Anda di halaman 1dari 11


4, AUGUST 1993


Robust Impedance Shaping Telemanipulation

J. Edward Colgate

Abstract-When a human operator performs a task via a

bilateral manipulator, the feel of the task is embodied in
the mechanical impedance of the manipulator. Ikaditiody, a
bilateral manipulator is designed for transparency; i.e., so that
the impedance reflected through the manipulator closely approximates that of the task. Impedance shaping bilateral control,
introduced here, differs in that it treats the bilateral manipulator
as a means of constructively altering the impedance of a task. This
concept is partidarly valuable if the characteristic dimensions
(e.g., force, length, time) of the task impedance are very different
from those of the human limb. It is shown that a general form
of impedance shaping control consists of a conventional powerscaling bilateral controller augmented with a real-time interactive
task simulation (i.e., a virtual environment). An approach to
impedance shaping based on kinematic similarity between tasks
of different scale is introduced and illustrated with an example. It
is shown that an important consideration in impedance shaping
controller design is robustness; i.e., guaranteeing the stability of
the operatodmanipulatodtasksystem. A general condition for the
robustness of a bilateral manipulator is derived. This condition is
based on the structured singular value (p). An example of robust
impedance shaping bilateral control is presented and discussed.



HUMAN performing a manual task will typically associate a feel with the task. For instance, the feel of
carving wood is very different when performed with a sharp
knife than with a dull one. Feel relates to the manner in which
the environment responds to the motions and forces generated
by the human. As such, feel can be described by a causal
dynamic operator that relates a response (output) variable
to a stimulus (input) variable. For instance, if the output
is taken to be force and the input velocity, the appropriate
operator is an impedance [24]. Different selections of input and
output variables may lead to admittance, hybrid, scattering, or
other descriptions of task feel. In this paper, unless otherwise
specified, the term impedance will refer generally to an
operator that embodies feel.
When a bilateral manipulator is inserted between a human
and environment, it will alter the feel of the task. In most
traditional applications of bilateral manipulation, a design goal
is to achieve transparency; i.e., minimize the extent to which
the bilateral manipulator alters the feel of the task. But it is also
possible to make intentional use of a bilateral manipulator as
a means of altering feel. This concept is particularly valuable
if the characteristic dimensions (e.g., force, length, time) of
Manuscript received November 13, 1991; revised March 4, 1993. This work
was supported in part by the Engineering Foundation, Grant RI-A-90-4 and
in part by the National Science Foundation, Grant MSS-9022513.
The author is with the Department of Mechanical Engineering, Northwestem University, 2145 Sheridan Road,Evanston, IL 60208.
IEEE Log Number 9212605.
A causal dynamic operator is one in which the present value of the output
may depend on present and past values of the input.

the task impedance are very different from those of the human
In this paper, the use of a bilateral manipulator to shape*
the perceived impedance of the environment is discussed. An
approach to impedance shaping is presented and illustrated
with an example. The example, however, also illustrates
a potential difficulty with impedance shaping. A bilateral
manipulator designed to shape the impedance of a particular
environment in a particular frequency range may, for another
environment or another frequency range, exhibit non-robust
behavior. Results from robust control theory, including the
structured singular value (p), are used to develop a powerful
stability criterion. Finally, an example of a robust impedance
shaping controller is given.
A. Power and Impedance Scaling in Bilateral Manipulation

This work is motivated largely by its application to the

class of bilateral manipulators in which master and slave
operate on very different length, force, and power scales. This
class includes strength-increasing man-amplifiers [3 11 or
extenders [27] as well as dexterity-increasing macro-micro
bilateral manipulators (MMBMs) [9]. [20], [25]. Unlike more
conventional master-slave manipulators, the major impetus for
such devices is not to enable remote manipulation, but rather
is to extend the capabilities of the human limb. As such, these
devices usually exhibit some measure of power and impedance
A basic understanding of power and impedance scaling may
be gained from Fig. 1. In this figure, an ideal bilateral
manipulator is shown connecting a 1-port operator and a 1port environment. The power transfer from the operator to
the manipulator is the product of effort and flow, $ 1 ~ 1 , while
the power transfer from the environment to the manipulator is
4 2 ~ 2[32]. These are related by the scale factors that define
the bilateral control law:




The ratio, k E / k + ,therefore, is known as the power scaling

factor. The impedance scaling factor, k,k+, is derived from
the relation of the impedance felt by the operator to the
impedance of the environment:

= k,k+Ze ( 4 2 )


where Zap, stands for apparent impedance. It is evident that,

with proper selection of k, and k+, the power and impedance
scale factors may be independently set.
2The term shape originates from the effect that the bilateral manipulator
has on the frequency response plots of linear impedance-in general, these
plots will be not only shifted, but reshaped as well.

1042-296X/93$03.00 0 1993 IEEE


Pon I

I- ,


devices were also studied in the 1980s; notably, the development of improved myoelectric controllers for prosthetic limbs



Bilateral Manipulator


Fig. 1. Block diagram of a one degree-of-freedom teleoperator system with

an ideal (dynamics-less) power scaling bilateral manipulator. q5 is a flow
variable and E is an effort; Yo is the admittance of the operator, and 2,
is the (linear) impedance of the environment. k+ and k, are dimensionless,
static scaling factors. The power flow convention used throughout this paper
can be seen in this figure: the efforts and flows associated with all physical
subsystems (i.e., operator, manipulator, etc.) are defined such that their inner
products are the power flows into the subsystems. Subsystems are connected
with common-flow energic junctions (the bond graph I-junction 1321 ) to
assure correct signs.

Before proceeding, it is perhaps worth noting that ordinary

passive hand tools may be classified as manipulators capable
of scaling impedance, but not power. For instance, the feel
of tightening a nut is certainly different with and without a
wrench. Also in this class are conventional micromanipulators
which employ transmissions (e.g., fine-pitch lead screws) to
scale down human motion^.^ The utility of such relatively
simple devices is evident as they are routinely used for
micropuncture of arteries less than 100 pm in diameter [21].
The idea of using bilateral manipulation to provide power
scaling appears to date to the early 1960s and the man
amplifiers built at General Electric [31]. Man amplifiers were
intended to give the ordinary man extraordinary strength,
presumably supplanting the need for an array of powered
tools, such as the hydraulic lift. The same period gave rise
to cybernetic prostheses, which were to give amputees benefits beyond those of body-powered prostheses, by providing
greater strength and more natural means of control [34].
Neither man amplifiers nor cybernetic prostheses met with
much initial success, judged in terms of user acceptance.
Common to both seemed to be the problem of mental fatigue:
they simply required too much concentration on the part of
the operator.
In the 1980s, the design of more natural human-machine
interfaces gave new life to both concepts. Kazerooni introduced the extender, which is similar to the man amplifier;
however, by being intimately connected to the powered limb
of the extender, the operator communicates with it via borh
power and information [27]. In effect, the extender is a
power-assist, not unlike a power steering system. In a careful set of experiments, Doubler and Childress demonstrated
that a very similar concept, called extended physiological
proprioception (EFT), provided significant benefits for the
control of upper-extremity prostheses [15]. It is important to
recognize that neither extenders nor EPP seek to provide a
limb replacement; rather, both seek to provide a very natural
limb extension, in much the same sense that a tennis racket
is an extension to the arm of an accomplished player. Other
approaches to achieving natural control of power-assistive


Power attenuating bilateral manipulators (MMBMs) appear

to be of more recent vintage. In 1987, Fukuda, er al. [20]
described a one degree of freedom system in which the master
and slave were parallel-jaw grippers. Stable gripping of both
hard and soft objects was demonstrated. In 1989, Hunter, et
al. [25] described a six degree of freedom system that was
developed for the mechanical testing of individual muscle
cells. In the authors laboratory, a four degree of freedom
MMBM is now under development as a tool for the study of
dexterity enhancement. The master manipulandum is described
in [29], and the slave is described in [22].
The historical review above suggests that the development
of power scaling bilateral manipulators may be considered an
ongoing effort to improve scaled manipulation via enhanced
sensory feedback. For instance, an MMBM is a natural step
beyond a fine-pitch lead screw: while both devices enable
high resolution positioning, only the MMBM provides useful
kinesthetic feedback. It has been rigorously demonstrated that
the information contained in scaled kinesthetic feedback is
useful (i.e., improves performance), at least in the case of
EPP control [ 151.
Impedance shaping is offered here as a potential evolutionary step beyond power scaling manipulation. The idea
underlying impedance shaping bilateral control is that a priori
information, in the form of a task model, may be used to
enhance performance further. In other words, the impedance
shaping bilateral controller seeks to improve performance by
adding to the sensor-based feedbacks (vision and force) a
model-based feedback (force). This idea has one clear precedent in the work of Herzog [23], who demonstrated improved
manual control when the impedance of a joystick was designed to approximate that of the plant being controlled. This,
however, was not an example of bilateral control, because
sensor-based force feedback was absent. Another precedent is
model-based control (including loopshaping techniques [ 1SI),
although the analogy to impedance shaping cannot be taken
very far. For instance, the ultimate design goals in impedance
shaping and loopshaping are rather different. In one respect,
however, there is notable similarity: an essential tradeoff in
both methods is between aggressiveness and robustness.
Impedance shaping is an approach to bilateral manipulation,
but is not a design methodology. It lacks, for instance, a
general approach to performance specification (this depends
on the task at hand and what measure of performance is
to be optimized). Nonetheless, a design methodology which
incorporates impedance shaping is entirely feasible. As an illustration of this, a specific methodology based upon kinematic
similarity is developed below and in Section IV.
A. Similarity-A

3The apparent impedance in this case is dominated by manipulators

intrinsic impedance (inetia, friction) and is not at all indicative of environment

Basis for Impedance Shaping Design

Similarity rules are commonly used in model testing [35].

For instance, a physical model and full-scale prototype are

. _..._



said to be geometrically similar if all length scale ratios (e.g.

height over width) are the same in both. Kinematic similarity
requires that length scale ratios are the same in model and
prototype and time scale ratios are also the same in model
and prototype. Dynamic similarity requires that this be true
of length, time, and force scale ratios. For linear systems
represented by transfer functions, dynamic similarity implies
that the poles and zeros of the model are the same as those
of the prototype.
Geometric similarity is enforced in most implementations of
telemanipulation. In other words, the length ratios experienced
at the master are in fixed proportion to those of the task.
This is consistent with the use of visual feedback, which
produces an image that is geometrically similar to the task.
Most implementations of bilateral manipulation attempt to
enforce dynamic similarity as well. For instance, the "ideal"
power-scaling manipulator illustrated in Fig. 1 generates an
apparent impedance which is simply that of the environment
times a constant. If these impedances are linear, it is clear
that the poles and zeros are unaffected. Of course, dynamic
similarity is impossible to achieve in practice due to the effect
of manipulator dynamics on the apparent impedance (this
point will be clarified in the examples that follow), but it is
commonly taken as a design goal, termed "transparency" [ 5 ] .
An obvious limitation of dynamic similarity, however, is
that the poles and zeros of the apparent impedance cannot be
altered relative to those of the environment. Stated differently,
the forces contributed by various physical effects (e.g., inertia,
stiffness, friction) cannot be independently scaled. Yet, in the
context of micromanipulation, it may sometimes be desirable
to scale inertial forces, for instance, to a greater extent than,
say, surface tension forces. This may be done in an effort to
make the microdynamical system feel like a macrodynamical
system. One rational approach to this type of scaling problem is to enforce (via impedance shaping bilateral control)
kinematic similarity but not dynamic similarity between the
apparent impedance and the actual impedance.
The example below illustrates the distinction between conventional (dynamically similar) bilateral manipulation and
impedance shaping (kinematically similar) bilateral manipulation. It also provides the beginnings of a design methodology
for impedance shaping control.



"Ideal"Impedance Shaping BilateralController

Fig. 2. Block diagram of an "ideal" impedance shaping bilateral controller
attached to an environment.

The force relation, however, is more complicated. Because

mass scales as length cubed, one sensible choice of the force
scale might be k, = a4,which would result in an impedance
scaling of k+kE = a3. This would make the perceived mass
equal to the expected mass of the object seen in the display,
assuming equivalent densities. The viscous forces on the mass,
however, vary according to length, so that the appropriate
scaling of the viscous component of the impedance would be
k+k, = a. The question of which of these scalings to select
has no clear answer. Any fixed scaling has the disadvantage
that the apparent impedance will be dynamically similar to
the environment impedance, and therefore will feel more
sluggish than a kinematically similar macrosystem. In other
words, a fixed scaling cannot alter the inherent time constant
( T ~= me/be)of the environment.
Now suppose ?at the MMBM incorporates an impedance
shaping term, AZe(s), as in Fig. 2. The following procedure
will enforce kinematic similarity but not dynamic similarity.
a Select a geometric scale factor, a. The time scale factor
is unity (because manipulation must occur in real-time,
any time scaling is prohibited).
a Construct a model of the environment. In this case:
Ze(s) = h e s

+ it?


+ abe


Select k+ = l / a .
Select the minimum k, necessary to match one term of
k,k+Ze(s) to the corresponding term of Z,"'(s). In this
case, k, = a2 leads to:

k,k+Z,(~)= (


where heand ie are estimates of me and be.

Determine the kinematically similar, scaled impedance.
In this case:
Z p ( s ) = a3yjLes

B. Design Example
To begin, assume that the MMBM may be represented,
as in Fig. 1, by the two scale factors k, and k+. Suppose,
further, that the environment consists simply of a small block
(mass me)immersed in a viscous liquid, and that the Reynolds
number is low enough that drag forces scale linearly with the
block's length scale and velocity [35](viscous drag coefficient
be). The MMBM is to be used for manipulating the block.
If a visual display system is used to magnify the image
of this microenvironment by a certain factor a, then a good
choice of the velocity scale factor is k+ = l / a . This choice
would ensure that the perception of a 1:1 relationship between
the position of the slave's endpoint and the position of the
master's handle is maintained, as in EPP (of course, in actual
implementation, position control would be required).



+ ab,

Y ~ , s


The damping term matches that of Z,"'(s).

Select the impedance shaping term AZe(s) according to:

AZe(s) = k~'k~'Z~'(s) - Z e ( s )


In this case:

AZe(s) = ( a 2- l)yjLes


The result of this design procedure is an apparent impedance

having the following form:

+ Z,(s)).

= tk+(AZe(s) ze(s))
= a ( @ - l)yjLes





Slave Controller

Bilateral Controller
Impedance Shaping Bilateral Controller
n = 100
bm= 1.0 N-dm
bs = 0.001 N-dm
Bs = 282 N-dm Ks = 8ooO N/m
0: = 25
m, = 0.5 kg
mS = 0.005 kg

Fig. 3. A one degree-of-freedom impedance shaping MMBM featuring master and slave dynamics.

As the estimate Ze(5) approaches the actual impedance Z, (s),

the apparent impedance will approach the kinematically similar impedance. It is important to recognize, however, that a
precise model of the actual impedance is not essential to the
success of the method. The apparent impedance in (8) consists
of two parts: the actual environment impedance scaled by a
fixed factor of a, and the impedance of a simulated inertia
- l ) h e s ) ,also scaled by a. The latter may be recognized
as a simple form of virtual environment; i.e., an interactive,
real-time simulation [2]. Note that this construction of Zapp(s)
will produce the desired result of amplifying inertia to a
greater extent than damping (thereby increasing the apparent
time constant of the environment), even if the-estimate heis
considerably in error, and even if the estimate be is completely
in error.
C. Discussion

Even if unknown a priori, the environment dynamics may

be estimated on-line. It is not uncommon for a person performing a delicate task to begin by gently probing the environment.
A micromanipulator might do something similar, in order to
build an environment model. Another possibility is that the
bilateral controller include a parameter adaptation scheme that
runs continuously. One difficulty with adaptive techniques is
that they are inherently nonlinear, making performance and
stability analyses difficult. In this paper only linear timeinvariant bilateral controllers are considered-the design and
analysis of nonlinear, time-varying controllers are subjects of
future research.
In the example below, a somewhat more realistic MMBM
model involving master and slave dynamics is considered. This
example illustrates the difficulty of maintaining robustness.

The example above demonstrates that impedance shaping

depends on the availability of an environment model. The force A. Example-Poor Robustness
generated by the environment is, in general, the outcome of
A block diagram of a somewhat realistic impedance shaping
multiple physical effects. If each of these effects is subject implementation is shown in Fig. 3. The master is modeled
to a distinct scaling law, then they must be separated by as a controlled force source acting on an inertia m, with
some means in order to implement impedance shaping. It is a viscous damping b,. The slave is modeled as a controlled
reasonable conjecture that a model is the only such means, force source acting on an inertia m, with viscous damping
unless internal states of the environment can be measured b,, connected to its load via a transmission of gear ratio n. A
PI velocity controller (effectively a PD position controller) is
The need for a model highlights a potential weakness of the closed around the slave. The impedance shaping controller is
impedance shaping concept: a standard reason for choosing similar to that in Fig. 2; however, the velocity feedback term
teleoperation rather than autonomous operation is that the is taken from the actual velocity of the slave which will, in
environment is unstructured. Nonetheless, every environment general, differ from the commanded velocity.
will exhibit some structure, and what information is available
Fig. 4 compares the apparent admittance (inverse of
should be used. For instance, in microelectronics assembly, impedance) of an environment (the one considered above with
masses and coefficients of friction may well be known a me = 5 grams, be = 1 N-s/m) felt through the bilateral
priori, and in microsurgery, tissue properties may be known. manipulator to the actual admittance of the environment, the
In general, impedance shaping may be made more or less admittance that would be felt through the ideal impedance
aggressive, according to the knowledge of the environment. shaping controller of Fig. 2, and the admittance that would
Furthermore, so long as the bilateral controller meets a non- be felt if the impedance shaping were turned off. Note that
restrictive robustness criterion presented in Section 111, its the impedance shaping controller has the effects of reducing
behavior will be benign, even if the estimated and actual the admittance (amplifying the impedance) and reducing the
environment dynamics differ radically (as may occur when comer frequency (this corresponds to amplifying inertia more
the slave suddenly contacts a rigid surface).
than damping); the non-shaping controller does only the




A precise definition of robustness is provided below, but it

is easily illustrated that this design is not robust. Although

the impedance shaping controller is designed for a specific
microenvironment, the characteristics of this environment are
subject to rapid and dramatic change. Perhaps the best illustration of this is making or breaking contact with a rigid
surface. To extend this example, suppose that contact is made
with a surface of stiffness IC, = 5 x lo4 N/m. The isolated
MMBM with impedance shaping control will remain stable
when contacting this surface; however, when the inertia of the
operator's arm is included, the entire system may lose stability.
As a specific example, suppose that the arm introduces an
inertia of m, = 2 kg. The system will exhibit unstable
oscillations at a frequency of about 125 rad/= (20 Hz)
which the operator will find disturbing, but in all likelihood,
absolutely uncontrollable.
B. Approach to Robustness








-. ..
. :1__

frequency ( d m - )


Fig. 4. Environment admittance (upper dashed line); apparent admittance

via bilateral manipulator with no shaping (dot dash); apparent admittance
via ideal impedance shaping bilateral manipulator pictured in Fig. 2 (lower
dashed); apparent admittance via bilateral manipulator pictured in Fig. 3
(solid). Notice that all apparent admittances are far less than the actual
admittance, however, the comer frequency (at which inertial effects become
important) is moved significantly only if impedance shaping is used. Notice
that a realistic impedance shaping design (solid) is close to ideal (dashed) for
a finite frequency range; at high enough frequency, master inertia dominates
the admittance of the realistic design.

The ideas presented in this section draw heavily on recent

(since 1982) advances in the field of robust control. The basic
approach is to identify the class of manipulator/environment
behaviors (admittances) with which a human can be expected
to interact successfully, and to identify the class of all possible
environment behaviors. Then conditions may be placed on the
manipulator behavior to ensure that, for environment behaviors
within the latter class, manipulator/environmentbehaviors will
always fall within the former class.
It is a matter of common observation that human operators have no difficulty maintaining stability when interacting
with all manner of passive tools and environments. Indeed,
even greater skill can be expected: some tools, such as
rotary sanders and floor waxers, exhibit what are obviously
not passive impedances, but can, with some training, be
operated quite effectively. Nonetheless, the broadest class
of tooYenvironment behaviors which can, at this point, be
analytically characterized as stable under human control, is
the set of passive behaviors. A reasonable choice for the class
of environment impedances is also that of passive impedances. This has been motivated elsewhere [13]. Therefore,
the following intuitive condition for stability/robustness will
be used: a bilateral manipulator is said to be robust iJ; when
coupled to any passive environment, it presents to the operator
an impedance (admittance) which is passive. Thus, while the
apparent impedance felt through the bilateral manipulator may
exhibit a radically different magnitude, or even shape in
the frequency domain, than the environment impedance, it
must at least be passive. The mathematical statement of this
condition, and the derivation of a simple test for it, follow the
preliminaries below.

former! The impedance shaping controller appears to work

well: its admittance is comparable to that of the ideal controller
over a substantial frequency range, though at high enough
frequency it is dominated by the inertia of the master (mm).
Of course, the high-frequency dominance of inertia is an
effect which must occur in any real hand tool or bilateral C. Mathematical Preliminaries
Impedances, Admittances, and Hybrid Matrices -The terms
It is important, however, to assess the robustness of the
impedance shaping controller before passing final judgment. impedance and admittance may be applied to nonlinear systems as well as linear systems. Impedances and admittances
4The nonshaping controller does in fact change the shape of the admitance are causal dynamic operators that map vectors of inputs (4 and
frequency response plots, as seen in Fig. 4. This change is due to the inherent E , respectively) to vectors of outputs ( E and 4, respectively):
dynamics of the master and slave manipulators which affect the apparent
impedance, but do not contribute to intentional shaping.

= Z(+)

+ = Y(&)




In the case of linear, time-invariant (LTI) systems, these

operators reduce to transfer function matrices:
4s) = Z(S)4(S)

4(s) =Y(S)E(S).


Associated with each pair ( E ; , 4i),where E ; and 4i are the ith

components of the vectors E and 4, respectively, is the notion
of a port. A physical system connects to its environment at
ports; the product ~ i d iis the power flowing into the system
at port i.
Not all multiport physical systems can be represented by
impedances or admittances. An example is the two-port transformer, which can be represented by the hybrid matrix:

but not by any impedance or admittance matrix [33]. It has

been demonstrated that, for any LTI multiport, a hybrid matrix
will exist [3]. The general form of a hybrid matrix is:

where qTqi is the total power input.

Passivity and Positive Real Matrices-A passive multiport
is one for which, given zero energy storage at t = 0 [14]:

2 0 V Q d t ) t, L 0.


This definition applies equally to nonlinear and linear systems. Its interpretation is simply that finite energy may never
be extracted from a passive system. The hybrid matrix of a
passive LTI multiport must be positive real:

H ( s ) stable ; H * ( s )+ H ( s ) 2 0 for Re{s} > 0. (14)

A test for positive realness is that the multivariable Nyquist
plots of H ( s ) must lie in the closed right half plane [8].
Scatrering Matrices-The interaction between physical systems is fundamentally bilateral; therefore, an effort ( E ) may
never exist independent of a flow ($), and vice versa. Efforts
and flows, however, are by no means the only pair of variables
that may be used to describe bilateral interaction. Any pair
related to E and 4 through a nonsingular linear transformation
is equally legitimate. Another useful set of variables is (G,
20 ), the inwave and outwave or scattering variables
[33]. Scattering variables relate to efforts and flows by


1 1 1
= 3 [ 1 -1][;]

LTI multiports may always be represented by scattering matrices:


w (s) = S ( s ) G ( s ) .


Passivity and Bounded Real Matrices: In terms of scattering

variables, a passive multiport is one for which, given zero
energy storage at t = 0 :

(GT7i?- f T 5 ) &
2 0 V $ ( t ) , t 2 0.


Fig. 5. (a) A generalized description of an operator-MMBM-environment

system in terms of scattering variables and scattering operators. (b) An
equivalent system description (except that the operator has been replaced
with a pseudo-operator).

This condition applies equally to linear and nonlinear systems.

The scattering matrix of a LTI multiport must bebounded real:
S(s) stable;

I - S*(s)S(s) 2 0

for Re{s}.


A test for bounded realness is that the maximum singular value

of S ( j w ) be less than or equal to one at all frequencies, or
equivalently, that the infinity norm of S ( j w ) be less than or
equal to one.
D. Robustness Criterion

The robustness criterion to be developed will apply to

bilateral manipulators that may be described by hybrid or scattering matrices; the assumption will also be made (temporarily)
that the environment is also linear. Suppose that the bilateral
manipulator and environment are described by the following
scattering matrices (see also Fig. 5(a)):
bilateral manipulator: f ( s ) =
environment: Se(s)


where P ( s ) is partitioned according to master ports (subscript

1) and slave ports (subscript 2). The intuitive robustness
criterion given at the beginning of this section may be stated
in terms of these matrices as shown as follows.
Robustness Criterion-Version

The P ( s ) contains no poles in the closed right half plane,

llp11+ P12Se(I- p22Se)-1pz111mI

v llSelloo I


The dependence upon the Laplace variable, s, is understood.

The expression on the left is simply the infinity norm of the



scattering matrix felt at the operator ports; it is a linear

fractional transformation of Se [17]. This statement of the
robustness criterion, however, is of little computational value,
because the norm must be computed for all n x n Se of infinity
norm less than or equal to one. A condition in terms of P ( s )
alone would be preferable.
The key to finding such a condition lies in recognizing that
the conditions sought are equivalent to the conditions for the
stability of the telemanipulator/environment when coupled to
a pseudo-operator which is passive, but otherwise arbitrary.
This can be understood by replacing the operator block in Fig.
5(a) with an arbitrary passive n-port, then asking the question,
under what conditions on P ( s ) is stability guaranteed? The
conditions are the same as those sought here, because a
necessary and sufficient condition to guarantee the stability5
of a LTI system coupled to an arbitrary passive n-port, is
that the system itself appear to be passive [8]. Thus, by
proving coupled stability-the stability of the telemanipulator
coupled to an environment and pseudo-operator which are
both passive but otherwise arbitrary-the apparent passivity
of the telemanipulator/environmentis implicitly proved. This
reasoning leads to the revised block diagram in Fig. 5(b), and
to a second statement of the robustness condition:
Robustness Criterion-Version 2

The P ( s ) is stable when coupled to any 2n-port that meets

the following criteria on structure and infinity norm:

Notice that IISstrctIlm
5 1 (the 2n-port is passive) if and only
if IISpseudol(,
5 1 and IISellm5 1 (the pseudo-operator and
environment are both passive).
Version is no
a condition on p ( s )done than Version
1; however, the problem c m now be recognized as a structured
uncertainty problem, which may be addressed by Doyles
structured singular value [16] or Boyds structured Lyapunov
functions [6]. The structured singular value of P , p ( P ) , is
defined as
0 if no A E X, solves det(I PA) = 0
PL(p) = ( A
J - ~ ( A det(l
) I + PA) = 0) )-I
where X,(s) is defined as the class of 2n-port block diagonal
above) with no restriction
matrices (same structure as Sstrct,
on infinity norm, and F(A) is the maximum singular value
of A. A useful coupled stability condition can be stated in
terms of p:

(s), above) iff

passive, block diagonal 2n-port (Sstrct


A third version of the robustness criterion follows immediately.
Robustness Criterion-Version 3

The matrix P ( s ) contains no poles in the closed right half

plane, and:

The proof of the coupled stability theorem, therefore the
robustness condition, is straightforward with the assistance of
the multivariable Nyquist Theorem [28], because the definition
of p is essentially tautological. It may appear, in fact, that
the introduction of the structured singular value has been
counterproductive, as all perturbations A E X, must be
considered. Certain properties of p ( P ) , however, make it
possible to arrive at a condition that may be computed in
terms of P ( s ) alone.
The useful properties are the following:

ii) p(DPD-) = p ( P ) where D has the form:


d l I0n x n

0X n ] > d l > d z



where 8 again refers to the maximum singular value. The first

from the definition Of p(>? the
Property follOws
second from the fact that D-AD = where A E xm
Together these properties
p ( M ) 5 a(DMD-1).

While this result is only a modest advance, it can be shown

that, if the right hand side is minimized over all allowable D ,
the equality holds. This result was first proved by Doyle, who
also showed that the optimization problem is convex, ensuring
the tractable computation of p ( P ) 1161. Thus, a useful form
Of the robustness condition is finally achieved as fOllOWS..
Robustness Criterion-Version 4

Coupled Stability Criterion for Block Diagonal 2n-Port


The P ( s ) contains no poles in the closed right half plane,


The 2n-port system with scattering matrix G(s) will be

guaranteed to remain stable when coupled to an arbitrary,
51n this paper, stability means that all the poles of the LTI system lie
in the closed left half plane. This includes what is sometimes known as
marginal stability, and does not guarantee asymptotic stability. A guarantee of
asymptotic stability requires that the pseudo-operatorenvironment be strictly
passive according to criterion given in [12].


p inf

where p = dl/dz.







Remark 1: For p = 1, the condition above would be

equivalent to the passivity of P. In macro-micro manipulation,
P will never be passive due to the amplification inherent in
the scale factors IC, and IC+(IC,/k+ > 1). Thus, the freedom
provided by the parameter ,l3 is essential to the utility of the
above condition.
In [6], Boyd and Yang demonstrate via Lyapunov and
absolute stability theory, that if D exists such that DPD-
is bounded real, then P will remain stable when coupled to
any nonlinear, passive, block-diagonal 2n-port; i.e., a causal
dynamic operator satisfying the passivity condition of (13),
and the structure condition of (21). This condition is more
conservative than Doyles only in that D is not frequency
dependent. The corresponding robustness condition is shown
as follows.

frequency (radsec)

Robustness Criterion-Version 5 (Nonlinear Environments)

The matrix P ( s ) contains no poles in the closed right half

plane, and p > 0 exists such that:

Fig. 6. Plot of the structured singular value ( p ) associated with the bilateral
manipulator pictured in Fig. 3. p must be less than or equal to one at all
frequencies to ensure robustness. This design is not robust.

( + ~ ( s ) / E ~ ( sin) Fig. 3) should, for connection to all admissible

be positive real. This admittance consists
is bounded real.
U of the sum of two parts: the intrinsic admittance of the
servo-controlled slave, and the admittance of the feedback
Boyd and Yang also show that an equivalent condition exists
loop via the bilateral controller, maskdoperator, and slave.
in terms of hybrid matrices. It is perfectly analogous, and can
The former is typically positive real while the latter, due to
be obtained by replacing P with H , and bounded real with
noncollocation, cannot be. However, in part due to the large
positive real in the statement above. A similar result, which
gear ratio, the admittance of the slave is very small (i.e., it is
does not use input-output descriptions at all, but instead relies
nonbackdriveable), thus the total admittance is dominated by
on a network representation of an operator-telemanipulatorthat part which is not positive real. This difficulty is exactly
environment system, is described by Anderson in [4].
analogous to the compliance control of a nonbackdriveable
Remark 2: In addition to nonlinear environments, the conrobot, which is discussed in [7].
dition is useful in dealing with nonlinear manipulator dynamOf the two factors, the more difficult to contend with
ics. It is often possible to treat a bilateral manipulator as a
is generally nonbackdriveability. In [lo], a similar example
pair of 2n-port nonlinear admittances (master and slave mechwhich includes noncollocation, but no gear ratio, is considered.
anisms) sandwiching a 2n-port LTI controller. This breakdown
It is shown that a low pass filter in the force feedback path is
allows the nonlinear mechanism dynamics to be absorbed into sufficient to provide robustness. Even so, the design process
the operator and environment descriptions.
employed is ad hoc and not readily generalized to more
In the next section, these results will be used to analyze
complex situations.
impedance shaping controllers.
Further strides can be made by eliminating both nonbackdriveability and noncollocation. An example of such a
design is illustrated in Fig. 7. The placement and role of the
impedance shaping term are now quite intuitive: as in the
The nonrobustness of the MMBM considered in Section III- ideal case (see Fig. 2), the shaping term is a real-time enviA can be verified by computing the structured singular value ronment simulation that augments the physical environment.
of its scattering matrix. The result is displayed in Fig. 6. An This term is readily implemented as part of the collocated
attempt was made to improve the robustness properties of this feedback of the slave; i.e., the slave control law becomes
MMBM by including various forms of compensation in both
the feedforward (velocity) and feedback (force) paths. Fairly
) - B8
- AZ,(s) 42(5)
extensive exploration failed to produce any design that would E,o,trol(S) = &PI & A Z ( ~ =
satisfy the robustness criterion over all frequencies, although
violations could be pushed to quite high frequency. This where 4 2 is the velocity of the slave, and E control is the
problem appears to stem from two factors: noncollocation [ 191 control force applied to the slave which is the sum of PI control
and nonbackdriveability. Noncollocation describes a situation (&PI)and impedance shaping ( E A Z ) terms.
The force feedback signal is based on the PI control effort
in which significant dynamics separates sensor and actuator
signals; nonbackdriveability describes the property of low rather than the force at the slave/environmentinterface, making
output admittance in a manipulator. In terms of these factors, it effectively collocated with the velocity of the master, 41.
the difficulty may be understood as follows: the admittance Note that, due to the low inertia and backdriveability of the
that the bilateral manipulator presents to the environment slave, the PI control effort will, under most circumstances,






Slave Controller


Slave Plant
Impedance Shaping Bilateral Controller

Bilateral Controller



a = 25
bm = 1.O N-s/m
br = 0.001 N-dm
Bs = 400 N-s/m Ks = 4OOO N/m

m, = 0.5 kg
ms = 0.005 kg

Fig. 7. An architecture for robust impedance shaping bilateral control.

approximate the interface force (&PI x 2 - E A Z ) . Collocated

force feedback has little effect on performance: the performance of this design is shown in Fig. 8, which compares
quite well to Fig. 4.
The real benefit of this approach is robustness, as illustrated
in Fig. 9. It is interesting to note that the optimal scaling p
is found to be p = a-1.5independent of frequency, so that
the design is robust for even nonlinear environments. It may
appear from Fig. 9, however, that the design is only marginally
robust, as the structured singular value is approximately unity
across the frequency spectrum. Somewhat surprisingly, this is
the desired result. As an illustration of this point, consider
a bilateral manipulator that behaves as a rigid body (many
hand tools behave approximately as rigid bodies). Because
such a manipulator would be lossless, its structured singular
value would be one at all frequencies. Greater robustness could
be achieved by adding internal dissipation, but this is not
necessarily desirable.
A final point regarding the architecture f f Fig. 7 is that
robustness will be guaranteed whenever AZ, is positive real
(it can then be lumped together with 2, without altering the
assumption of passivity). Thus, the shaping term should be selected to supplement the passive dynamics of the environment,
but not subtract from those dynamics.
The principal contributions of this paper have been the
introduction of impedance shaping bilateral control and the
development of a powerful robustness criterion for bilateral
manipulation. A design approach based on kinematic similarity
has emerged through the examples. This approach may be
summarized as follows.
Select a scale factor a and design an impedance shaping
term (AZ,) according to the procedure given in Section
11-B. The shaping term should be positive real.
Using a backdriveable slave manipulator, implement bilateral control as indicated in (29) and Fig. 7 (collocated
position and velocity feedback for servo control of slave;
collocated feedback for impedance shaping; and constant
scale factors).
While this approach may serve as the basis for a formal
methodology, issues stemming from master and slave dynam-

frrquency ( d s e c )

Fig. 8. Environment admittance (upper dashed line); apparent admittance
via ideal impedance shaping bilateral manipulator pictured in Fig. 2 (lower
dashed); apparent admittance via bilateral manipulator pictured in Fig. 7
(solid). Notice similarity to Fig. 4; the revised control architecture has not
degraded performance.

ics, multiple degrees of freedom, and discrete time controller

implementation need to be thoroughly addressed. Moreover,




[ l ] C. J. Abul-Haj, Elbow-prosthesis emulation: A technique for the
quantitative assessment of an assistive device, Ph.D. dissertation,Mass.
Inst. Technol., Cambridge, MA, 1987.
[2] B. D. Adelstein, A virtual environment system for the study of human
arm tremor, Ph.D. dissertation, Mass. Inst. Technol.,Cambridge, MA,,
[3] B. D. 0. Anderson, R. W. Newcomb and J. K. Zuidweg, On the
existence of H matrices, IEEE Trans. Circuit Theory, vol. CT-13, no.
1, pp. 109-110, 1966.
[4] R. J. Anderson and M. W. Spong, Asymptotic stability for force
reflecting teleoperators with time delay, in P m . IEEE Int. Con$
Robotics Automat., 1989, pp. 618-625.
[5] C. Andriot and R. Foumier, Bilateral control of teleoperators with
flexible joints by the H infinity approach, in SPIE Con$ 1833: Telemanipulutor Technol., Boston, 1992, pp. 80-91.
[6] S. Boyd and Q. Yang, Structured and simultaneous lyapunov functions
for system stability problems, In?. J. Control, vol. 49, no. 6, pp.
2215-2240, 1989.
[7] E. Colgate, On the intrinsic limitations of force feedback compliance controllers, in Robotics Res.-l989, K. Youcef-Toumi and H.
Kazemni, Eds. New York: ASME, 1989.
[8] I. E. Colgate, Coupled stability of multiport systems-Theory and
experiments, J. Dyn. Syst., Meus. Contr., in press.
[9] -,
Power and impedana scaling in bilateral manipulation, in
Proc. IEEE Int. Con$ Robotics Automat., 1991, pp. 2292-2297.
[ 101 -,
Robust impedance shaping via bilateral manipulation, in Proc.
American Control Con$, 1991, pp. 3070-3071.
[I I] J. E. Colgate, Stability of manipulators interacting with unstructured and structured dynamic environments, In Proc. IMACS/SlCE Int.
Symp. Robotics, Mechutronics. and Manu$ Syst., Kobe, Japan, 1992 pp.
Strictly positive real admittances for coupled stability, J.
[12] -,
Franklin Inst., vol. 329, no. 3, pp. 429444, 1992.
[13] J. E. Colgate and N. Hogan, Robust control of dynamically interacting
systems, In?. J. Control, vol. 48, no. 1, pp. 65-88, 1988.
[I41 C. A. Desoer and M. Vidyasagar, Feedbnck Sysrems: Input-Output
Properties. New York: Academic, 1975.
[15] J. A. Doubler and D. S. Chddress, An analysis of extended physiological proprioception as a prosthesis control technique, J. Rehub. Res.
undDevelop., vol. 21, no. 1, pp. 5-18, 1984.
[I61 J. C. Doyle, Analysis of feedback systems with structured uncertainties, in Proc. Inst. Elect. Eng., Purt D, vol. 129, no. 6, 1981, pp.
[17] J. C. Doyle, Structured uncertainty in control system design, in in
Proc. 24th Con$ Decision and Contr., 1985, pp. 260-265.
[ 181 J. C. Doyle and G. Stein, Multivariable feedback design: concepts for a
classicaYmodem synthesis, IEEE Trans. Automat. Contr., vol. 26, pp.
4-16, 1981.
[19] S. D. Eppinger and W. P. Seering, Three dynamic problems in robot
force control, IEEE Trans. Robotics Automat., vol. 8, no. 6, pp.
751-758, 1992.
[20] T. Fukuda, K. Tanie and T. Mitsuoka, A new method of master-slave
type of teleoperation for a micro-manipulator system, in Proc. IEEE
Micro Robots und Teleop. Workshop, Hyannis, MA, 1987.
[21] M. R. Glucksberg, Personal communication.
[22] K. W. Grace, J. E. Colgate, M. R. Glucksberg and J. H. Chun, A six
degree of freedom micromanipulator for ophthalmic surgery, in Proc.
IEEE Int. Con$ Robotics Automat., 1993, pp. 630-635.
[23] J. H. Herzog, Manual control using the matched manipulator control
technique, IEEE Trans. Man-Machine Syst., vol. MMS-9, no. 3, pp.
56-60, 1%8.
[24] N. Hogan, Controlling impedance at the madmachine interface, in
Proc. IEEE In?. Con$ Robotics Automat., 1989, pp. 1626-1631.
[25] I. W. Hunter, S. Lafontaine, P. M. F. Nielsen, P. J. Hunter and
J . M. Hollerbach, A tele-microrobot for manipulation and dynamic
mechanical testing of single living cells, in Proc. IEEE Micro Electro
Mech. Syst. Con$, 1989, pp. 102-106.
[26] S. C. Jacobsen and R. W. Mann, Control systems for artificial arms,
in Proc. IEEE Con$ Man, Syst., and Cybernetics, 1973.
[27] H. Kazerooni, Human machine interaction via the transfer of power and
information signals, IEEE Trans. Syst,, Man. Cybern., vol. SMC-20,
pp. 450-463, 1990.
[28] J. M. Maciejowski, Multivariable Feedback Design. Reading, MA:
Addison-Wesley, 1989.
[29] P. A. Millman and J. E. Colgate, Design of a four degree-of-freedom
force-reflecting manipulandum with a specified forcdtorque workspace,
in Proc. IEEE Int. Con$ Robotics Automat., 1991, pp. 1488-1493.









frequency (ndlsec)

Fig. 9. Structured singular value plot for bilateral manipulator pictured in

Fig. 7. This design is robust, because p is slightly less than 1 at all frequencies.

the notion of kinematic similarity is only one, and by no means

an optimum, approach to impedance shaping.
Ultimately the utility of the impedance shaping concept,
and of any other approaches to enhancing operator dexterity,
must be borne out experimentally. An MMBM comprising
a four degree-of-freedom master manipulandum, a six degreeof-freedom slave manipulator, and a transputer-based real time
controller, is being developed for this purpose [30]. Because
the impedance shaping term is essentially a virtual environment, this system is also able to execute real-time virtual
environment simulations, outputting forces of interaction via
the master manipulandum.
The robustness criterion presented here is broadly applicable to problems in telemanipulation; however, it remains
conservative in the following respects: it assumes that the
environment is unstructured, and it requires that the apparent
impedance felt through the bilateral manipulator be passive.
As pointed out in the discussion of impedance shaping, some
understanding of environment structure generally does exist,
and should be exploited. The appropriate restriction on the
apparent impedance merits experimental investigation, and it
may be found that the restriction to passivity can be relaxed.
Unfortunately, removing that which is conservative in the
problem statement does not mean a more powerful robustness
criterion will be found. The mathematics may simply not be
tractable. To address this issue, a generalized description of
physical (bilateral) interaction, based on neither hybrid nor
scattering descriptions, is being developed. The idea is to
select that description which optimally encodes the available
information (e.g., passivity, environment dynamics, etc.) into
a simple measure such as the infinity norm. This would enable
the robustness theory presented here to be extended to a significantly larger class of environment, operator, and manipulator
behaviors. Preliminary results have been presented in [ 1 I].
The author also thanks Ambarish Goswami for his careful
reading of this manuscript.



P. A. Millman, M. C. Stanley, P.E. Grafing and J. E. Colgate, A system

for the implementation and kinesthetic display of virtual environments,
in SPIE Con$ 1833: Telemanipulator Technology,, Boston 1992, pp.
R. S. Mosher, From handiman to hardiman, Soc. of A~tomor.Eng.
Trans., vol. 16, pp. 588-597, 1967.
H. M. Paynter, Analysis and Design of Engineering Systems. Cambridge, M A MIT Press, 1961.
H. M. Paynter and I. J. Busch-Vishniac, Wave-scattering approaches
to conservation and causality, J. Franklin Inst., vol. 325, no. 3, pp.
295-313, 1988.
N. Weiner, Cybernetics. New York Wiley, 1961.
F. M. White, Fluid Mechanics. New York McGraw-Hill, 1979.

J. Edward Colgate received the S.B. degree in

physics and the S.M. and Ph.D. degrees in Mechanical Engineering from tie Massachusetts Institute of
Technology, Cambridge, MA.
Since 1988, he has been an Assistant Professor
in the Department of Mechanical Engineering at
Northwestem University, Evanston, IL.His areas of
interest include haptic interface to teleoperators and
virtual environments, robotic tools for microsurgery,
electrostatic actuators, and modeling and control of
dynamic systems.