometric
concepts. This mechanism has no
stops or impediments to prevent FIGURE 2 Raytheon’s Phalanx Block 1B close-in weapon system. The Phalanx 1B system is a
double-gimbal mechanism (DGM) composed of a 20-mm M61A1 Gatling gun, advanced
the joints from rotating freely search-and-track Ku-band radar, and an integrated forward-looking infrared (FLIR) system. The
through 2π and more radians. close-in weapon system provides naval defense against antiship missile, surface craft, and low-
While many of the practical speed aircraft threats. Photos courtesy of Raytheon Company.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 45


Physically Inspired Controls.” The first section addresses crucial in understanding why the gimbal mechanism cannot
topological information that can be obtained from the be globally asymptotically stabilized by means of continu-
DGM’s topological configuration space. This information is ous control logic. In the second section, a geometric layer is
added to the gimbal topology. While topological considera-
tions of the DGM address the possibility of achieving con-
trol objectives, the mass distributions of the mechanism’s
components are key to choosing a natural control logic.
Specifically, the metric geometry discussions aid in deter-
mining the direction and magnitude of control inputs for
inertia-based, proportional-derivative (PD) control design.
The final section links the previous sections by using topo-
logical and geometric information to view the double gim-
bal as a pointing and tracking mechanism. One issue
discussed in this section is the problem of gimbal lock. In
short, this tutorial article provides mathematical clarification
and physical insight into global issues regarding gimbal
motion and control designs that take into account the natur-
FIGURE 3 Tracking in a circular corridor. The problem of tracking al tendencies of the gimbal.
on S1 can be viewed as a game of chase between two objects,
Red (the tracker) and Blue (the evader), confined to a circular
corridor. A natural control logic for Red is to assess the distance
KINEMATICS AND TOPOLOGY
to Blue in both directions, compare them, and then accelerate in The DGM cannot be globally asymptotically stabilized
the shorter of the two directions. This logic is problematic when by means of continuous control logic [7]. The underlying
Red and Blue are both stationary and antipodal to one another. In cause of this limitation lies with the kinematic con-
this case, Red’s control logic is undefined since the distances to straints on the motions of the gimbals imposed by the
Blue in both directions are equal. Red’s indecisiveness can be
solved with a coin flip, where heads means go left and tails
revolute joints. For an intuitive explanation of how kine-
means go right. While the introduction of a coin removes the matics affect global control design, we give the following
indecision problem, a problematic discontinuity is created. That example.
is, at the point of indecision, Red may have to instantaneously Consider two objects, Red and Blue, each with constant
change direction from left to right. This situation is caused by the mass and both confined to the unit circle as shown in Fig-
kinematic constraints of the circular corridor, which are related to
the topological holes of the configuration space.
ure 3. The unit circle, denoted S1 , is the set of points in R2
given by

S1  {(x, y) = (cos(θ), sin(θ)) : θ ∈ (0, 2π]}. (1)

Suppose Red’s objective is to track down Blue. An obvi-


ous control strategy for Red is to assess the distance to
Blue in both directions and accelerate in the shorter direc-
tion. A problem arises when Red and Blue are both at rest
φ and at opposite sides of a diameter, that is, antipodally
positioned. In this case, Red’s control logic is undefined
since there does not exist a unique shortest direction to
Blue. Red’s indecisiveness can be solved with a coin flip,
where heads means go left and tails means go right. While
the introduction of a coin removes the indecision problem,
a problematic discontinuity is created. That is, at the point
of indecision, Red may have to instantaneously change the
direction of its acceleration from left to right.
Red’s control troubles stem from being confined to the
circular corridor S1 without being able to traverse the inte-
rior of the circle, which can be viewed as a hole in the plane
θ of the circle. Although the hole of the circle is the source of
Red’s difficulty, from the perspective of Blue, the hole is a
FIGURE 4 Rotations of the double-gimbal mechanism (DGM). The
welcome feature, which serves as a physical obstacle
positive angles φ and θ represent counterclockwise rotations of the behind which to hide and evade capture. As discussed
DGM about the inner and outer gimbal axes, respectively. below, holes are also a problem for controlling the DGM.

46 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


The Configuration Space of the DGM where θ0 ∈ (0, 2π − θ0 ) and φ0 ∈ (0, 2π − φ0 ) are varia-
When viewed as separate, free rigid bodies, the spatial tions in the outer and inner gimbal axis angles, respective-
configuration, that is, the orientation in Euclidean three- ly. Having the notion of nearness in a topological space,
dimensional space, of each of the three components of the we gain access to the techniques of [8] and [9], which can
DGM are unconstrained. However, since the components be applied to the DGM.
are assembled into the DGM using the revolute joints, and, The continuous control logic obstruction of the circular
since the base is fixed to, for example, the ground, the spa- corridor and, by extension, the DGM are related to the
tial configurations are kinematically constrained. In fact, topological property of connectedness as defined in [8].
for the assembled DGM shown in Figure 4, the kinematic Essentially, two subsets of a topological space are discon-
constraints of the revolute joints restrict the gimbal nected when none of the points of one subset are near the
motions to rotations through the angles θ and φ about the other subset. Two subsets that are not disconnected are
outer and inner gimbal axes, respectively. said to be connected. Although the circular corridor S1 is
For the idealized DGM, which has no mechanical stops to connected, within the context of R2 the circle is connected
prevent the bodies from rotating through 2π and more radi- in a way that divides the remainder of the plane, that is,
ans, the angular displacement of each gimbal can be repre- the complement of S1 , into an “inside,” the bounded part
sented as a point on the circle S1 . Consequently, the that we refer to as a hole, and an “outside,” the unbound-
configurations of both gimbals together can be represented ed part. Consequently, the rest of the plane is disconnected
by the set T2 = S1 × S1 . Since each point on T2 corresponds by the way the circle is connected. The inside of S1 , the
to a pair of gimbal angles θ and φ that determine the config- hole, interferes with a straightforward control law.
uration of the DGM, the configuration space of the DGM is Since the product topology on the torus T2 is defined
homeomorphic to the torus in Figure 5. We focus solely on using the induced topology on the circle S1 ⊂ R2 , and
the idealized DGM and its configuration space T2 for the because the topological space S1 contains holes, we expect
remainder of the article. the topological space T2 to also contain holes, which
obstruct global continuous control design for the DGM.
Analysis of the DGM Kinematic Constraints As shown in [7], compactness of a configuration space
While the kinematic constraints of the DGM are encapsu- without boundary, which, for a topological configuration
lated in its configuration space, that is, the set T2 , a topolo- space contained in R3 , is equivalent to the configuration
gy T [8] on T2 is required to solidify the intuitive notion of space being closed and bounded, is a topological property
holes introduced in the circular corridor example above.
A topology T on a set X is a collection of subsets of X
that satisfy the union and intersection properties given in
[8], [9], and [11]. Define the set

O(θ0 , θ0 )  {(cos(θ), sin(θ)) : θ0 < θ < θ0 + θ0 }, (2)

where θ0 ∈ (0, 2π) and θ0 ∈ (0, 2π − θ0 ), and where we


slightly abuse notation by referring to θ0 as both a geomet-
ric angle and an angle measurement. The meaning of this
notation can be inferred from context. The set O(θ0 , θ0 ) is
an open subset of S1 relative to the induced topology on
S1 ⊂ R2 [8]. In fact, the collection of all open arcs of S1 form
a basis for this topology, and the standard topology for the
set T2 = S1 × S1 is obtained from the product topology [8].
With this topology, the set T2 and its elements gain a notion
of nearness as defined in [9]. Essentially, an element p of a FIGURE 5 Double-gimbal mechanism (DGM) configuration space.
set X is near a subset A ⊂ X when the intersection of every Since the rigid bodies comprising the DGM are interconnected by
open set containing p with A is nonempty. By the definition revolute joints, which impose kinematic constraints on the motions
of the gimbals, only rotations through the angles θ and φ about the
(2) of O(θ0 , θ0 ), the point p0 = (cos(θ0 ), sin(θ0 )) ∈ S1 is
inner and outer gimbal axes, respectively, are allowed. For the ide-
near the subset O(θ0 , θ0 ) for all θ0 ∈ (0, 2π − θ0 ). alized DGM, that is, a DGM without stops or impediments to prevent
For the torus T2 , the gimbal configuration correspond- either of the gimbals from rotating through 2π rad and beyond, the
ing to (θ0 , φ0 ) is near the subset of configurations angular displacement of each gimbal can be represented as a point
on the circle S1 . Consequently, the configurations of both gimbals
together can be represented by the set T2 = S1 × S1 . Since each
O(θ0 , θ0 , φ0 , φ0 )  {(θ, φ) : θ0 < θ < θ0 + θ0 and point on the torus corresponds to two gimbal angles that determine
the configuration of the DGM, the configuration space for the DGM
φ0 < φ < φ0 + φ0 } , (3) is homeomorphic to the torus shown above.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 47


that is sufficient to prevent a mechanism, specifically, the ascertaining the possibility of global asymptotic stabiliza-
DGM, from being globally asymptotically stabilizable tion by smooth control logic.
using continuous control logic. The Betti numbers of a topological space X, denoted by
Compactness of the configuration space is not, howev- βk (X) and defined in [9] and [10], are topological proper-
er, a necessary condition for a mechanism to not be global ties that indicate the presence of k-dimensional topological
asymptotically stabilizable. Consider, for example, the holes. The topological holes are obstructions to global
pendulum and cart system, wherein the pendulum is asymptotic stability. The Betti numbers are also called con-
attached to the cart by means of a revolute joint and the nectivity numbers [9].
cart is connected to the lab bench by means of a prismatic The circle S1 has two Betti numbers given by β0 (S1 ) = 1
joint. As in the DGM case, the configuration space of the and β1 (S1 ) = 1. The Betti number β0 (S1 ) = 1 indicates that
pendulum is S1 , while the configuration space of the cart is the circle is connected. From the perspective of connected-
R1 . Consequently, the configuration space of the pendu- ness, β1 (S1 ) = 1 indicates the maximum number of points
lum and cart system is R1 × S1 . This configuration space is that can removed from the circle while leaving the remain-
represented, both as a set and as a topological space, by a ing space connected. Removing one point of the circle
right circular cylinder. Because of the component R1 , the leaves a punctured circle, which can be viewed as a line
cylinder is infinite in extent and therefore is not compact. segment and which is connected. However, removing a
Nonetheless, because of the component S1 , there is a hole second point from the circle, that is, removing one point
in the space that acts as an impediment to global asymptot- from the line segment, yields two disconnected line seg-
ic control objectives by means of smooth control logic. ments. From the hole perspective and within the context of
Consequently, the holes, rather than the compactness, of a R2 , the Betti number β1 (S1 ) = 1 is an indication of the hole
topological space are the key features to focus on when created by the circle, which divides the remainder of R2 ,
that is, the complement of S1 , into an outside, the
unbounded part, and an inside, the bounded part, which
we refer to as a one-dimensional hole.
The DGM configuration space has three Betti numbers
given by β0 (T2 ) = 1, β1 (T2 ) = 2, and β2 (T2 ) = 1. The Betti
number β0 (T2 ) indicates that the torus is connected. Simi-
lar to the circle, the Betti number β2 (T2 ) = 1 indicates that
the surface of the torus divides R3 into an inside, which is
(a) (b) the bounded part called the two-dimensional hole, and an
outside, which is the unbounded part. The Betti number
β1 (T2 ) = 2 indicates that the torus has two distinct
one-dimensional holes. From the perspective of connected-
ness, β1 (T2 ) = 2 indicates that a maximum of two “cuts’’
can be made along closed curves on the torus while leav-
ing the remaining set connected. By cutting along the red
circle shown in Figure 6, the resulting surface shown in
Figure 6(b) can be straightened into a cylinder, which is a
connected set. Cutting along the white circle shown in
Figure 6(a), which is also a vertical line on the cylinder, the
resulting surface shown in Figure 6(c) can be straightened
into the plane shown in Figure 6(d), which is still connect-
(c) (d) ed. However, by making a third cut along the black curve
on the torus shown in Figure 6(b), which is also the black
FIGURE 6 Illustration of the Betti number β1 (T2 ) = 2. The Betti line in Figure 6(d), the plane becomes disconnected.
number β1 (T2 ) = 2 indicates that the torus has two distinct one-
dimensional holes. From the perspective of connectedness, METRIC GEOMETRY
β1 (T2 ) = 2 indicates the maximum number of “cuts’’ that can be
made along closed curves on the torus and yet leave the remain-
We add to the DGM topological configuration space a
ing space connected. By cutting along the red circle in (a), the metric, which defines an intrinsic distance function that
resulting surface in (b) can be straightened into a cylinder, which determines control inputs that are consistent with the
is a connected set. By cutting along the white circle in (a), which mechanism’s mass distribution. For an intuitive example
is also a vertical line on the cylinder, the resulting surface in (c) and explanation of how the mass distribution affects con-
can be straightened into the plane in (d), which is still a connect-
ed set. However, making a third cut along the black curve on the
trol design, we add an extra layer to the control decision-
cylinder in (c), which is also the black line in (d), the plane making process for the chasing problem along the circular
becomes disconnected. corridor. Suppose now that the masses of the red and blue

48 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


beads change in correspondence with their position along A metric g can be written in coordinate form as
the corridor. The changing mass of Red and Blue can be
represented as hills along the corridor shown in Figure 7. g(vq , wq ) = vqT M(q)wq , (5)
The presence of hills requires that we modify the constant
mass, or level corridor, control strategy. where q = (θ, φ) is the coordinate representation of a
Recall that, in the level case, an obvious control logic is point q in the configuration space T2 , and
for Red to assess the distance to Blue, in both directions, vq = [v1 (q) v2 (q)] is a coordinate representation of a vec-
determine which path is shorter, and then to proceed in tor vq in the tangent space to the torus at q embedded in
that direction. However, in the hilly scenario, a modifica- R3 as shown in Figure 8. The coordinate vector wq is
tion to this decision-making process is required. Namely, defined similarly to vq . Furthermore, for each q, M(q) is a
Red cannot rely solely on the circular arc distance but positive-definite matrix. When M(q) is the configuration-
rather must use a geometric notion of distance that dependent mass matrix of the DGM and we include a
accounts for its changing mass. Essentially, if Red mea- scalar of 1/2, the metric g from (5) is called the kinetic ener-
sures that one direction is closer in circular arc distance gy metric of the DGM.
but also perceives many intervening hills, then it may be Consider now, and for the remainder of this article, a
better for Red to take the longer but less hilly route to DGM that is perfectly balanced in both gimbal axes, which
intercept blue. means that the center of mass of each gimbal lies along its
Even in the hilly situation there are still points of indeci- respective spin axis. Furthermore, assume that the princi-
sion. Confinement to a circle, which is a kinematic con- pal frame of the inner gimbal includes unit vectors direct-
straint, guarantees the existence of these points. Unlike the ed along its axis of rotation and the line-of-sight (LOS)
level corridor, however, the points of indecision, that is, direction determined by the unit vector
points of discontinuous control logic, are generally not ⎡ ⎤
cos(φ) cos(θ)
those positions antipodal to the target. We return to the
LOS  ⎣ cos(φ) sin(θ) ⎦ . (6)
points of control indecision in the control section.
sin(φ)
The main point of this example, which carries over to
PD control design for the DGM, is that a natural measure Given these assumptions, the mass matrix of the DGM
of distance on a mechanism’s topological configuration reduces to
space is one that takes the mass distribution of the mecha-
nism into account when determining the directionality of a
control impetus. This measure is achieved by using the
mass distributions of the mechanism’s components to
define an energy metric on the mechanism’s configuration
space. This metric on T2 provides the extra structure need-
ed to construct the intrinsic distance function necessary to
define the configuration error for an inertia-based PD con-
trol logic for the DGM.

Metric Geometry of the DGM


The configuration space of the DGM, that is, the torus T2 ,
can be realized as the image of the mapping
T : R × R → R3 given by
⎡ ⎤
(R + r cos(φ)) cos (θ)
T(θ, φ) = ⎣ (R + r cos(φ)) sin(θ) ⎦ , (4) FIGURE 7 Tracking in a circular corridor with hills. Tracking on the
r sin(φ) Riemannian manifold S1 can be viewed as a game of chase
between two objects, Red (the tracker) and Blue (the evader), con-
where θ, φ specify the orientation of the outer and inner fined to a circular corridor within which Red’s mass varies as a func-
tion of its location. The changing mass of Red can be represented
gimbals, respectively, as shown in Figure 4. While the torus as hills along the circular corridor. Unlike the tracking control logic
has the structure of a two-dimensional topological subman- presented in Figure 3, Red must not rely solely on the circular arc
ifold of R3 [11], [12], the torus does not yet have a metric distance but rather must use a geometric notion of distance that
geometry structure that accounts for the inertia properties accounts for its changing mass. That is, Red needs to update its
of the gimbals. To remedy this shortcoming we introduce a control logic with the concept of intrinsic distance, which depends
on a description of hills along the corridor. A Riemannian geometry
metric g that defines a mass- and configuration-dependent induced on the configuration space by the mass distributions of the
weighted dot product of velocity vectors for each point q on mechanism’s constitutive bodies allows Red to define the required
T2 [11], [12]. intrinsic distance.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 49


d σ (b)

2 d σ (b)

wr

φ 1
wr
vq
vq d σ (a)

d σ (a) 0
dτ 0 1 2
θ
(a) (b)

FIGURE 8 Configuration and coordinate plane notation. (a) The black configuration curve is denoted by σ (τ ), where the starting point of the
curve is the configuration q = σ (a), and the ending point is r = σ (b). The yellow vector vq and the green velocity vector (dσ /dτ )(a) are
both vectors in the blue tangent plane to the double-gimbal torus (DGT) at the configuration q. Similarly, the red vector wr and the blue
velocity vector (dσ /dτ )(b) are both in the green tangent plane to the DGT at r . (b) The black coordinate curve σ (τ ) is a curve between
q = σ (a) and r = σ (b) for which σ (τ ) = T (σ σ (τ )), where T is the mapping given by (4). Corresponding in color to the vectors
vq , (dσ /dτ )(a), wr , and (dσ /dτ )(b) on the DGT are the vectors in the (θ, φ) coordinate plane denoted by vq , (dσ σ /dτ )(a), wr , and
(dσσ /dτ )(b), respectively. The correspondence between, for example, the configuration vector vq and the vector vq is given by
DT(q)vq = vq , where DT(q) is the Jacobian of T evaluated at q.

  defined by the mass matrix (7) is a Riemannian manifold


m11 (q) m12 (q)
M(q) = [12], [13] that we call the double-gimbal torus (DGT).
m12 (q) m22 (q)
An intrinsic distance function between each pair of con-
 
Iz + Jz cos2 (φ) + Jx sin2 (φ) 0 figurations q, r on the DGT is defined using the energy met-
= , (7)
0 Jy ric defined by the mass matrix (7). Let σ (τ )  T(σ (τ )) be a
curve on the torus such that σ (a) = q and σ (b) = r, where T
where Iz and Jx , Jy, Jz denote the moments of inertia of the is defined by (4), and where σ (τ ) = [θ(τ ) φ(τ )]T is a curve
outer and inner gimbals, respectively, about the outer gim- in the gimbal coordinate plane as shown in Figure 8. For
bal axes of rotation, the LOS, the inner gimbal axis of rota- the parameter τ ∈ [a, b] define the arc length of a curve seg-
tion, and the remaining principal axis of the inner gimbal. ment σ (τ ) on the DGT by
The topological manifold T2 along with the energy metric
 b dσ dσ
lσ (q, r)  g , dτ
a dτ dτ

 b 2 2
dθ dφ
= m11 + m22 dτ , (8)
a dτ dτ

where dσ σ /dτ = [dθ/dτ dφ/dτ ]T is the velocity vector to


the coordinate curve σ (τ ). Given the arc length function
lσ (q, r), the intrinsic distance between the gimbal configura-
(a) (b)
tions q = σ (a) and r = σ (b) is defined to be
FIGURE 9 Illustration of the intrinsic distance function on the double-
gimbal torus (DGT). Let σ (τ ) = T (σ σ (τ )) be the red curve on the dist(q, r)  inf lσ (q, r), (9)
σ
DGT, where σ (τ ) = [θ(τ ) φ(τ )] = [0 τ ] and where T is given by (4).
Furthermore, let q = T (σ σ (0)) and r = T (σσ (π)) be the configura- where the infimum is taken over all smooth curves σ con-
tions of the double-gimbal mechanism (DGM) represented by the
necting q and r. See Figure 9 for an example of the intrinsic
starting point of all the curves in (a) and the ending point of all the
curves in (b), respectively. The arc length l σ (q, r ) of the curve σ (τ ) distance function on the DGT.
is computed by (8) to be Jy π . Of all the other curves between q For the DGT the “inf” in (9) can be replaced by “min.”
and r , the red curve yields the smallest of all the lengths, and there- That is, for each pair of gimbal configurations, there is at
fore the intrinsic distance between q and r is Jy π . That is to say, least one curve segment connecting the elements that
for the DGT, the “inf’’ in (9) can be replaced by “min’’ since, for each
yields the intrinsic distance between them. These curves
pair of elements of the DGT, there is at least one curve segment
between the configurations q and r , the red curve in the above are called minimal geodesics. In general, geodesics are
case, that yields the intrinsic distance. only critical points of the energy functional [14]

50 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


 d2 φ cos(φ) sin(φ)(Jz − Jx ) dθ 2
b σ dσ
dσ σ + = 0, (14)
E(σ )  g , dτ, (10) dτ 2 Jy dτ
a dτ dτ
d2 θ 2 cos(φ) sin(φ)(Jz − Jx ) dθ dφ
− = 0. (15)
dτ 2 Iz + Jz (cos(φ))2 + Jx (sin(φ))2 dτ dτ
while the minimal geodesics are the minimizers of (10). In
coordinate form the geodesic equations on the DGT are
given by The solution σ (τ ) = [θ(τ ) φ(τ )]T to (14) and (15), which is
a geodesic, is a free motion, that is, a time evolution of the
2
d2σ i i
σ j dσ
dσ σk DGM without external forces.
+ jk = 0, (11) Clairaut’s relation given in [15] can be used to obtain
j,k =1
dτ 2 dτ dτ
qualitative solutions to geodesic equations for the DGT.
The parallel and meridian double-gimbal geodesics obtained
where i = 1, 2 and σ (τ ) = [σ σ 1 (τ ) σ 2 (τ )] = [θ(τ ), φ(τ )] . from Clairaut’s relations are the curves for which, respec-
Furthermore, for i, j, k = 1, 2, the Christoffel symbol jki is tively, the inner gimbal angle φ and outer gimbal angle θ
given by are held constant. Also obtained from Clairaut’s relation
are geodesics that are bound between two parallels as well
2 as geodesics that asymptotically approach a parallel. The
1
jk 
i
min (mnk, j + m jn,k − m jk,n ) , (12) geodesics on the DGT are summarized in Table 2 with plots
2 n =1
given in Figure 10. For instance, the geodesic curve that
asymptotically approaches φ = π/2 corresponds to the free
where min is the (i, n) entry of the inverse of the mass matrix motion of Figure 11. For this geodesic, the DGM asymptoti-
M(q) = [mij(q)] from (7), and (·),k denotes partial differentia- cally points straight up with zero control input.
tion with respect to the kth coordinate. Using kinetic energy as When the inertias of the DGM satisfy

the Lagrangian L(q, q̇) = (1/2) 2j,k =1 m jk (q)q̇ j q̇k , where
q = [q1 (τ ) q2 (τ )] and q̇ = [q̇1 (τ ) q̇2 (τ )] , the geodesic Jz = Jx , (16)
equations (11) can also be derived from the Euler-Lagrange
equations not only does the mass matrix (7) reduce to

∂L d ∂L  
− = 0, (13) Iz + Jz 0
∂qi dτ ∂ q̇i M= , (17)
0 Jy

where i = 1, 2.
To illustrate (12), we compute the Christoffel symbol but so do the geodesic equations (14) and (15) for the DGM
2 . Using the mass matrix entries of (7), and noting that simplify considerably to
11
(·),1 and (·),2 denote partial differentiation with
respect to θ and φ , respectively, and using
m22 = 1/Jy and m12 = m21 = m21 = 0, definition TABLE 1 Geometric data for the double-gimbal torus (DGT). Each
(12) with i = 2, j = 1, and k = 1 yields entry mi j of the mass matrix is substituted into (12) to find the
Christoffel symbols ij k . The functions ij k are first substituted into
(31) to find the curvature functions Rinj k· , which are then

multiplied
by the functions mnl to obtain the functions Ri j kl =mnl Rinj k· . A
2 1 2 double-gimbal mechanism (DGM) with inertias satisfying Jz = Jx
11 = m2n (mn1,1 + m1n,1 − m11,n )
2 n =1 has zero curvature and thus is called the flat DGM. The geodesics of
the flat DGM indicate that the inner and outer gimbal free motions
1 21 decouple, which allows each gimbal to rotate independently of the
= m (m11,1 + m11,1 − m11,1 ) other. If Jz = Jx , then the DGM system has curvature, which implies
2
1 22 that coupled free motions of the DGM can occur, as shown in Figure
+ m (m21,1 + m12,1 − m11,2 ) 10(b) and (c).
2

cos(φ) sin(φ)(Jz − Jx ) Nonzero Entries


= .
Jy mi j m11 = I z + Jz cos2 (φ) + Jx sin2 (φ), m 22 = Jy

cos(φ) sin(φ)(Jz − Jx ) cos(φ) sin(φ)(Jz − Jx )


i
jk
1
12 =− , 2
11 =
Table 1 provides the remaining nonzero Iz + Jz cos2 (φ) + Jx sin2 (φ) Jy
Christoffel symbols for the DGT.
((Jz − Jx ) cos4 (φ) + (2 Iz + 2 Jx ) cos2 (φ) − Iz − Jx )(Jz − Jx )
Using data from Table 1, the geodesic equations Ri j kl R1212 =
Iz + Jz cos2 (φ) + Jx sin2 (φ)
(11) for the DGT are

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 51


TABLE 2 Geodesics on the double-gimbal torus (DGT). For a double-gimbal mechanism (DGM) with inertia data given in the
left column, some of the geodesic types (parallel, meridian, bound, or asymptotic) are indicated. Each geodesic on the DGT is
a minimal energy motion, that is, a time evolution of the DGM in the absence of external forces. The second line of the table
describes the geodesics on the flat DGT. The fact that all parallels and meridian curves are geodesics indicates that all inner
and outer gimbal motions are uncoupled from each other. That is, each gimbal can spin independently of the other. However,
for the nonflat DGT, with geodesics described by the third and fourth lines of the table, the presence of bound and asymptotic
geodesics indicates a nontrivial coupling of the inner and outer gimbal motions shown in Figure 10(b) and (c).

Inertia Data Geodesic Type


Parallel Meridian Bound Asymptotic
φ = k1 , θ ∈ [0, 2π] θ = k2 , φ ∈ [0, 2π] φ1 ≤ φ ≤ φ 2 approach φ = k3
I z, Jz, Jx , Jy > 0 all all none none
Jz = Jx
I z, Jz, Jx , Jy > 0 φ = 0, φ = π/2 all present φ = π/2, φ = 3π/2
Jz > Jx φ = π, φ = 3π/2
I z, Jz, Jx , Jy > 0 φ = 0, φ = π/2 all present φ = 0, φ = π
Jz < Jx φ = π, φ = 3π/2

d2 φ inner and outer gimbal dynamics are decoupled, which


= 0, (18) implies that the inner and outer gimbal can rotate indepen-
dτ 2
d θ
2 dently of the other. For example, given a flat DGM with a
= 0. (19) nonzero initial velocity for the inner gimbal and zero ini-
dτ 2
tial velocity for the outer gimbal, the motion of the inner
gimbal does not induce motion of the outer gimbal.
In this case the DGM is called the flat DGM. An exam- For inertia-based PD control design of the DGM we
ple of a flat DGM is obtained by replacing the mirror in require the notion of a transport mapping [16] Tr→q , which
Figure 1 with a uniform sphere. The solutions to the equa- is a smooth mapping that carries a vector vr in the tan-
tions of motion (18) and (19) for the flat DGM are straight gent space to the torus at the configuration r to the vector
lines in the (θ, φ) coordinate plane. These straight lines vq = Tr→q(vr) in the tangent space to the torus at the con-
mapped onto the torus using T from (4) are the geodesics figuration q.
on the flat DGT summarized by the first row of Table 2. In The metric geometry of the DGT provides a specific,
fact, all solutions to the geodesic equations for the flat and natural choice of a transport map, namely, the parallel
nonflat DGT are mapped by T given in (4) to curves on the transport map Pr→q along a curve between configura-
torus. Consequently, it is not the “shape” of the torus that tions r and q. In coordinates, the parallel transport equa-
determines whether or not the DGT is flat or nonflat but tions, which characterize Pr→q , are the linear system of
rather the types of geodesics that are allowed on the torus; ordinary differential equations evaluated along a coordi-
see Table 2. The flat gimbal geodesics indicate that the nate curve σ (τ ) given by

(a) (b) (c)

FIGURE 10 Geodesics on the double-gimbal torus (DGT). The red curves are the gimbal geodesics σ (τ ) = [θ(τ ) φ(τ )] mapped under T
given in (4) and plotted on the surface T(θ, φ) with θ, φ ∈ [0, 2π). The parallel, asymptotic, and bound geodesics are typical of inertias
that fit the description of the second line of Table 2. For the torus and geodesics (a), (b), (c) above, the inertias are
I z = 2, Jz = 2, Jx = 1, Jy = 4 with initial conditions [θ = 0, φ = π/2, θ̇ = .8164, φ̇ = 0], [θ = 0, φ = 0, θ̇ = .6123, φ̇ = .3535] , and
[θ = 0, φ = 1, θ̇ = .7794, φ̇ = 0] for the parallel, asymptotic, and bound geodesics, respectively. Every geodesic on the DGT is a free
motion, that is, a time evolution of the double-gimbal mechanism (DGM) without external forces.

52 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


FIGURE 11 Asymptotic motion. The asymptotic gimbal geodesic in Figure 10(b) translates into a motion of the double-gimbal mechanism
(DGM) for which the inner gimbal asymptotically points straight up. Since gimbal geodesics account for the mass distribution of the system,
the corresponding motions are the natural tendencies of the DGM. By using the minimal energy gimbal geodesics, energy-efficient control is
possible by working with, rather than against, the gimbals’ natural tendencies.

dvi 2 σj
dσ Using (21), the coordinate vector vr is mapped to the
i
= − jk vk , (20) velocity vector of the outer gimbal in the configuration
dτ j,k =1

r = T(r) = [R 0 r]T to obtain
⎡ ⎤
such that v = [v1 v2 ] is the vector in the (θ, φ) coordinate 0

plane for which vr = (a) = DT(r)vr = ⎣ R dτ

(a) ⎦ , (24)

0
σ (τ ))vσ (τ ) = vσ (τ ) ,
DT(σ (21)
which is the yellow vector tangent to the black curve
where (22), as shown at the bottom of the page, is the Jaco- σ (τ )  T(σ
σ (τ )) on the torus in Figure 12. Parallel transport-
bian of the mapping T from (4) evaluated along the coordi- ing the vector vr along the coordinate curve σ (τ ), and
nate curve σ (τ ) = [θ(τ ) φ(τ )]T . mapping the resultant vectors to the torus using DT(σ σ (τ )),
The parallel transport equations (20) indicate that the defines parallel transport of vr = dσ /dτ (a) along the curve
components v1 , v2 of a coordinate vector v change with σ (τ ), which yields the blue vectors along the black curve in
respect to the parameter τ according to a term involving Figure 12. The fact that the parallel-transported vector
i
jk . Since the Christoffel symbols involve not only the mass changes along the curve σ (τ ) indicates that the vector
matrix entries, which are, as in the gimbal case (7), configu- varies with the tangent, or velocity, direction. Specifically,
ration dependent, but also the derivatives of the mass- for the gimbal in the configuration represented by
matrix entries with respect to the configuration variables, it q = σ (b) = [π (π/2)]T , the coordinate velocity vector vq is
is possible that jki = 0 for some i, j, k. As a consequence, a given by
parallel-transported vector can vary with the parameter τ .  
In contrast, Euclidean transport Ir→q of vectors along σ
dσ dθ
dτ (b) .
vq = (b) = (25)
curves in R3 leaves the vector constant from point r to dτ 0
every other point q. That is, Euclidean transport moves vec-
tors from point to point without rotating them. Using (21), the velocity vector of the gimbal in the con-
We now compare the parallel transport map Pr→q figuration q = T(q) = [−R 0 r]T is given by
with the Euclidean transport map Ir→q . For a purely
outer gimbal motion of the DGM represented by the coor- ⎡ ⎤
0
dinate curve σ (τ ) = [θ(τ ) (π/2)] , where τ ∈ [a, b], the dσ
vq = (b) = DT(q)vq = ⎣ −R dτ (b) ⎦ ,

(26)
coordinate velocity vector of the gimbal at the point dτ
0
r = σ (a) = [0 (π/2)] is given by
which is the green vector in Figure 12. In contrast, trans-
 dθ 
dτ (a) .
dσ porting the velocity vector vr using Euclidean transport
vr = (a) = (23)
dτ 0 Ir→q yields the black vectors in Figure 12. The fact that

⎡ ⎤
−R sin(θ(τ )) − r cos(φ(τ )) sin(θ(τ )) −r sin(φ(τ )) cos(θ(τ ))
DT(σ (τ )) = ⎣ R cos(θ(τ )) + r cos(φ(τ )) cos(θ(τ )) −r sin(φ(τ )) sin(θ(τ )) ⎦ , (22)
0 r cos(φ(τ ))

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 53



the black vectors are constant along the curve σ (τ ) indi- 0, n = i,
δki(vn )  (29)
cates that the Euclidean transported vector is unrelated to vk , n = i.
the outer gimbal motion. That is, while parallel transport
incorporates aspects of the dynamics of the gimbal, The relationship between the geodesic equations
Euclidean transport does not. Therefore, using Euclidean (11) and the parallel transport equations (27) can be
transport as a specific transport mapping in this case is seen as follows. When vi = dσ σ i/dτ , that is, when the
unphysical. vector v is chosen to be the velocity vector dσ σ /dτ along
We can rewrite the parallel transport equations (20) as the curve σ (τ ), the parallel transport equations (27)
become
D
vi = 0, (27)
dτ 2 σ j dσ
D σi
dσ d2σ i i
dσ σk
= + jk = 0, (30)
dτ dτ j,k =1
dτ 2 dτ dτ
where D/dτ is the covariant derivative along the curve σ (τ )
defined by
which are the geodesic equations of (11). The geodesic
D 2
d σj
dσ equations expressed in (30) are the analogue of Newton’s
 + i
jk δki, (28)
dτ i, j,k =1
dτ dτ law F = ma for F = 0 on a Riemannian manifold.
To explain the name flat DGM when (16) is satisfied, we
where the action of the operator δ on v is defined compo- consider the Riemann curvature tensor [14] given by
nentwise by

2  
ρ ρ η ρ ρ η ρ
Rγ αβ·  γβ,α + γβ ηα − γ α,β − γ α ηβ . (31)
η =1

For a pictorial explanation of Riemann curvature on the


torus, see Figure 13. For a DGM with inertias satisfying
(16), the DGT data from Table 1 indicate that not only does
the Riemann curvature of the DGT vanish but so does the
covariant derivative D/dτ equal the usual derivative d/dτ .
It follows that the geodesic equations for the flat DGM are
given by (18) and (19). The absence of curvature in the case
of the flat DGM decouples the inner and outer gimbal
motions, which means that each gimbal is free to rotate
independently of the other. The nonzero curvature of the
DGT indicates nontrivial coupling between the inner and
outer gimbal motions described by the various types of
geodesics listed in Table 2.
FIGURE 12 Parallel transport Pr →q versus the usual Euclidean
transport Ir →q . For a purely outer gimbal motion of the double- PHYSICALLY INSPIRED
gimbal mechanism (DGM) represented by the black curve

σ (τ ) = T (σ σ (τ )) with σ (τ ) = [θ(τ ) (π/2)] , where τ ∈ [a, b] , the
CONTROL DESIGN FOR THE DGM
velocity vector of the outer gimbal in configuration σ (a) = [R 0 r ]T In this section we focus on pointing and tracking control
is (dσ /dτ )(a) = [0 R (dθ/dτ )(a) 0]T , which is the yellow vec- design for the DGM. The pointing objective for the DGM is
tor. Parallel transporting the velocity vector (dσ /dτ )(a) to align the line-of-sight direction vector (6)
along the curve yields the blue vectors. The fact that
the parallel-transported vector rotates along the ⎡ ⎤
curve indicates that the vector varies with the tan- cos(φ) cos(θ)
gent direction. Specifically, the velocity vector of the LOS = ⎣ cos(φ) sin(θ) ⎦ , (32)
o u t e r g i m b a l a t t h e point σ (b) = [−R 0 r ]T is the green vec- sin(φ)
tor (dσ /dτ )(b) = [0 − R (dθ/dτ )(b) 0]T . In contrast, transport-
ing the velocity vector (dσ /dτ )(a) using Euclidean transport Ir →q
yields the black vectors. The fact that the black vectors are con-
with a reference pointing direction
stant along the curve indicates that the transported vector is
unrelated to the outer gimbal motion. That is, while parallel ⎡ ⎤
transport incorporates aspects of the dynamics of the gimbal, cos(φr) cos(θr)
Euclidean transport does not. Therefore, using Euclidean trans- ⎣ cos(φr) sin(θr) ⎦ . (33)
port as a transport map in this case is unphysical. sin(φr)

54 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


The double-gimbal mechanism is a multibody mechanical device composed
of three rigid bodies interconnected by two revolute joints.

Since each LOS vector can be associated with a point on t = LOS × g


the unit sphere S2 defined by = [−Kp(cos(ψ) sin(α) sin(φ) + cos(φ) sin(θ) sin(ψ))] î
S2  {(x, y, z) = (cos(φ) cos(θ), cos(φ) sin(θ), sin(φ)) : + [Kp(cos(α) cos(ψ) sin(φ) − cos(θ) cos(φ) sin(ψ))] ĵ
θ, φ ∈ (0, 2π]}, (34) + [Kp cos(φ) cos(ψ) sin(α − θ)] k̂, (36)

the configuration space of the pointing directions for the


DGM is S2 . Pointing control design for the DGM can be can be interpreted as the torque applied to the LOS vector rel-
viewed as setpoint stabilization on S2 . To achieve this con- ative to the inertial frame {î = [1 0 0]T , ĵ = [0 1 0]T ,
trol objective using physically inspired control forces, we k̂ = [0 0 1]T }. To determine how the gravity-motivated torque
design a setpoint controller for the DGM based on the (36) translates to the DGM, we project the torque vector t onto
analogue of gravity acting on a spherical mass-bob pendu- the inner and outer gimbal axes of rotation to obtain
lum. Essentially, we view LOS control design for the DGM
as the development of a potential-based control design on F = [Kp cos(ψ) cos(φ) sin(α − θ)
T2 , which induces setpoint stabilization on S2 .
− Kp(cos(α − θ) cos(ψ) sin(φ) − cos(φ) sin(ψ))].
The control objective for tracking control design of
the DGM is to asymptotically stabilize a configuration (37)
curve q(t) on T2 to a reference configuration curve r(t)
on T2 . To achieve this control objective, we again use A potential generating function V for the force field F
gravity as a motivation to derive a potential function on on T2 given by (37) is a function whose negative differen-
S1 and then extrapolate the design to the DGM configu- tial −dV equals F. Such a potential is given by
ration space T2 = S1 × S1 . In a second tracking control
design for the DGM, we proceed along the lines of [16] V(θ, φ) = −LOS · g
and [17] by specifying energy-based, coordinate-inde- = Kp(sin(φ) sin(ψ) − cos(α − θ) cos(φ) cos(ψ)).
pendent generalized PD control forces on T2 . All of the
(38)
control designs in this section are free from the limita-
tions of generalized coordinates. For a
detailed example of the limitations imposed
by a chosen set generalized coordinates, see
“Circular Corridor Revisited.”

Gravity-Based Pointing-System
Design for the DGM
To begin pointing design for the DGM, let
g ∈ S2 be represented by the vector

⎡ ⎤ (a) (b)
cos(ψ) cos(α)
g = Kp ⎣ cos(ψ) sin(α) ⎦ , (35) FIGURE 13 Parallel transport illustration of curvature on the torus. From a pictorial
sin(ψ) standpoint, the torus has curvature if there exists a closed curve such that the paral-
lel transport of an initial vector vi around this curve returns a final vector v f = vi .
Consequently, the difference between an initial vector vi and the parallel-transported
where ψ ∈ [−π/2, π/2], α ∈ [0, 2π], and Kp vector v f is essentially a measure of curvature. If vi = v f then the torus is flat, while if
is a scalar. The vector g can be interpreted vi = v f then the torus is not flat. (a) The parallel transport of the blue vector vi
as a gravity-like, uniform force field of around the black closed curve yields the green vector v f , which is not equal vi , and
thus the torus has nonzero curvature. (b) Parallel transporting the blue vector vi
strength Kp. By imagining that a unit point
around the black closed curve yields the green vector v f . The fact that the v f = vi
mass is located at the end of the LOS direc- seems to indicate that the torus is flat. However, the flatness conclusion is erroneous
tion vector, the vector t given by since the black curve circumnavigates a hole in the torus.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 55


 
The fact that a DGM pointing design based on the ∂F ∂F
dF(p) = (p) (p) = [0 0]. (39)
potential (38) cannot be globally stabilizing follows from ∂c1 ∂c2
Morse theory [14], which relates the Betti numbers of T2
to bounds on the number of equilibrium points on T2 The Hessian Hess f (p) of a real-valued function f eval-
when the controller is derived from a smooth potential. A uated at p can be expressed in coordinates by
critical point p of a smooth, real-valued potential func-
tion f on, for example, a two-dimensional submanifold ⎡ ∂2 F ∂2 F ⎤
of R3 is a point p that satisfies df (p) = 0. In terms of coor- ∂c21
(p) ∂c1 ∂c2 (p)
dinates (c1 , c2 ), the function f can be expressed as Hess F(p) = ⎣ ⎦. (40)
∂2 F ∂2 F
F(c1 , c2 ), and the condition df (p) = 0 can be written as ∂c2 ∂c1 (p) ∂c22
(p)

Circular Corridor Revisited

T
o select a suitable set of generalized coordinates for the track- flat
u = m11 ū (S6)
er and target confined to R2 , we center the corridor at the ori- 1
hilly
gin of the coordinate plane and employ polar coordinates (r, θ). u = m11 (θ)ū + (5 cos (5θ) + 2 cos (2θ))θ̇ 2 , (S7)
2
Fixing the radial coordinate imposes a circular kinematic con-
straint, where the angle of rotation θ acts as the single coordinate where ū is a new control parameter that can be selected to achieve
needed to describe the configuration of the system. That is, after a desired effect. For the choices (S6) and (S7), the control systems
imposing kinematic constraints, a single degree of freedom (S4) and (S5), respectively, reduce to the dynamic equations for a
remains for this system with positive values of θ indicating coun- unit-mass particle given by θ̈ = ū. Letting (θd , θ̇d ) denote the
terclockwise rotation. phase variables for a reference configuration, we can choose
To formulate dynamics, we express the tracker’s kinetic
energy in terms of its phase variables (θ, θ̇). Letting m11 (θ) ū = θ̈d − Kp e − Kd ė , (S8)
denote the configuration-dependent mass of the particle, the

kinetic energy of the particle is given by where e = θ − θd and where K p and K d are the proportional-
derivative control design parameters. The selection of (S8) for
1 1
K = g(θ̇, θ̇) = m11 (θ)θ̇ 2 . (S1) the new control parameter renders the closed-loop error system
2 2
in the form of a mass-spring-damper error equation given by
Assuming no potential forces, the Lagrangian is L = K . The
dynamics are obtained by equating to zero the resultant of all ë + Kd ė + Kp e = 0 (S9)
forces imparted on the tracker including the inertial forces. The
inertial force I is given by with damping constant K d and spring constant K p . Using (S9) we
can choose the damping and stiffness for both global asymptotic
d ∂L ∂L
I =− + . (S2) stability and response. From the traditional control perspective, we
dt ∂ θ̇ ∂θ
have thus obtained a complete solution to both the flat and hilly
Letting u denote the generalized force along the θ direction, the circular corridor chasing problems.
full dynamics are given by I + u = 0, that is, We now look deeper into the coordinate control formulation
above to address several points of concern stemming from the
d ∂L ∂L
− = u. (S3) subtle use of coordinates. For instance, where are the indecision
dt ∂ θ̇ ∂θ
points, those points of discontinuous control logic, introduced in the
For the flat corridor these dynamics are circular corridor examples? Each term in (S9) is smooth in the tra-
ditional approach and therefore it seems that discontinuous control
m11 θ̈ = u , (S4)
logic is unnecessary. The traditional perspective seems to accom-
while, for the hilly corridor with metric given by, for illustration pur- plish what we intuitively understand to be impossible, namely, the
poses, m11 (θ) = 2 + sin(5θ) + sin(2θ), the dynamics are construction of a smooth controller that renders the desired trajec-
tory a global asymptotically stable equilibrium. To find the flaw in
1
m11 (θ)θ̈ + (5 cos (5θ) + 2 cos (2θ))θ̇ 2 = u. (S5) the traditional control approach we retrace the above steps.
2
Recall that the first step of Lagrange’s method is to intro-
From the traditional Lagrangian perspective, these mathematical duce a generalized coordinate, namely, the real-valued polar
expressions model the system’s dynamics. coordinate θ to denote position. However, the configuration
Proceeding with the computed torque control design, we space for this system is S1 and not the real line R. The corre-
select a control acceleration u that mimics the inertial force for spondence between the two spaces is given by
both the flat and hilly corridors. That is, f : R → S1 : θ → [cos θ sin θ], which is not a homeomorphism

56 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


The critical point p of f is nondegenerate if Hess F(p) is only an equilibrium with Morse index zero can be stable.
nonsingular or, equivalently, has nonzero eigenvalues. All other indices correspond to equilibria with saddle or
Consequently a nondegenerate critical point p is isolated, unstable equilibrium behavior.
which means that p has a neighborhood that contains no Setting the differential dV of the potential (38) to
other critical points. A function f is called a Morse func- zero yields, after some simplifications, the equilibrium
tion if all of its critical points are nondegenerate. The conditions
Morse index of the critical point, which is the number of
negative eigenvalues of the Hessian evaluated at a critical
point p of f , determines the stability of the critical point cos(ψ) cos(φ) sin(α − θ) = 0 , (41)
when f is employed as a potential function. For instance, cos(ψ) sin(φ) cos(α − θ) = cos(φ) sin(ψ) . (42)

since holes are topological invariants. In other words, the circle


has a hole, while R does not. Further, note that since f is peri-
odic, it is not one to one over R even though it is continuous.
Restricting the domain of f defines a new map
h : [−π, π) → S1 : θ → [cos θ sin θ], which is one to one and
onto as desired but, unfortunately, lacks continuity at −π . That
is, the circle has to be torn to create the segment, which means
that S1 and R are not topologically the same, and thus R is not FIGURE S1 Tracking in a rectilinear corridor. In a traditional view
sufficient for capturing the kinematic constraints of the system. of the tracking problem in the circular corridor examples, the
The traditional view of kinematics has the virtue of expediting problem is indirectly turned into one of tracking in an infinite rec-
tilinear corridor. Since the target must lie either to the left or the
analysis but, without care, can lead to problems. Due to the
right of the intercepter (but not both) there is no directional ambi-
nature of the topology of R, we can devise a smooth controller guity for the intercepter. Furthermore, the perceived distance to
that yields a global asymptotic equilibrium on the real line a fixed target position is proportional to the effort it takes to inter-
because the controller is never confused about the direction to cept the position when moving at a fixed speed. Clearly, there is
the target—it lies either to the right or the left as shown in Figure a natural control logic to be used in this scenario, and none of
the control issues that arise in the case of the circular corridor
S1. However, the controller in this case is stabilizing the refer-
are present.
ence parameter curve θr : R → R rather than the configuration
reference curve r : R → S1 . To illustrate, suppose the parame- Another point of concern is the singularity of the generalized
ter and configuration curves are, respectively, the constant coordinate approach. Notice that although (π/2), and for that mat-
curves t → (π/2) and t → [0 1]. Then the controller stabilizes ter, (π/2) + 2πm for any integer m, contain the point [0 1], these
(π/2), but fails to stabilize [0 1]. Notice, if θ = (5π/2), then the parameter values correspond to different proportional feedback
controller pushes off this value and toward (π/2) and yet, forces, namely, −2K p πm. Consequently, the controller, pertaining
f(5π/2) = [0 1]. That is, the tracker can be sitting on top of the to actual configurations, is a multivalued function. From the per-
target and yet pass it by, not even remaining in its vicinity. The spective of the configuration space, this feedback force field is not
configuration is not stable, for, were the configuration [0 1] to be well defined. This phenomenon is called wrapping with closed-
stabilized by the controller, then all initial configurations near loop instability of configuration and energy inefficiency the result-
[0 1] would stay near [0 1].These topological considerations indi- ing artifacts. The only way to make this controller well defined and
cate that the equilibria are not globally asymptotically stable. continuous is to restrict its use to an actual coordinate chart by
When we add energy considerations to the configuration restricting f and the coordinate controller to the interval
space of a mechanism there are subtle, and yet real situations, (−(π/2), (3π/2)). This restriction indicates that the control logic
in which the traditional coordinate controller is ill-behaved. For does not cover all situations since f is locally a homeomorphism
instance, reconsider the flat corridor. If the target is again con- between the punctured circle S1 − {[0 − 1]} and this subspace of
stantly positioned at (π/2) and we are approximately at R. This restriction erases not only topology problems but also, in
−(π/2) + ε where ε is a small positive number, headed toward the case of the flat corridor, the energy problems.
−(π/2), with momentum to spare, then, unless the control These examples illustrate that, in most instances, configura-
gains are set for very high bandwidth response, we pass by tion manifolds of a mechanism are often both not flat and not
−(π/2). Logically, we realize that we are confined to a circular without holes. Ignoring the unique features of a mechanism’s
corridor, which means we can reach the goal by continuing on configuration space results in decision making that runs contrary
without wasting any effort to turn around. However, the coordi- to intuition and therefore may fail to attain the objectives estab-
nate controller, contrary to logic and with wasted effort, requires lished as the basis for decision making. We therefore take topol-
that we must turn around. ogy and geometry into account when designing control logics.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 57


The solutions to (41) and (42) are the critical points based control designs on T2 than the above design that has
given by the following cases: eight equilibria. A four-equilibria potential on T2 is given
in the next section.
Case 1) ψ = ± π2 , φ = ± π2 , θ = free ,
Proposition 1
Case 2a) [θ, φ] = [α, ψ], [θ, φ] = [α + π, −ψ + π],
Let V be a Morse function on the torus, the configuration
Case 2b) [θ, φ] = [α, ψ + π], [θ, φ] = [α + π, −ψ],
manifold of the DGM. Then the number nk of critical
Case 2c) [θ, φ] = [α + π2 , π2 ], [θ, φ] = [α + π2 , − π2 ],
points of Morse index k must satisfy
Case 2d) [θ, φ] = [α − π2 , π2 ], [θ, φ] = [α − π2 , − π2 ].

n0 ≥ β0 (T2 ), n1 ≥ β1 (T2 ), n2 ≥ β2 (T2 ) ,


In Case 1, the set of equilibria is infinite, and, since
these equilibria are not isolated, the potential function (38) where β0 (T2 ) = 1, β1 (T2 ) = 2, and β2 (T2 ) = 1 are the Betti
is not a Morse function. However, Case 2 yields a Morse numbers of the torus.
function having eight equilibria, namely, two minimizers
(Case 2a, stable points), four inflections (cases 2c and 2d, PD Control Design on Riemannian Manifolds
unstable points), and two maximizers (Case 2b, unstable In the rest of this article we focus on fully actuated mecha-
points). It follows from (32) that the two stable and two nisms as addressed in [16], [18], and [19], wherein the con-
unstable equilibria on T2 from cases 2a and 2b induce a stitutive elements of the augmented PD control law [20]
single stable and single unstable equilibrium on S2 , respec- are abstracted to a mechanism’s Riemannian manifold. We
tively. The stable equilibrium on S2 becomes asymptotical- formally state, interpret, and then apply a geometrically
ly stable with the introduction of the dissipative force inspired stabilization theorem to the DGM.
Fd = −K d[θ̇ φ̇], where K d is a positive constant. That is, for The following notation is employed in Theorem 1. The
most initial orientations of the DGM with zero initial operator D/dt is the covariant derivative operator along
velocity, the force F − Fd aligns the LOS with the uniform the curve q(t) defined analogously to (28). For a configura-
gravity field. tion error function ϕ(·, ·), the notation ϕr(q) and ϕq(r) indi-
The additional equilibria in cases 2c and 2d correspond cates that r and q are fixed but arbitrary configurations
to the physical phenomenon known as gimbal lock. That implying that ϕr(·) and ϕq(·) are functions of only q and r,
is, when the outer gimbal is oriented orthogonally to the respectively. The notation dφr(q) is thus the differential of
uniform force field, as indicated in Case 2c by θ = α + π/2, the function φr(·) with respect to q. The symbol g (q)(·)
and when we concurrently direct the inner gimbal to be denotes the operator g (q)(vq)  g(vq, ·), where g(·, ·) is the
orthogonal to both the outer gimbal orientation and that of metric defined in coordinates by (5).
the uniform field, the DGM reacts with a constraint reac-
tion force commensurate with the torque t of (36). As a Theorem 1
result, the DGM is locked. Consider the control system
Construction of the potential (38) requires two distinct
topological spaces, the sphere S2 and the torus T2 . The D
g (q) q̇ = F, (43)
topological differences are indicated by the fact that they dt
have different Betti numbers. In particular, the Betti num-
bers of S2 are β0 (S2 ) = 1 , β1 (S2 ) = 0 , and β2 (S2 ) = 1 , and let r(t), t ∈ [0, ∞) be a twice differentiable reference
whereas the Betti numbers of T2 are β0 (T2 ) = 1, β1 (T2 ) = 2, trajectory. Furthermore, let ϕ be a configuration error
and β2 (T2 ) = 1. The above gravity-based pointing design function, let Tr→q be a transport map compatible with ϕ,
started with the definitions of both the LOS (32) and gravi- let K d(q) be a dissipation function represented by a posi-
ty (35) as points on S2 . Since the DGM physically imple- tive-definite matrix, and define the control input
ments the setpoint design, and since the configuration F = FFB + FFF , where
space of the DGM is the torus, we are required to project
the torque (36) onto the torus. Since there does not exist a
homeomorphism between the sphere and torus topologies, FFB = −d ϕr(q) − K d(q)ė , (44)
  
it is not surprising that the pointing design based on (35)
 D d 
yields gimbal-lock solutions. FFF = g (q) Tr→q ṙ +  (Tr→q ṙ) . (45)
dt dt q fixed
The following proposition relates Morse theory on T2
with the Betti numbers of T2 . This proposition is a special
case of a more general theorem given in [5], which indi- Then the curve q(t) = r(t) is stable with Lyapunov function
cates that the minimal number of equilibria of any Morse
function on T2 is four. Therefore, there are better, at least 1
Wtotal (q, q̇, r, ṙ) = ϕ(r, q) + g(ė, ė) . (46)
from the standpoint of number of equilibria, potential- 2

58 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


In addition, if the configuration error function ϕ is uniformly Tracking Control Design for the DGM
quadratic [16, p. 536] with constant L and if additional We now apply Theorem 1, the generalized augmented PD
boundedness assumptions [16, pp. 540–541] are met, then the control design, to the DGM. Using gravity as motivation,
curve q(t) = r(t) is exponentially stable with Lyapunov func- we design an energy-based, configuration error function
tion Wtotal from all initial conditions (q(0), q̇(0)) such that for the circular corridor problem and then extrapolate the
design to the DGM.
Wtotal (0) < L . (47) Consider the circle S1 embedded in a uniform gravity
field of strength Kp as shown in Figure 14(a). The uniform
For further detail relating to Theorem 1, see [16, chap. field induces a gravity field on the circle by projection as
11]. This result can be interpreted as follows. The error shown in Figure 14(b). The projected field indicates how a
function ϕ is similar to a Hookean stiffness potential since, mass particle confined to the circle, such as the mass bob of
in the vicinity of r, the properties of ϕ ensure that it is a a pendulum, experiences the influence of the uniform
potential well, which attracts the mechanism configuration gravity field. As shown in Figure 14, the force field on the
q to the reference trajectory r. Requiring that the configura- circle induces two isolated equilibria for the particle,
tion error function ϕ be uniformly quadratic in a ball of
fixed radius about each point r further supports the poten-
tial well interpretation.
As illustrated in the section “Metric Geometry” using
the parallel transport map, a transport map Tr→q can be
used to compare velocities in different tangent spaces to
a Riemannian manifold. Given curves r(t) and q(t) in a
Riemannian manifold, a velocity error is given by

ė = q̇ − Tr→q ṙ . (48)

We emphasize that not any transport mapping suffices


since Tr→q must interact with the chosen configuration error
function ϕ in a compatible manner, as defined by the relation


d ϕq(r) = −Tr→q (d ϕr(q)) . (49)
(a)
In (49) the transport map Tr→q has been “pulled back,”
denoted by Tr→q∗ , so that the transport map can operate on

the differential d ϕr(q).


The positive-definite matrix K d(q) in Theorem 1 acts as
a viscous damping coefficient to dissipate the relative
motion between r(t) and q(t).
Given the explanations of ϕ, Tr→q , and K d(q), Theorem 1
is a geometric analogue of an augmented PD control
design, which in the setting of Rn and for fully actuated
simple mechanical control systems [16], can be expressed as

F = q̈T M(q) + q̇T C(q, q̇) , (50)

where F = FFB + FFF is given by (b)

FFB = −eT Kp − ėT K d(q) , (51) FIGURE 14 Physically inspired control of the double-gimbal mecha-
nism (DGM). For the uniform gravitational field on R2 of (a), the
FFF = ṙT C(q, q̇) + r̈T M(q) , (52) induced nonlinear field on S1 in (b) has two equilibria, one stable
and one unstable. When a particle is placed on the hoop, the pro-
where e  q − r. The augmented PD control design jected force field indicates how the particle experiences the uniform
gravity field due to its confinement to the circle. This behavior can
(50)–(52) simplifies to the error system
be used to design a physically inspired tracking controller on the
circle by simply altering the direction of the ambient force field. This
approach can be extrapolated to T2 = S1 × S1 , the configuration
ëT M(q) + ėT (C(q, q̇) + K d(q)) + eT Kp = 0. (53) space of the DGM.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 59


The focus in this article is the idealized DGM, which we use as the primary
example to illustrate general topological and geometric concepts.

 
namely, a stable equilibria at the bottom point [0, −1] of r , and Hess r (r) = α 10 01 . Since Hess r (r) is positive def-
the unit circle and an unstable equilibria at the top point inite, the local potential  is an attractive potential well. Con-
[0, 1] of the unit circle. sequently, ϕ defined by (56) is a configuration error function.
The induced field of Figure 14(b) can be generated by a It can further be shown that this configuration error function
potential on S1 , the configuration space of the mass parti- is uniformly quadratic [16, p. 540] and therefore meets the
cle confined to the circle. The coordinate-free potential at a conditions of Theorem 1. Since dq = −dr and because
point q on the circle is Tr→q = Ir→q , that is, the Euclidean transport mapping
defined in the geometry section, the configuration error
v(q) = Kp(1 − Cos(q)), (54) function of (56) is compatible with the transport mapping.
We have now realized the augmented PD control
where Cos is the circular function cosine defined directly on design for the DGM. The coordinate expression of the
S1 . Letting θ denote the radian measure of the angle sub- feedback part of the controller is given by
tended by [0, −1], that is, the bottom point of the circle, and
the configuration q of the mass particle, then the parameter- FFB = − dr (q) − ėT K d
 
ized expression of the potential v(q) from (54) is given by = − α sin(θ − θr) sin(φ − φr)
 
V(θ) = Kp(1 − cos(θ)). (55) − θ̇ − θ̇r φ̇ − φ̇r K d , (59)

The potential V(θ) is related to the projected force field which, together with the feedforward compensation, ren-
in Figure 14(b) by −dV = −Kp sin(θ)dθ . ders the reference curve r(t) locally exponentially stable in
With the potential V(θ) in mind, we develop an aug- the closed loop. The key step in the design of (56) is to first
mented PD control law for the DGM. Let q(t) = [θ(t) φ(t)]T select a model potential with the desired properties at a
and r(t) = [θr(t) φr(t)]T denote, respectively, the coordinate specific configuration q defined by
representations of the configuration and reference with
respect to the standard (θ, φ) parameterization of the torus. V(q) = 2 − cos(θ) − cos(φ) , (60)
The candidate configuration error function ϕ is defined by
its coordinate expression  given by which, as shown in Figure 15, has an ideal well at
q = [0 0]T . The potential (60) can then be used to define a
(q, r) = α(2 − cos(θ − θr) − cos(φ − φr)), (56) configuration error function (56) by substituting q − r for
q to ensure that the potential varies with the reference
where α > 0 is a scalar that models stiffness. We use configuration r.
Euclidean transport in the parameter space to induce par- As with the pointing design based on the potential
allel transport on T2 and choose K d(q) to be a symmetric, (38), Morse theory indicates that the DGM tracking
positive-definite matrix with constant entries. design based on the potential (60) cannot be globally sta-
We now show that ϕ is a configuration error function bilizing. In fact, equating the differential of the potential
compatible, in the sense of (49), with the transport map- (60) to zero, we find four isolated critical values, which
ping. The coordinate function  is symmetric, proper, and correspond to the (θ, φ) parameter vectors [0 0]T , [0 π]T ,
bounded from below. When r is held fixed, the differential [π 0]T , and  [π π]T .  Evaluating the Hessian
)
dr and Hessian at q are given by HessV r (q) = α cos(θ
0
0
cos(φ)
at these values, respectively,
the critical points are of Morse index 0, 1, 1, and 2, from
 
dr (q) = α sin(θ − θr) sin(φ − φr) , (57) which we conclude the existence of one stable equilibri-
  um, two saddles, and one unstable equilibrium. The sta-
cos(θ − θr) 0 bility properties of these equilibria are visually confirmed
Hess r (q) = α . (58)
0 cos(φ − φr) in Figure 15. Proposition 1 indicates that not only are
these equilibria expected from the control design, but this
Setting q = r, we find that r (r) = 2 − 1 − 1 = 0 , design is among the best achievable potential-based con-
dr (r) = α[ 0 0 ], which indicates that r is a critical value of trol designs on the torus. By best, we mean that the

60 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


We present a geometric approach to the DGM that provides a solid
foundation for the global mechanical behavior of this device.

design based on (60) achieves the minimal number of using methods analogous to those given in this section,
four equilibria, whereas all other designs must have at the controlled dynamics for the nonflat DGM reduce to a
least four equilibria. For example, the potential (38) on T2 spring-damper system along the geodesics defined by the
has eight equilibria. metric geometry of the nonflat DGM [21].
As the model potential (60) moves with the reference r, In the case of the flat DGM, which is characterized by
Proposition 1 is still applicable, and therefore the minimal inner-gimbal inertias that satisfy Jz = Jx , the controlled
number of predicted equilibria is maintained. dynamics simplify considerably because not only does the
parallel transport map Pr→q reduce to the Euclidean trans-
Geometric PD Control on the Flat DGT port map Ir→q , but so do all covariant derivatives reduce
Although the generalization of the augmented PD con- to regular derivatives. The controlled dynamics (43) for the
troller is a physical control design, it does not initially flat gimbal DGM are given by
account for the mass distribution of the DGM. As a
result, the control functions only to provide reference D dq
M = q̈T M = F , (61)
stability with no natural preference given to how stabili- dt dt
ty is achieved. The goal of this section is to incorporate
the metric geometry structure of the DGM into the aug- where M is the generalized mass matrix from (17),
mented PD control strategy, which we then implement q(t) = [θ(t) φ(t)]T is the coordinate vector representing
on the flat DGM. the configuration q of the DGM at time t, and F is a vec-
The focus of this section is the flat DGM because the tor of input torques. Let r(t) = [θr(t) φr(t)]T be a smooth
geodesic equations on the flat DGT have simple closed- reference trajectory that the DGM configuration curve
form solutions, which, for appropriate choices of configu- q(t) is required to asymptotically approach. To com-
ration error function and compatible transport map, pute the proportional term dφr(q) of the feedback force
simplify the implementation of Theorem 1 to the aug- FFB from (44), we choose the geometric configuration
mented PD control design (53). Indeed, for the flat DGM, error function
(53) is based on the Euclidean
model of state error given by
e  q − r. In contrast, as shown
Phi 2 Phi 2
in the section “Metric Geome-
0 0
try,” the geodesic equations on
−2 −2
the nonflat DGT are consider-
ably more complex, and thus 3 3
numerical solutions to the geo- 2 2
desic equations must be
obtained to implement the con- 1 1
trol design of Theorem 1 on the 0 0
nonflat DGT. Furthermore, the
−1 −1
formulation of the controlled 0
−2
dynamics of Theorem 1 for the 0
2
nonflat DGM requires a non- Theta 4
2 Theta
Euclidean model of state error. 6
Consequently, the controlled (a) (b)
dynamics for the nonflat DGM
do not reduce to the augmented FIGURE 15 A double-gimbal mechanism (DGM) control potential. One approach to designing
PD control design (53). Howev- a configuration-error function for the generalized augmented PD controller is to first con-
struct a model potential V(q) and then to define V(q − r), which ensures that the potential
er, for appropriate choices of varies with the reference trajectory r. We illustrate in (a), with alternative view in (b), the
configuration error function and model potential for the DGM that results from extrapolating the physically inspired design on
compatible transport map, and S1 described in Figure 14 to the configuration space T2 = S1 × S1 of the DGM.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 61


Topology and geometry play a crucial role in the analysis of the global,
nonlinear properties of mechanical systems.

β error function (67), but so also does the velocity error func-
(r(t), q(t)) = dist2 (r(t), q(t)) , (62)
2 tion reduce to
ė = q̇ − ṙ. (70)
where β > 0 is a scalar and where dist is the intrinsic dis-
tance function defined on the DGT by (9). The configura- A geometric Rayleigh dissipation torque for the flat
tion error function of (62) can be simplified by solving the DGM is given by
boundary value problem
FD = −α(q̇ − Pr→q ṙ)T M
D dσ d 2σ = −α(q̇ − ṙ)T M , (71)
= = 0, (63)
dτ d τ dτ2
σ (0) = r(t), σ (1) = q(t) . (64) where α > 0 is a scalar. The control forces FP and FD sum
together to define the feedback force FFB defined in
The solution σ (τ, t) to (63) and (64) is, under appropri- Theorem 1. To complete the specification of the control
ate conditions discussed in the next section, the unique, forces from Theorem 1, we define the feedforward term
minimal geodesic curve between r(t) and q(t) given by
T
D
FFF = (Pr→q ṙ) M = r̈T M , (72)
σ (τ, t) = r(t)(1 − τ ) + q(t) τ. (65) dt

It follows from (65) that, for each configuration q(t) and which accelerates the gimbal trajectory q(t) toward the ref-
reference r(t), erence r(t). Applying the control torque F = FFF + FP + FD ,
the error system for the flat DGM is
σ

= q(t) − r(t), (66)

(ë + α ė + βe)T M = 0 , (73)
which, when substituted into (9), leads to the simplified
configuration error function where e = q − r.
For setpoint control to the constant configuration r, the
β control forces (68), (71), and (72) based on the intrinsic dis-
(r(t), q(t)) = (q(t) − r(t))T M (q(t) − r(t)) . (67)
2 tance function on the DGM are torques that force the gim-
bal along a geodesic between r and q in the coordinate
Holding r = r(t) fixed, the components of the negative dif- plane. The response of the gimbal can be set to under-
ferential of (67) with respect to q = q(t) determine damped, critically damped, or overdamped by tuning the
scalars α and β.
FP = dr (q) = −β(q − r)T M , (68) The geometric control design methodology given above
for the flat DGM can be adapted to the coupled, that is
which is the proportional term in (44) of the PD control nonflat, DGM. While there may not be a closed form solu-
design from Theorem 1. tion to the resulting geodesic boundary value problem or
For the velocity q̇ of the configuration curve q(t) to parallel transport equation, accurate solutions can be
match the velocity ṙ of the reference curve r(t), a Rayleigh achieved through numerical approximation. Consequent-
damping torque is used to dissipate energy between the ly, as in the flat DGM, the nonflat DGM can be compelled
gimbal configuration q(t) and reference configuration r(t). to move as a spring-damper along geodesics defined by
By parallel transporting the reference velocity vector ṙ to the metric geometry of the DGT [21].
q(t) for each t, the velocity vectors q̇ and Pr→q ṙ can be
subtracted to obtain a velocity error function Cut Locus for the DGT
The inertia-based PD control design from the preceding
ė = q̇ − Pr→q ṙ. (69) section depends on finding a unique minimal geodesic
between the desired configuration r and the mechanism
For the flat DGM, not only is the Euclidean transport configuration q. Figure 16 illustrates the existence of gim-
map compatible, in the sense of (49), with the tracking bal configurations connected by two minimal geodesics,

62 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008


for example, the red point and
any point on the antipodal ring.
As a result, the Hookean poten-
tial ϕ in (62) is not a configura-
tion error function on the whole
DGT. Consequently, the PD con-
trol law based on this potential is
not globally defined. We must
restrict the domain of definition
of the control law and search for
local stability results. The solu-
tion to this technical issue is dis- FIGURE 16 Two views of the estimated injectivity radius and Loki computation of the cut locus for
the double-gimbal torus (DGT). The blue geodesics emanate from a point p for a distance i (T2 )
cussed in the following section. given by the Klingenberg injectivity radius result, which defines the red curve. The green geodes-
For the Hookean potential of ics emanate from the point p and reach the two black rings called the cut locus C( p), which are
(62) the set of problematic config- determined by the software Loki. For inertia-based tracking control of the double-gimbal mecha-
urations q for each reference con- nism (DGM), the region described by the red curve is a conservative estimate of the region of
figuration r comprise the cut locus local exponential stability. For setpoint-to-setpoint maneuvers, the black rings are problematic ref-
erence configurations for inertia-based PD control logic.
C(r) of r. In the case of the circu-
lar corridor, the cut locus of r is
its antipode, which is the point of control indecision. In the determined by i(T2 ) for which an energy- and distance-
case of the hilly corridor, the cut locus is not necessarily based control law provides local exponential stabilization
the antipodal point because, in general, C(r) is mass for a given DGM reference r. The extent of the stability
dependent. That is, the existence of the cut locus is dictated region is determined by the magnitude of i(T2 ). A result
by the topology of S1 , while a description of the cut locus of Klingenberg [22] shows that i(Q) is the smaller of either
is determined by the metric geometry of S1 . half the length of the shortest closed geodesic or π divid-
Only by avoiding the cut locus of the DGT can the dif- ed by the least upper bound on the sectional curvature.
ferential of ϕ from (62) be computed as For the DGT the sectional curvature is the Ricci curvature
Ric obtained from the Riemann curvature R by
dσ 
−αd ϕr(q) = −αg  (∇ ϕr) = −αdist(r, q) g  , (74) Ricij  2k =1 Rkikj· where i, j = 1, 2. The Klingenberg result
dt
requires knowledge of all of the closed geodesics on the
where σ is the unit-speed geodesic segment connecting q DGT. Although a characterization of all of the closed geo-
to r and g  is the Riemannian metric viewed as function desics on the DGT is not available, we suspect that in
acting on a vector. The extent of the cut locus for each DGM applications the meridian geodesics shown in Table 2
point on the reference trajectory
r(t) limits the stability region for
the geodesic-based PD tracking
control design. One way to avoid
indecision points in a geodesic-
based control design is to find
the smallest distance to the cut
loci over all points of the configu-
ration space. This distance,
denoted by i(Q), is called the
Time = 0 s Time = 1.9 s Time = 11.9 s Time = 17.9 s Time = 27.9 s
injectivity radius of the Riemannian
manifold Q and determines the
domain on which ϕ of (62) is in
fact a configuration-error func-
tion. See [21] and [22] for details
on computing the injectivity
radius of a Riemannian manifold.
The compactness of the con-
figuration space T2 of the DGM FIGURE 17 Setpoint-to-setpoint tracking on the double-gimbal torus (DGT). The geodesic seg-
ments given in black are the minimal energy motions between two successive double gimbal con-
is sufficient to guarantee that figurations. Since the fixed reference configurations are not points on the cut locus of any point
i(T2 ) > 0. Consequently, there along the geodesics, the inertia-based PD control design is valid and proceeds in an energy-effi-
exists a region of phase space cient manner.

AUGUST 2008 « IEEE CONTROL SYSTEMS MAGAZINE 63


are the shortest closed geodesics. These geodesics corre- [2] M.K. Masten, “Inertially stabilized platforms for optical imaging systems
spond to the inner gimbal spinning through 2π rad at a tracking dynamic targets with mobile sensors,” IEEE Contr. Syst. Mag., vol.
28, no. 1, pp. 47–64, 2008.
fixed outer gimbal angle. Since the inner gimbal usually has [3] H.G. Wang and T.C. Williams, “Strategic inertial navigation systems
the lesser inertia of the two gimbal components, any other high-accuracy inertially stabilized platforms for hostile environments,” IEEE
closed geodesic, which involves motion of the outer gimbal, Contr. Syst. Mag., vol. 28, no. 1, pp. 65–85, 2008.
[4] H. Choset, K. Lynch, S. Hutchinson, G. Kantor, W. Burgard, L. Kavraki,
requires more energy. Taking half the length of a meridian and S. Thrun, Principles of Robot Motion: Theory, Algorithms, and Implementa-
geodesic corresponding to a purely inner gimbal motion, tions. Cambridge, MA: MIT Press, 2005.
and comparing to the bound of the sectional curvature, we [5] C. Nash and S. Sen, Topology and Geometry for Physicists. New York: Acad-
emic, 1983.
conjecture half the meridian length to be a good estimate of [6] T. Frankel, The Geometry of Physics: an Introduction, 2nd ed. Cambridge,
the injectivity radius for most practical DGMs; see Figure 16. U.K.: Cambridge Univ. Press, 2004.
Shifting the focus away from the general tracking prob- [7] S.P. Bhat and D.S. Bernstein, “A topological obstruction to continuous
global stabilization of rotational motion and the unwinding phenomenon,”
lem to setpoint-to-setpoint maneuvers, the cut locus of the Syst. Control Lett., vol. 39, pp. 63–70, 2000.
DGT determines the stability region of DGT for the inertia- [8] J.R. Munkres, Topology, 2nd ed. Englewood Cliffs, NJ: Prentice-Hall, 1999.
based PD control design of the previous section. The soft- [9] M. Henle, A Combinatorial Introduction to Topology. New York: Dover, 1979.
[10] L.S. Pontryagin, Foundations of Combinatorial Topology (transl: by F.
ware package Loki [23] can be used to compute the cut locus Bagemihl and W. Seidel). New York: Dover, 1952. .
of an abstract, two-dimensional Riemannian manifold. This [11] R. Bishop and S. Goldberg, Tensor Analysis on Manifolds. New York:
software requires only that the metric coefficients gij or a Dover, 1980.
[12] M.P. DoCarmo, Riemannian Geometry. Cambridge, MA: Birkhäuser, 1992.
parameterization of the two-dimensional manifold be given. [13] B. O’Neill, Semi-Riemannian Geometry with Applications to Relativity. San
For the DGT with metric coefficient mij given in (7), the cut Diego, CA: Academic, 1983.
locus of an arbitrary point is two rings as shown in Figure 16. [14] J. Milnor, Morse Theory. Princeton, NJ: Princeton Univ. Press, 1963.
[15] M.P. DoCarmo, Differential Geometry of Curves and Surfaces. Englewood
It follows that setpoint-to-setpoint control is indeed possible Cliffs, NJ: Prentice-Hall, 1976.
when the two rings are avoided for every point on the geo- [16] F. Bullo and A.D. Lewis., Geometric Control of Mechanical Systems. New
desic segments connecting the two setpoints, see Figure 17. York: Springer-Verlag, 2005.
[17] A.M. Bloch, Nonholonomic Mechanics and Control. New York: Springer-
A sequence of energy-efficient, setpoint-to-setpoint opera- Verlag, 2003.
tions for a nonflat DGM is illustrated in Figure 17. [18] F. Bullo, R.M. Murray, and A. Sarti, “Control on the sphere and reduced
attitude stabilization,” in Proc. Symp. Nonlinear Control Systems, vol. 2, Tahoe
City, CA, June 1995, pp. 495–501.
CONCLUSIONS [19] F. Bullo and R.M. Murray, “Tracking for fully actuated mechanical sys-
Topology and geometry play a crucial role in the analysis of tems: A geometric framework,” Automatica, vol. 35, pp. 17–34, 1999.
the global, nonlinear properties of mechanical systems. The [20] R.M. Murray, Z. Li, and S. Sastry, A Mathematical Introduction to Robotic
Manipulation. Boca Raton, FL: CRC, 1994.
double-gimbal system provides an elementary but techno- [21] R. Fuentes, G. Hicks, and J. Osborne, “Intrinsic control design for
logically important example of how these mathematical abstract machines,” in Lecture Notes in Control and Information Sciences, vol.
fields are intertwined with dynamics and control design. 366. New York: Springer-Verlag, 2007.
[22] M. Berger, A Panoramic View of Riemannian Geometry. New York: Springer-
Since the simplest mechanical system fits within the general Verlag, 2003.
framework of mechanics, we need only recognize that such [23] R. Sinclair and M. Tanaka, “Loki: Software for computing cut loci,”
mathematical concepts are extensible. The most difficult Exper. Math., vol. 11, no. 1, p. 25, 2002.

challenge confronting the mainstream application of geome-


try and topology in mechanics and control is its accessibili- AUTHOR INFORMATION
ty, not its relevance, to the engineering community. Jason M. Osborne (jason.m.osborne@gmail.com) received a
Ph.D from North Carolina State University in mathematics
ACKNOWLEDGMENTS in 2007 and a B.S. in mathematics and physics from East
The authors were supported by Raytheon Missile Systems, Tennessee State University. His research interests include
General Dynamics, NSF (DMS-0306017), Air Force Office of geometric control of holonomic and nonholonomic
Scientific Research, and the Air Force Research Laboratory. mechanical systems. He can be contacted at 208 Love Build-
We thank these fine organizations for their patronage. The ing, 1017 Academic Way, Tallahassee, FL 32306-4510 USA.
first author would like to thank the Space Scholars Program Gregory P. Hicks is a principal aeronautical engineer
at the Air Force Research Laboratory for the opportunity to with General Dynamics Advanced Information Systems.
begin the research of this article. The first author would fur- He received the Ph.D. from North Carolina State Universi-
ther like to thank Carlos Miranda for his invaluable help ty in 2003. From 2003 until 2005 he was a National
and patience in the three-dimensional rendering of the Research Council research associate with the United States
cover image. Without his expertise in Lightwave 3D, the Air Force Research Laboratory. His scientific interests lie
cover image would not have been possible. with the fields of dynamics and controls.
Robert Fuentes is a senior systems engineer at Raytheon
Missile Systems. He received his Ph.D. from the University
REFERENCES
[1] J.M. Hilkert, “Inertially stabilized platform technology concepts and prin- of Colorado at Boulder. His research interests include adap-
ciples,” IEEE Contr. Syst. Mag., vol. 28, no. 1, pp. 26–46, 2008. tive and geometric methods in mechanics and control.

64 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2008

You might also like