! !

! !

Stability of Haptic Displays

D. W. Weir and J. E. Colgate

This chapter reviews the issue of instability in haptic devices, as well as

the related concept of Z-width. Methods for improving haptic display per-
formance (expanding the Z-width) are also discussed.

8.1 Definitions
Haptic displays can be considered to be devices which generate mechanical
impedances. ‘Impedance’ here is defined as a dynamic relationship between
velocity and force. The behavior of the haptic display depends on the
virtual environment being rendered. For instance, if the desired behavior
is that of a point mass, the haptic display must exert forces proportional
to acceleration. Similarly, if the desired behavior is that of a spring, the
haptic display must exert forces proportional to displacement [Colgate and
Brown 94].
Passivity has proved to be a useful tool for studying both the stability
and performance of haptic displays. A one-port system is passive if the
integral of the power extracted over time does not exceed the initial energy
stored in the system. For a translational mechanical system, power is the
product of force (f ) and velocity (ẋ), with the sign convention that power
is positive when energy flows into the system. Typically, the initial energy
is defined to be zero, resulting in the following inequality:
! t
f (τ )ẋ(τ )dτ ≥ 0, ∀t ≥ 0 (8.1)
0

A passive system, coupled with any other passive system, is necessarily

stable. Ordinary physical objects, such as springs, masses, and dampers,
are passive, and common experience suggests that humans remain stable
when interacting with passive systems. Therefore, the human user is typi-
cally considered a passive impedance, particularly at high frequencies above
the bandwidth of voluntary motion. If a haptic display rendering an ar-
bitrary virtual environment can be guaranteed passive, then the complete
system will be stable when the display is coupled with the human operator.
This property frees the designer from having to analyze the interaction of

151

! !

! !
 ! !

! !

152 8. Stability of Haptic Displays

the haptic display and virtual environment with the human operator under
all possible configurations.
In the real world, objects interact according to a set of physical laws
that govern their behavior. In the virtual world, this interaction is only
approximated. Even though the approximate behavior may be very close
to the real behavior, the implications of these errors can be profound. In-
stability and limit cycle oscillations are two common ways in which haptic
interactions deviate from their physical counterparts, both of which result
from non-passivity. Small amplitude limit cycle oscillations can be partic-
ularly problematic even if they do not escalate to gross instability because
human tactile perception is extremely sensitive to vibrations in the 100Hz
to 1kHz range [Bolanowski et al. 88]. Maintaining passivity is one way,
albeit sometimes restrictive, of ensuring that virtual objects behave in a
stable manner when interacting.
Everyday interaction with common objects involves experiencing a wide
range of impedances. Moving in free space implies almost zero resistance
to motion, while interacting with tables, walls, and other massive objects
provide almost complete resistance to motion. The challenge is to design
a haptic interface that can display as wide a range of dynamic impedances
as possible.
The dynamic range of impedances that can be rendered by a haptic dis-
play while maintaining passivity is termed its Z-width. Since a display with
larger Z-width will usually render “better” feeling virtual environments, Z-
width may be viewed as a measure of quality for the haptic display.
As a final note, we should mention that haptic displays are often re-
ferred to as “impedance type” or “admittance type.” Impedance displays
measure the endpoint motion and output a force or torque in response.
Admittance displays measure the applied force or torque and output a
motion. Both systems respond according to the (imperfectly) simulated
physics of the virtual environment being rendered. This chapter will ad-
dress both impedance and admittance displays, but will focus primarily
on impedance causality displays. Note, however, that for either type, the
notions of passivity and Z-width are equally valid.

8.2 Designing for Passivity

Expanding the impedance range of a haptic display as a method for im-
proving performance begins with passivity. Maintaining passivity places
severe restrictions on virtual environment stiffness and damping; therefore,
a number of techniques have been developed to facilitate haptic rendering
of high impedance environments.
Due to the nature of impedance causality haptic displays, the lower

! !

! !
 ! !

! !

8.3. Passive Rendering of a Virtual Wall 153

bound on impedance is generally limited by the quality of force sensing and

feedback, and the mechanical design. Often, impedance causality displays
feature low inertia designs enabling low impedance renderings. The upper
bound on passive impedance can be limited by sensor quantization, sampled
data effects, time delay, and noise [Colgate and Schenkel 97]. Thus, most
research efforts have focused on increasing the maximum impedance that
can be displayed as a way of increasing the Z-width of haptic displays.
A number of methods exist to increase the maximum passive impedance
of a haptic interface. These fall into a number of broad categories: con-
trollers, physical mechanisms, and electrical mechanisms. The category of
controllers includes virtual couplings and passivity observers. Virtual cou-
plings act as mediators between the haptic display and the virtual environ-
ment. Passivity observers and passivity controllers function by adjusting
the energy present in the system to maintain passivity. Mechanical meth-
ods are generally the most direct, whereby physical dissipation is added to
the mechanism to expand the passive impedance range of a haptic display
by counteracting the effects of energy leaks. Electrical methods are a blend
of physical methods implemented electrically and controller approaches im-
plemented using analog electronics.
In a slightly different category are psychophysical techniques that act
to alter the user’s perception of the impedance range of the haptic display.
These include methods such as rate hardness and event-based rendering.

8.3 Passive Rendering of a Virtual Wall

8.3.1 A Simple Passivity Result
Haptic displays are sampled-data systems, i.e., they combine a continuous-
time mechanical system with a discrete-time controller. The effects of sam-
pling, even assuming ideal sensors and actuators in the continuous-time
plant, cause a haptic display to lose passivity.
[Colgate and Schenkel 97] derive an analytical passivity criterion for a
simple 1 degree of freedom (DOF) haptic interface, as shown in Figure 8.1.
The discrete-time controller models a virtual wall, including a unilateral
constraint operator and includes analog-to-digital (A/D) and digital-to-
analog (D/A) converters in the feedback loop. A block diagram for this
sampled-data system is shown in Figure 8.2. The unilateral constraint is
chosen as a fundamental building block for virtual environments, because
it models a simple form of contact and collision between two objects.
A necessary and sufficient condition for passivity of the sampled data
system in Figure 8.1 is:

! !

! !
 ! !

! !

154 8. Stability of Haptic Displays

displacement
x,v actuator
transducer

D/A program A/D

Clock

Figure 8.1: A simple 1 DOF haptic display [Colgate and Schenkel 97].

operator
-

f v
u +
1 - e -Ts - 1 1
s ms + b s
zero order hold

xs x
H(z)
T

unilateral constraint

Figure 8.2: Block diagram of a haptic display and operator sampled-data

system [Colgate and Schenkel 97].

T 1
b> ${(1 − e−jωT )H(ejωT )} for 0 ≤ ω ≤ ωN (8.2)
2 1 − cos(ωT )
where b is the physical damping present in the mechanism, T is the sampling
rate, H(z) is a pulse transfer function representing the virtual environment,
and ωN = Tπ is the Nyquist frequency [Colgate and Schenkel 97].
The result can be simplified to an analytical expression relating the

! !

! !
 ! !

! !

8.3. Passive Rendering of a Virtual Wall 155

sampling rate, virtual stiffness, virtual damping, and dissipation within

the haptic display. [Colgate and Schenkel 97] analyze a wall consisting of
a virtual spring and damper in mechanical parallel, together with a uni-
lateral constraint operator. A velocity estimate is obtained from backward
difference differentiation of the position sensor data. This results in the
following transfer function within the wall:

z−1
H(z) = K + B (8.3)
Tz
where K > 0 is the virtual stiffness and B is the virtual damping coefficient
(B is allowed to be positive or negative). Equation 8.2 combined with
Equation 8.3 simplifies to the following passivity condition:
[Colgate and Schenkel 97]

KT
b> + |B| (8.4)
2

The physical damping present in the mechanism must be sufficient to

dissipate the excess energy created by errors introduced by sampling in the
discrete-time controller, commonly referred to as “energy leaks.”

8.3.2 Importance of Damping

The physical damping present in the haptic display is critically important,
due to its role in counteracting the energy generation from errors intro-
duced by sensing and discrete-time control. [Colgate et al. 93] expand on
the passivity bound of Equation 8.4 and provide simulation data showing
how maximizing sensor resolution and minimizing sampling rate improves
performance. Colgate and co-authors also introduce the concept of adding
physical damping to the system in order to increase the limits of virtual
stiffness and virtual damping that can be passively achieved [Colgate et al.
93, Colgate and Schenkel 97, Colgate and Brown 94].
The implications of Equation 8.4 are somewhat counterintuitive: to
increase the maximum impedance of a haptic display, increase the vis-
cous damping in the mechanism in order to maintain passivity. The ad-
dition of physical damping can dramatically increase the maximum pas-
sive impedance a device can render. When low impedances are rendered,
virtual damping in the discrete-time controller can be negative, masking
the increased physical damping in the device. However, simulated or vir-
tual damping cannot substitute for real, physical dissipation in the mecha-
nism [Colgate and Brown 94]. Physical damping can be added to the haptic
interface through a variety of techniques that will be discussed in Part II.

! !

! !
 ! !

! !

156 8. Stability of Haptic Displays

first detected extra-wall position which

toggles off control law

0.5 manipulandum displacement y(t)

displacement and force

-0.5

-1

-1.5
commanded force fk(t)
fk = 1 * y k

-2
1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3
time
∆ta ∆tb

Figure 8.3: The effect of sampling: actual position, sampled position, com-
manded force [Gillespie and Cutkosky 96].

8.3.3 Virtual Wall as a Benchmark

The virtual wall is the standard haptic task. Since most interaction with
virtual environments can be simplified to interaction with a virtual wall of
varying stiffness and damping, the virtual wall is commonly used as a per-
formance benchmark for haptic interfaces. For example, see [Colgate and
Brown 94], [Gillespie and Cutkosky 96], [Zilles and Salisbury 95], [Adams
and Hannaford 99], and [Abbott and Okamura 05].
Due to the nature of sampling, simulating the behavior of a stiff virtual
wall is a difficult task. To characterize the general problem, consider the
following example. As a general rule, there is always some penetration of
the position of the haptic display into the virtual wall. As a consequence,
at the next sampling interval, the discrete controller detects the wall pene-
tration, and the virtual environment computes large output forces normal
to the wall surface. This large force has a tendency to rapidly push the
haptic display outside of the virtual wall into free space. This situation
now reverses, where at some future sampling interval, the position of the
haptic display is outside the virtual wall, so the forces return to zero. This
sequence is depicted in Figure 8.3. Oscillations arise when this cycle of
free space and wall penetration is repeated. Sampling prevents detecting
the exact time when the haptic display contacts the surface of the virtual
wall, and position sensing resolution has the effect of quantizing penetra-

! !

! !
 ! !

! !

8.3. Passive Rendering of a Virtual Wall 157

Velocity filter, High resolution encoders

600
Physical damper engaged (1 KHz)
Physical damper engaged (100 Hz)

500
Virtual Stiffness (Nm/rad)
No physical damping (1 KHz)
No physical damping (100 Hz)

400
300
200
100
0

-0.5 0 0.5 1 1.5 2

Virtual Damping (Nm-sec/rad)

Figure 8.4: Typical Z-width plot illustrating maximum passive impedance

range [Colgate and Brown 94]. (! c 1994 IEEE)

tion distance into the virtual wall, both of which are destabilizing effects.
These errors can lead to energy generation and active, non-passive behav-
ior. These effects will be further addressed in in the next section.
The virtual wall is also traditionally used to characterize the impedance
range, or Z-width of haptic interfaces. Z-width is often displayed us-
ing virtual stiffness-virtual damping plots, showing the maximum passive
impedance boundary as the stiffness and damping vary, typically under a
variety of conditions, as shown in Figure 8.4.
However this method does not show how the Z-width varies according
to frequency. It also does not show the minimum stable impedance that can
be rendered. The importance of this is illustrated in the following example.
If a single haptic display has maximum and minimum impedances of Zmin
and Zmax , respectively, then two of them in mechanical parallel will have a
maximum impedance of 2Zmax , increasing the boundary on the K-B plot.
The minimum impedance is also increased to 2Zmin , so the system Z-
width has not changed, but this is not apparent on the K-B plot. This lack
of minimum impedance information makes it difficult to compare various
haptic interfaces.
For these reasons, a more useful figure of merit and way of displaying
Z-width information may be a set of curves, showing the extremes of both
impedance and admittance as a function of frequency while maintaining
passivity.

! !

! !
 ! !

! !

158 8. Stability of Haptic Displays

Figure 8.5: Mapping between actual position and quantized position, with
sensor resolution ∆ [Abbott and Okamura 05]. (!c 2005 IEEE)

8.4 Extensions to the Passivity Framework

8.4.1 Quantization and Time Delay
The most common position-sensing technique for haptic displays is the
use of optical encoders. One consequence of optical encoders is that po-
sition information is quantized based on the encoder resolution. Other
position-sensing techniques are also frequently quantized, such as analog
potentiometers that are sampled by a finite resolution analog-to-digital
converter. Such a position signal would also be subject to electrical noise,
but that will be not be addressed here. The distinction between sampling
and sensor quantization should be emphasized. Sampling introduces un-
certainty with respect to when events occur and what happens between
sampling intervals. Sensor quantization causes a loss of information due
to sensing only discrete changes in the value of a signal, as indicated in
Figure 8.5. The actual position can lie anywhere between two quantized
position measurements. Sensor quantization is independent of the sampling
frequency.
In [Abbott and Okamura 05], position quantization and Coulomb-plus-
viscous friction in the haptic device are explicitly modeled, as shown in
Figures 8.3, 8.5, and 8.6. Analyzing the worst-case scenarios of compressing
and extending a virtual spring, representing the virtual wall with a haptic
display, results in this passivity condition:
" #
2b 2fc
K ≤ min , (8.5)
T ∆

! !

! !
 ! !

! !

8.4. Extensions to the Passivity Framework 159

Figure 8.6: Model of haptic device rendering a virtual wall used by [Abbott
and Okamura 05]. (! c 2005 IEEE)

where b is the viscous damping in the mechanism, T is the sampling time,

∆ is the position quantization interval, and fc is the Coulomb friction. The
haptic display is assumed to consist of a mass plus friction, and the virtual
wall consists of a unilateral constraint. The first part of the inequality,
2b
T , is the same as Equation 8.4 when the virtual damping is equal to
zero. The stiffness is limited by the physical damping in the system, which
must be sufficient to dissipate at least as much energy as the energy leaks
introduced by sampling. The second term of the inequality, 2f ∆ , relates the
c

Coulomb friction in the device to the encoder resolution. It should be noted

that normally one of the terms is the dominating effect and provides the
limiting factor for passive virtual stiffness. In the experimental verification
of this passivity condition presented by [Abbott and Okamura 05], the
maximum virtual stiffness limited by damping and sampling rate, 2b T , is
almost two orders of magnitude smaller than the Coulomb friction limited
virtual stiffness, 2f
∆ .
c

Consider the following simplified conceptual derivation of the passiv-

ity criterion in Equation 8.5 to provide an intuitive understanding of the
passivity limit of virtual stiffness. Imagine compressing an ideal spring
with constitutive law F = kx. The energy stored in the ideal spring after
compressing a distance ∆x = xk+1 − xk = vT during one sampling period
is:

1
E= k∆x2 (8.6)
2

! !

! !
 ! !

! !

160 8. Stability of Haptic Displays

 






  

Figure 8.7: Detail of energy leak due to sampling.

Due to sampling, the force of the virtual spring remains constant be-
tween sampling intervals, as shown in Figure 8.7. Equation 8.7 is the re-
sulting energy leak due to sampling, while at the same time, Equation 8.8
is the energy dissipated by viscous damping (assuming constant intersam-
ple velocity). In order to maintain passivity, the energy dissipated must
be greater than the energy introduced by the energy leak (Equation 8.9);
therefore, it is possible to calculate the maximum passive virtual stiffness,
given the sampling rate and the physical dissipation (damping) in the me-
chanical system (Equation 8.11).

1
Eleak = K(vT )2 (8.7)
2
Edissip = bv 2 T (8.8)
Eleak ≤ Edissip (8.9)
1
Kv 2 T 2 ≤ T bv 2 (8.10)
2
2b
K ≤ (8.11)
T
A similar derivation can be made for the virtual stiffness limit due to
friction and quantization interval. Continuing with the conceptual example
of rendering an ideal spring, the position of the haptic display can change
to a distance equal to the quantization interval, ∆, without being sensed.
This would introduce an energy leak equal to the compression of the ideal
spring by a distance ∆ (see Equation 8.12). The friction in the mechanism
must dissipate at least as much energy as that introduced by the energy
leak, which is the work done by the friction force (Equation 8.13). This
inequality leads to a maximum passive virtual stiffness given the position

! !

! !
 ! !

! !

8.4. Extensions to the Passivity Framework 161

sensing quantization and the friction in the mechanism (Equation 8.16).

1
Eleak = K∆2 (8.12)
2
Edissip = fc ∆ (8.13)
Eleak ≤ Edissip (8.14)
1 ∆
K∆2 ≤ fc ( )T (8.15)
2 T
2fc
K ≤ (8.16)
∆
Equation 8.5 can be nondimensionalized by dividing by 2K. The two
resulting terms, β and σ, are used as axes to define a nondimensional plane
depicting stability regions according to behavior, shown in Figure 8.8. This
is a graphical way of depicting Equation 8.5. [Diolaiti et al. 06] analyze a
similar system with the added inclusion of time delay and introduce a new
nondimensionalized velocity parameter, ξ. ˙

b
β := (8.17)
KT
fc
σ := (8.18)
K∆
x
position ξ := (8.19)
∆
t
time τ := (8.20)
T
˙ ) ẋT
velocity ξ(τ = (8.21)
∆
One advantage of this plot is the identification of varying types of insta-
bility between regions of the plane. ξ˙ defines a new type of behavior: it is
the maximum allowed velocity of the haptic display, faster than which the
small effect of Coulombic friction and virtual environment parameters can
cause instability. The stability boundaries at β = σ = 12 correspond to the
effective dissipation limits for ensuring passivity, with β representing the
effective limit for viscous dissipation and σ corresponding to the effective
limit for Coulombic dissipation.
Quantization also limits performance through velocity estimation. Con-
sider, for example, a slowly changing position signal with a very fast sam-
pling rate. The finite difference method for estimating velocity is:
yk − yk−1
v̂k = (8.22)
T

! !

! !
 ! !

! !

162 8. Stability of Haptic Displays

ξ̇max +1 E Locally
2
Stable
A Globally
Stable (Passive)

D Loc.
Unstable −ξ˙max
1
2 C Globally B Limit
Unstable Cycles

1
2 β

Figure 8.8: Dimensionless stability plane with characteristic regions for

b
zero delay where β := KT and σ := K∆c ˙ ) = ẋT [Diolaiti et al. 06].
and ξ(τ ∆
(!
c 2006 IEEE)

If at sample times tk−2 and tk−1 the position information remains con-
stant, v̂k−1 = 0. However, if at sample time tk the position increases by one
quanta, δ, then the resulting velocity suddenly jumps to a very large value,
v̂k = Tδ . This rapidly varying velocity estimate can lead to instability. One
common method to reduce this effect is to low-pass filter the resulting veloc-
ity signal, thereby smoothing out the jumps. With increasing sample rate,
filtering becomes more imperative to obtain velocity signals. This presents
a trade-off, however, as increased filtering leads to increased time delay and
phase distortion, which can cause instability. The precision of the velocity
estimate improves with decreased sample rate, as illustrated in Figure 8.9.
However the reliability of the signal decreases due to the longer time delay.
This has the effect of averaging the velocity over a longer period of time,
or over a number of samples, as shown in Figure 8.9 and Equation 8.23.
n−1
1$ yk − yk−n
v̂k = v̂k−j = (8.23)
n j=0 nT

Fixed filters, such as a Butterworth filter, compute velocity from a

!
weighted sum of the raw velocity signal, v̂j , and past filtered velocity esti-
mates, v̂j .

! !

! !
 ! !

! !

8.4. Extensions to the Passivity Framework 163

Figure 8.9: Effect of window length on the variance of velocity [Janabi-

Sharifi et al. 00]. (!
c 2000 IEEE)

n
$ n
$
!
v̂k = bj v̂k−j + aj v̂k−j (8.24)
j=0 j=1

where aj and bj are the filter coefficients, and n is the order of the filter. As
n increases, the filter becomes more like an ideal low-pass filter; however the
delay and phase distortion are also increased. An additional subtlety is that
the signal is filtered along with the noise, so that heavy filtering leads to
poor transient response. To address this, [Janabi-Sharifi et al. 00] introduce
a velocity filtering technique that relies on a first-order adaptive window
length. The basic concept is that, when position signals are changing slowly,
the window should be long to provide a precise estimate of the velocity.
However, when the position is rapidly changing, the window length should
be short to improve velocity reliability and prevent introduction of excessive
delay. The window criterion exists to determine whether the slope of a
straight line reliably approximates the derivative of the signal between two
samples, xk and xk−n . If the noise, d, in the position signal can be assumed
to be uniformly distributed, such that d = &ek &∞ ∀k, then mathematically,
the adaptive window problem becomes finding a solution for the largest
possible window length n that satisfies the following:

|yk−i − Lyk−i | ≤ d, ∀i ∈ {1, 2, . . . , n} (8.25)

where Lyk−i = an + bn (k − i)T, given that (8.26)
kyk−n + (n − k)yk
an = and (8.27)
n
n
$ $n
n yk−i − 2 iyk−i
i=0 i=0
v̂k = bn = (8.28)
T n(n + 1)(n + 2)/6
The solution for the window length, n, is found iteratively where the
window grows from n = 1 until the window no longer fits the enclosed
data; then the previous n is used to compute the velocity estimate. bn is
the slope of a line that is a least-square approximation that minimizes the
error in the velocity signal [Janabi-Sharifi et al. 00].

! !

! !
 ! !

! !

164 8. Stability of Haptic Displays

8.4.2 Nonlinearities
Nonlinearities are an important consideration for haptic displays in virtual
environments. Essentially, almost all useful virtual environments are non-
linear in that impedances change dramatically upon contact with objects in
the virtual environment. [Miller et al. 00] analyze the passivity of nonlin-
ear delayed and non-delayed virtual environments. The authors establish
a passivity criterion relating the haptic display and human operator, the
virtual coupling, and the virtual environment for both delayed and non-
delayed environments. Virtual couplings will be introduced in more detail
in Part II of this chapter.
Again, the physical dissipation in the mechanism is a critical param-
eter. In addition to the passivity criterion, a key result is a limit to an
environment parameter, α, measuring the lack of passivity exhibited by the
virtual environment. It can be expressed as a function of inertia, damping
and stiffness parameters. α is related to the physical dissipation in the
system, δ, and is modulated by the impedance of the virtual coupling γ, if
present [Miller et al. 00]:

α < δ (8.29)
δγ
α < (8.30)
δ+γ

Many common haptic devices also have nonlinear kinematics. Through

an analysis of system dynamics, [Miller et al. 04] shows how the nonlinear
transformation from joint space to task space for a haptic display also
affects passivity. This result can be summarized by the following inequality:

δm ≥ J T δJ (8.31)
where δm represents the joint space dissipation, J is the haptic interface
Jacobian, and δ is the task space dissipation required for passive rendering
of the desired virtual environment.

8.5 Control Methods

8.5.1 Virtual Coupling
Virtual coupling is one of the basic techniques for rendering virtual en-
vironments in haptics, introduced by [Colgate et al. 95] and amplified
by [Miller et al. 00,Adams and Hannaford 98] and others. The virtual cou-
pling connects the haptic display and the virtual environment, and consists

! !

! !
 ! !

! !

8.5. Control Methods 165

"Virtual Coupling"
K

operator

haptic display Passive

Tool Simulation

Figure 8.10: The virtual coupling [Colgate et al. 95]. (!

c 1995 IEEE)

of a virtual spring in virtual damper in mechanical parallel, as shown in

Figure 8.10.
The virtual coupling is advantageous because it simplifies the problem
of ensuring stability. Using a virtual coupling to establish stability of the
haptic display, which is a sampled-data system, it is only necessary to
satisfy the following two conditions:
1. Select the virtual coupling parameters, such that a virtual wall with
these parameters would be passive.
2. Make the virtual environment discrete-time passive.
Condition 2 is simpler to achieve than analyzing the complete sampled-
data system to ensure passivity. Separating the discrete-time passivity of
the virtual environment from the rest of the system frees the designer from
concerns regarding the interaction between the virtual environment and
the haptic display and human operator. The virtual coupling, however,
has the effect of reducing the maximum environment impedance to match
the passivity limits of the haptic display, which are generally lower than
the impedances of the virtual environment.
Virtual environments rendering mass require the use of discrete time
integrators which typically are not passive, making condition 2 difficult to
meet. [Brown and Colgate 98] analyzed various discrete time integration
techniques in the context of establishing the lower bound of virtual mass
that can be rendered while maintaining passivity. The value of minimum
mass required for passive rendering depends on the form of integrator used.
The work of [Miller et al. 00] generalized these results by explicitly
modeling the non-passivity of the human and haptic interface as well as the

! !

! !
 ! !

! !

166 8. Stability of Haptic Displays

Figure 8.11: The haptic display system with a virtual coupling [Miller et al.
00]. (!
c 2000 IEEE)

virtual environment (Figure 8.11). As shown in Equation 8.30, the virtual

coupling increases the allowed lack of passivity in the virtual environment
while still maintaining overall system passivity.
[Adams and Hannaford 98] introduced the use of a virtual coupling
network to analyze and guarantee system stability. Using this technique,
elements of the haptic display are typically modeled as a series of inter-
connected, two-port elements in a network, shown in Figure 8.12. The
virtual coupling introduced by [Colgate et al. 95] and the coupling behav-
ior of the god-object introduced by [Zilles and Salisbury 95] are subsets
of this more general two-port coupling network approach. Coupling net-
work results are shown for both admittance and impedance architectures.
This technique was then applied to a 2 degree of freedom haptic display in
both impedance and admittance configurations, showing passivity results
derived experimentally and theoretically for both conditions [Adams et al.
98].

8.5.2 Passivity Observers and Controllers

[Gillespie and Cutkosky 96] introduced a technique for stabilizing virtual
walls by compensating for the energy leaks due to the zero order hold as
well as the asynchronous switching times associated with sampling. Asyn-
chronous switching times arise because the haptic display generally does
not enter or exit the virtual wall exactly at a sampling time; typically the
transition from “outside” to “inside” the virtual wall occurs in between
sampling intervals (Figure 8.3). These two sources of error are treated sep-
arately. The goal is to design a digital controller to cancel the effects of
these induced energy leaks, stabilizing the system. Figure 8.13-A shows a

! !

! !
 ! !

! !

8.5. Control Methods 167

vh − vd vd − ve
-1
human + haptic virtual + virtual
operator Fh display coupling Fe environ-
- Fd - ment

haptic
interface
Fig. 3. The haptic interface for impedance display case
Figure 8.12: The virtual coupling as a two-port element in a network
[Adams and Hannaford 98]. (!c 1998 IEEE)

sampled-data system simulation of a ball bouncing on a surface, without

sampling correction.
The dominant behavior of the zero order hold can be approximated
as a half sample delay. By designing a controller that predicts the state
of the system one half sample forward in time, the majority of the error
introduced by the zero order hold can be canceled. At sample time t =
kT , the controller predicts the state at t = KT + T /2 and then renders
the virtual environment using the predicted system state. Figure 8.13-
B shows simulation results using a half sample prediction algorithm of
a sampled-data system rendering a bouncing ball. It can be seen that
modeling the zero order hold as a half sample delay improves the rendering
during the majority of the time the ball is in contact with the virtual wall.
However, during the last sample and while in contact with the wall, the
algorithm introduces an error, computing a force pulling back toward the
wall. This error occurs because the ball exits the virtual wall between
sampling intervals due to the secondary effect of sampling, asynchronous
switching times.
To address this second concern, a model of the system to predict thresh-
old crossing times using state information is also incorporated. Conceptu-
ally, this is estimating ta and tb in Figure 8.3, given a model of the known
properties of the system and virtual environment being rendered. Dead-
beat control is then used to compensate for the energy leaks caused by these
asynchronous switching times. Figure 8.13-C shows the final improvement
after correcting the half sample delay and using deadbeat control to correct
for asynchronous switching times. Note that correcting for the effects of
sampling is independent of the added problem of sensor quantization.
Expanding on this work and the work on virtual coupling networks,
[Hannaford et al. 01] introduced passivity observers (POs) and passivity
controllers (PCs) for stabilizing haptic interaction with virtual environ-
ments. Passivity observers analyze system behavior and track the energy

! !

! !
 ! !

! !

168 8. Stability of Haptic Displays

0.5 reference ball displacement

0.08
sampled data ball displacement
0.4 0.06
displacement (m) and scaled wall force (N)

displacement (m) and scaled wall force (N)

0.04
0.3 sampled data ball displacement

0.02
0.2 reference ball displacement

0
0.1
-0.02

0
-0.04

-0.1
f_wall -0.06 f_wall

-0.2 -0.08
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
time (seconds) time (seconds)
sampled data ball displacement
0.1
reference ball displacement

0.05
displacement (m) and scaled wall force (N)

-0.05

-0.1 f_wall

-0.15

-0.2

-0.25
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
time (seconds)

Figure 8.13: (a) Sampled-data system simulation of a bouncing ball. (b)

Half sample prediction simulation results of a bouncing ball. (c) Bouncing
ball simulation with sampling and zero order hold correction algorithm
active. [Gillespie and Cutkosky 96]

Figure 8.14: One-port network with passivity controller [Ryu et al. 04].
(!
c 2004 IEEE)

flow between elements to estimate errors introduced into the sampled-data

system. Passivity controllers act to dissipate this excess energy by adjust-
ing the impedance between elements in the system (Figure 8.14). They
effectively inject additional damping to dissipate energy.
One of the main advantages of POs is that the PC does not modify the

! !

! !
 ! !

! !

8.5. Control Methods 169

desired system impedance unless an energy correction is necessary. Unlike

the virtual coupling, which constantly moderates the feel of the virtual
environment, the passivity controller adds damping only when necessary
to counteract energy leaks. This can potentially lead to better feeling
virtual environments.
The earliest POs assumed that velocity and force was constant between
samples, but more recent passivity observers presented in [Ryu et al. 05],
based on [Ryu et al. 04] and [Stramigioli et al. 02], show that this assump-
tion can be relaxed. The resulting passivity observer for an impedance
causality device takes the following form:

 
$k
Eobsv (k) =  f (tj−1 )(x(tj ) − x(tj−1 )) + f (tj )(x(tj ) − x(tj−1 )) (8.32)
j=0

The bracketed term of Equation 8.32 represents the exact energy input
to the discrete-time virtual environment from time 0 to time tk , and the
second term is an estimate of the energy input one time step ahead, and
based on the assumption that the velocity does not change during that time
step. If the dynamics of the controller are much faster than the dynamics of
the mechanical system, then the predictive second term in Equation 8.32
is typically not necessary. If at any time the observed energy, Eobsv , is
negative, then the sampled-data system may be contributing to instability.
It is then the job of the PC to modify the impedance of the network to
dissipate the excess energy.
To further improve the performance of PO/PC systems and maintain
the perception of a good feeling virtual environment, the excess energy
should be dissipated smoothly. [Ryu et al. 05] introduced a PO that
smoothly corrects for energy leaks by modeling the behavior of a reference
system and comparing that to the observed behavior, shown in Figure 8.15.
For simple virtual environments, a model of the energy flow into the vir-
tual environment can be explicitly calculated to act as the reference energy.
However, most interesting virtual environments are nonlinear, making an
exact calculation of the energy flow into the virtual environment very diffi-
cult. In this case and in the case of designing a general passivity observer,
a simple energy model can be used to reference the behavior. One imple-
mentation of such an energy tracking reference is the numerical integration
of the power flow into the virtual environment, where the force is computed
given the observed position information.
In the case of a continuous and lossless one-port network system, the
energy input to the system should be equal to the energy stored, S, plus
the energy dissipated, D:

! !

! !
 ! !

! !

170 8. Stability of Haptic Displays

! t
f (τ )ẋ(τ )dτ = S(t) + D(t), ∀t ≥ 0 (8.33)
0
This leads to the following PC algorithm for the one-port network with
impedance causality shown in Figure 8.14 [Ryu et al. 05]. In this case, the
PO (Eobsv in step 4, Equation 8.34) uses the modeled energy, instead of
the one step ahead predicted energy in Equation 8.32.
1. x1 (k) = x2 (k) is the input.
2. ∆x(k) = x1 (k) − x1 (k − 1)
3. f2 (k) is the output of the one-port network.
4. The actual energy input at step k is:
k
$
Eobsv (k) = f1 (j − 1)∆x(j) (8.34)
j=0

5. S(k) and D(k) are the amount of stored energy and dissipated energy
of the virtual environment at step k, respectively.
6. The PC control force to make the actual input energy follow the
reference energy is calculated:
)
−(Eobsv (k)−S(k)−D(k))
∆x(k) if W (k) < 0
fP C (k) = (8.35)
0 if W (k) ≥ 0

where W (k) = Eobsv (k) − S(k) − D(k).

7. f1 (k) = f2 (k) + fP C (k) is the output.
Another improvement to the passivity observer gained by following the
energy of a reference system is the problem of resetting. Consider the case
of a virtual environment that is both highly dissipative in certain regions
and active in other regions. The active region requires the passivity con-
troller to add damping to maintain stability. If the user spends a long time
in the dissipative region before contacting the active region, a large accu-
mulation of positive energy in the passivity observer can be built up during
interaction with the dissipative region. This is very similar to the problem
of integrator wind-up. Upon switching to the active region, the passivity
observer may not act until the net energy becomes negative, causing a delay
while the accumulated excess of passivity is reduced. During that delay,
the system can exhibit unstable behavior. If the passivity observer tracks
a reference energy system, this problem of resetting can be avoided.

! !

! !
 ! !

! !

8.6. Extending Z-Width 171

Figure 8.15: Effect of passivity controller and modeled reference energy

[Ryu et al. 05]. (!
c 2005 IEEE)

Another method of tracking and dissipating energy leaks is presented

by [Stramigioli et al. 02]. This work uses a port-Hamiltonian method
for estimating these sampled-data system errors. The key aspect of all of
these energy leak and passivity controllers is determining the inaccuracy
introduced by the discrete-time approximation of the continuous system,
so that the controller can dissipate this excess energy.

8.6 Extending Z-Width

This section first extends the passivity criterion in Equation 8.2, and gives
insight into passivity design with frequency-dependent damping. Then,
mechanical and electrical methods of implementing high frequency damping
are reviewed.

8.6.1 Frequency-Dependent Passivity Criterion

A system in feedback with an uncertainty set consisting of all possible
passive behaviors must itself be strictly passive to guarantee closed loop
stability [Colgate and Hogan 88]. We use this fact to establish the strict
passivity of the haptic display model in Figure 8.16. Specifically, we replace
the block representing the human operator with a block containing the
uncertainty set Σ. Σ is the set of all linear, time-invariant (LTI), passive

! !

! !
 ! !

! !

172 8. Stability of Haptic Displays

Fh

Haptic Display
6
Fm - vh
Z h1 ( s )

1  e sT xh 1
H ( z)
s T s

Virtual Environment (digital)

Figure 8.16: Model of a haptic system. Zh is the impedance of the haptic

display hardware; H(z) is the (linear) virtual environment; Σ is the uncer-
tainty set that we use to replace the human operator. Note that we assume
the actuator force Fm and human force Fh are collocated.

Im

6

Re

Figure 8.17: The set of all possible LTI passive impedances occupies the
right half Nyquist plane.

operators that map vh to Fh . It is well-known that such an operator must

be positive real; i.e., in the Nyquist plane, the real part (representing energy
dissipation) must be non-negative. Thus, Σ can be represented by the half-
plane shown in Figure 8.17.
The task, therefore, is to prove the stability of the system illustrated in

! !

! !
 ! !

! !

8.6. Extending Z-Width 173

Figure 8.16. Doing so establishes the strict passivity of the haptic display.
In this section, we only outline the proof, which uses Nyquist theory. The
basic strategy is to write the closed loop characteristic equation as 1 +
A(s)∆(s) = 0, where ∆(s) is the uncertainty set consisting of the unit disk.
If the open loop (uncoupled) system is stable, then a sufficient condition
for closed loop (coupled) stability is:

1 + A(jω)∆(jω) (= 0 ∀ω, ∀∆ (8.36)

or, equivalently,

|A(jω)| < 1 ∀ω (8.37)

This is a version of the Small Gain Theorem [Desoer and Vidyasagar 75].
Straightforward manipulation shows that the sufficient condition for the
closed loop stability of Figure 8.16 is:

n=∞
1 − e−jωT $ 1
1−H(ejωT ) (= 0
T n=−∞
[Z(jω + jnωs ) + Σ(jω + jnωs )](ω + nωs )2
(8.38)
where ωs = 2π/T is the sample rate. Consider the sum Z + Σ. Because
Σ has an arbitrary imaginary part, the imaginary part of Z contributes
nothing further. The real part of Z, however, shifts Σ to either the right
or left, depending on sign. In the cases of interest, ${Z(jω)} > 0, which
shifts Σ to the right. Moreover, 1/(Z + Σ) is easily found to be a circular
disk centered on the real axis and tangent to the origin as well as the point
(1/${Z(jω)}, 0). If this disk were frequency independent, we could factor
it out of the infinite sum, but in general this is not the case. Here, we
will make the assumption that ${Z(jω)} is non-decreasing with frequency,
meaning that the amount of damping in the haptic display remains fixed
or grows with increasing frequency. With this assumption, it is apparent
that:

1 1
⊂ ∀n (= 0 (8.39)
Z(jω + jnωs ) + Σ(jω + jnωs ) ${Z(jω)} + Σ(jω)

And Equation 8.38 is satisfied whenever Equation 8.40 holds true:

n=∞
$
1 − e−jωT 1
1 + H(ejωT ) (= 0 (8.40)
T [${Z(jω)} + Σ(jω)] n=−∞ (ω + nωs )2

The infinite sum can be solved analytically, yielding:

! !

! !
 ! !

! !

174 8. Stability of Haptic Displays

T (1 − e−jωT )
1 + H(ejωT ) (= 0 (8.41)
2(1 − cos(ωT ))[${Z(jω)} + Σ(jω)]
Or, in terms of the unit disk, ∆:

T (e−jωT − 1) 1 + ∆
1 + H(ejωT ) (= 0 (8.42)
4(1 − cos(ωT )) ${Z(jω)}
For compactness, we define:

T (e−jωT − 1)
r(jω) = (8.43)
4(1 − cos(ωT ))
The assumption of uncoupled stability enables us to rewrite Equation 8.42
as:

r(jω)H(ejωT )
1+ ∆ (= 0 (8.44)
${Z(jω)} + r(jω)H(ejωT )
which is the form of Equation 8.36. Thus, stability requires:
* *
* r(jω)H(ejωT ) *
* *
* ${Z(jω)} + r(jω)H(ejωT ) * < 1 (8.45)

This can be manipulated into the following form:

T
${Z(jω)} + ${(e−jωT − 1)H(ejωT )} > 0 (8.46)
2(1 − cos(ωT ))
ωs
for 0 ≤ ω ≤ ωN = (8.47)
2
Equation 8.46 may be compared to the result of [Colgate and Schenkel
97], also presented here as Equation 8.2. In the event that ${Z(jω)} has a
fixed value of b, the results are the same. Equation 8.46 is therefore a more
general result than previously reported, but subject to the non-decreasing
assumption.

8.6.2 Insights into Passivity and Damping

The passivity criterion in Equation 8.46 is slightly more general than Equa-
tion 8.2 in that it allows for frequency-dependent physical damping, but
only under the assumption that the physical damping is a non-decreasing
function of frequency.
The criterion in Equation 8.46 lets us, in effect, sum together the phys-
ical damping (first term) and virtual damping (second term). At each

! !

! !
 ! !

! !

8.6. Extending Z-Width 175













     


Figure 8.18: Physical, virtual, and total damping levels for the system of
Figure 8.19a and the virtual wall of Equation 8.3 with m/b = 0.1, KT /b =
1, B/b = 0.5. Note the excess of total damping at low frequency required
to achieve positive damping at the Nyquist frequency.

       
 


 
(a) (b)

Figure 8.19: (a) Model of a haptic display having inertia m and viscous
damping b. (b) Addition of the mass m2 gives rise to “high pass” damping.

frequency from zero to the Nyquist frequency, the sum (total damping)
must be positive to ensure passivity.
Figure 8.18 shows, as an example, the physical, virtual, and total damp-
ing for the haptic display pictured in Figure 8.19a, and implements the
virtual wall of Equation 8.3. It is evident that, in order to ensure passivity
at the Nyquist frequency, a considerable excess of damping is required at
low frequencies.
The negative virtual damping at high frequency is caused principally by
the phase delay of the backwards difference differentiator used to compute

! !

! !
 ! !

! !

176 8. Stability of Haptic Displays

1
Physical Damping
Virtual Damping
Total Damping
damping levels 0.5 Total Damping w/o Filter

-0.5

-1
0 0.2 0.4 0.6 0.8 1
frequency (w/wN)

Figure 8.20: Physical, virtual, and total damping for the system of Fig-
ure 8.19a, with the same parameters as Figure 8.18 and the addition of a
first order low pass velocity filter having a cutoff frequency one-fifth of the
Nyquist frequency. Note that the improved total damping at high frequen-
cies is offset by reduced total damping at low frequencies. Nonetheless, it
is evident that the physical damping could be reduced or the virtual wall
impedance increased without loss of passivity.

velocity. This effect can be minimized by filtering. For instance, if we com-

bine a first order low-pass digital filter with the differentiator, and set the
cutoff frequency at one fifth the Nyquist frequency, we obtain Figure 8.20.
The high frequency negative damping has been reduced but at a cost. The
extra phase lag introduced by the filter causes negative virtual damping to
occur at lower frequencies. This is a good illustration of why high order
velocity filters are rarely used in haptics: the cost of added phase delay
often out-weighs the benefits of magnitude roll-off. To the best of the au-
thor’s knowledge, no theory of optimal filter design for haptics (other than
the work of [Janabi-Sharifi et al. 00] reviewed previously, which is aimed
at handling quantization), has been developed.
A second approach to improving Z-width is to replace the simple, fixed
damper of Figure 8.19a with a high frequency damper, such as the one in
Figure 8.19b. By connecting the distal end of the damper to a floating
inertia rather than to ground, the effective physical damping, (${Z(jω)})
approaches zero at low frequency. Figure 8.21 shows that the combination
of “high pass” damping and velocity filtering enables a significantly higher
impedance virtual wall to be implemented passively than for the naı̈ve

! !

! !
 ! !

! !

8.6. Extending Z-Width 177













     


Figure 8.21: Physical, virtual, and total damping for the system of Fig-
ure 8.19b, with the same low pass velocity filter as Figure 8.20, and a
higher impedance virtual wall. Parameters are m/b = 0.1, m2 /b = 0.01,
KT /b = 1.8, B/b = 0.9. Positive total damping is maintained at all fre-
quencies without significant excess at any frequency.

design of Figures 8.18 and 8.19a.

8.6.3 Mechanical Methods

The direct approach of adding a mechanical viscous damper to the haptic
interface to increase the maximum passive impedance of the system works
well, as demonstrated by [Colgate and Brown 94]. The maximum pas-
sive virtual stiffness and damping are limited by the physical dissipation
in the mechanism by Equation 8.4. The additional physical damping is
counteracted using digital control; the damper torque is measured and a
low-passed version of this torque is added to the motor command. This
masks the user’s perception of damping at the low frequencies of human
voluntary motion, but improves system stability and passivity at high fre-
quencies where discrete-time control is ineffectual and energy leaks are most
problematic.
There are some practical problems with typical physical dampers, such
as temperature dependence, fluid leakage, and Coulomb friction generated
in fluid seals. Figure 8.4 shows the increased impedance range when phys-
ical damping is added to the haptic display and when the sampling rate
is increased. Magnetic dampers using eddy currents also work with the

! !

! !
 ! !

! !

178 8. Stability of Haptic Displays

Figure 8.22: Design of a 1 DOF Haptic Display with motor (on the left)
and fluid filled viscous damper (on the right) connected via a removable
steel tape [Brown 95].

added benefit of being able to turn off the damping, when rendering low
impedances [Gosline et al. 06]. It is also possible to use mechanical brakes
to dissipate energy and mimic the behavior of a damper in order to provide
the necessary dissipation in the mechanism [An and Kwon 06], although
the slow dynamic response of magnetic brakes may limit their performance.

8.6.4 Electrical Methods

More recently, a variety of techniques emerged that take advantage of
analog components for rendering continuous time behavior. This method
strives to avoid the difficulties presented by mechanical dampers, but still
incorporates the dramatic performance improvements afforded.
One such method of electrically increasing the Z-width of a haptic dis-
play is to design an analog motor controller that locally monitors each joint
and controls the coupling stiffness and damping [Kawai and Yoshikawa 02]
in order to maintain passivity. A schematic of this is shown in Figure 8.23.
In this way, the joint stiffness and damping are continuously controlled,
while the virtual environment is updated and commands the joint coupling
parameters at the sampling intervals, as shown in Figure 8.24. The increase
in passivity and system Z-width using these analog impedance controllers
is shown in [Kawai and Yoshikawa 04].

! !

! !
 ! !

! !

8.6. Extending Z-Width 179

Figure 8.23: Outline of analog feedback control corresponding to two joints

[Kawai and Yoshikawa 04]. (! c 2004 IEEE)

Another class of controllers takes advantage of the motor’s natural dy-

namics. Since electric motors are gyrators, a damper on the mechanical side
of the motor acts as a resistor on the electrical side of the motor [Karnopp
et al. 00]. [Mehling et al. 05] used a resistor and capacitor in parallel
with the motor to add frequency dependent electrical damping to a haptic
display, as illustrated in Figure 8.25.
The amount of electrical damping added is a function of the motor
torque constant Kt , the motor winding resistance Rm , and the external
resistor R1 :

Kt2
beq = (8.48)
R1 + Rm
This technique can be quite effective as illustrated in Figure 8.26. [Mehling
et al. 05] used an R-C cutoff frequency of 2.6 Hz, providing significant
damping at higher frequencies where the haptic display is likely to be un-
stable or exhibit limit cycle oscillations and above the frequencies of human
voluntary motion. It is important to note that the capacitor acts as ap-
parent inertia on the mechanical side of the motor. For this reason, the
R-C time constant of the electrical damper must be selected carefully. The
resistance must be small enough to provide useful damping at the frequen-
cies of interest, while keeping the capacitance small enough to cause only
a modest impact on apparent inertia at low frequency.
Clearly, maximum physical damping, b, is provided when R1 goes to

! !

! !
 ! !

! !

180 8. Stability of Haptic Displays

. Conventional haptic device.

. Proposed haptic device.

Figure 8.24: Conventional and electrically coupled hybrid haptic device

[Kawai and Yoshikawa 04]. (!
c 2004 IEEE)

zero, i.e. the motor is “crowbarred.” However, this creates problems with
driving the motor; any voltage applied bypasses the motor. The winding
resistance, Rm , also sets an upper bound on the electrical damping that
can be achieved in this configuration. There is also a practical limit to how
large the capacitance can be in addition to the added apparent inertia at
low impedance.
To increase electrical damping beyond the limit of Equation 8.48, it
is possible to design a circuit to cancel the effect of the motor winding
resistance, Rm [Diolaiti and Niemeyer 06]. Such a circuit in the motor
amplifier allows the motor winding resistance to be reduced dramatically;
however, due to noise and thermal effects, Rm cannot be canceled com-
pletely. Due to gyration, the motor winding inductance acts like a spring
on the mechanical side. [Diolaiti and Niemeyer 06] take advantage of this
by combining wave variable control with a circuit to cancel Rm , leaving

! !

! !
 ! !

! !

8.6. Extending Z-Width 181

Damper
B
Motor Motor
Kt, Rm Kt, Rm

Motor + - + -

+ - R1 R1 C
(a) (b) (c)

Figure 8.25: A mechanically damped system (a) and two electrically

damped systems; one without (b) and one with frequency depen-
dence (c) [Mehling et al. 05]. (!
c 2005 IEEE)

the springlike inductance to couple the physical world with the virtual en-
vironment. The benefit is that for common DC motors, such as those in
the Phantom haptic display, the resulting effective spring constant of the
inductance is much higher than the maximum passive stiffness that can be
attained using feedback and digital control. This technique also requires
recasting the digital controller in the form of wave variables, as shown in
Figure 8.27.
Extending Diolaiti and Niemeyer’s work, it is possible to use analog
circuitry to estimate the back EMF (electromotive force or voltage) of the
winding, by canceling both the resistance and the inductance of the motor
windings. The back EMF of the motor is proportional to velocity, so feeding
this signal back to the motor inside the current control amplifier provides
electrical damping. One caveat is that prior knowledge of the parameters
and dynamics of the motor is required in order to design such circuitry, and
dynamic tuning of the parameters is necessary to compensate for heating
in the windings.

8.6.5 Psychophysical Methods

In addition to analytical and quantitative methods for increasing the max-
imum passive stiffness that can be rendered by a haptic display, there are a
variety of psychophysical techniques available to improve human perception
of stiff virtual surfaces.
[Salcudean and Vlaar 97] developed a rendering method for virtual

! !

! !
 ! !

! !

182 8. Stability of Haptic Displays

Average Stability Boundaries for Tested Levels of Electrical Damping

300
(1) No Electrical Dam ping
(2) Electrical Dam ping = 0.00755 Nm s/rad
(3) Electrical Dam ping = 0.0151 Nms /rad

250
Virtual Stiffness (Nm/rad)

200

(3)
150

100 (2)

50

(1)

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

Virtual Damping (Nms/rad)

Figure 8.26: Z-width plot of the average stability boundary for each level
of electrical damping. Dashed lines indicate plus or minus one standard
deviation [Mehling et al. 05]. (!
c 2005 IEEE)

+
u √ eA eW
2R _
KL
Analog F F FH
R kT ẋ m
Circuit W ẋ
c
+ √
v _ 2R
i
Wave Transform

Figure 8.27: The wave transform connecting the virtual environment to

the electrical domain is implemented with an analog circuit [Diolaiti and
Niemeyer 06]. (! c 2006 IEEE)

walls using a “braking pulse” that occurs upon contact with the wall bound-
ary. The force of the pulse is designed to bring the haptic display to rest
as quickly as possible, ideally in one sampling period. This corresponds to
a very high level of damping when crossing the wall boundary, but since
the high level of damping is not sustained, it does not lead to instabil-
ity that would occur with a similar level of virtual damping in a constant

! !

! !
 ! !

! !

8.6. Extending Z-Width 183

vin

Position
Proportional

Transient
Force

Time

Figure 8.28: Schematic of position and force in event-based haptic display

[Kuchenbecker et al. 06]. (!
c 2006 IEEE)

parameter virtual wall. After the braking pulse, as the user remains in
contact with the virtual wall, the rendering method consists of the stan-
dard spring-damper virtual wall with virtual stiffness and damping gains
set, such that they are stable. This results in behavior that is similar to an
object colliding with a real wall and increases the perceived wall stiffness.
[Lawrence et al. 00] introduced the concept of “rate-hardness” as a way
of quantifying human perception of our virtual surfaces. Rate-hardness is
the ratio of initial rate of change of force versus initial velocity upon pene-
trating the surface. Human perception studies indicate that rate-hardness
is a more relevant perceptual hardness metric then absolute mechanical
stiffness when rendering virtual surfaces. This is likely due to the rela-
tively poor performance of the human kinesthetic sense when in contact
with stiff walls. When a human is already in contact with a stiff virtual
wall, the change in position relative to the change in force when haptically
querying the wall is very small.
If the user is allowed to dynamically test the wall through tapping, for
example, human perception is much better at distinguishing varying surface
hardness. It seems that tapping elicits high frequency force differences
which can be perceived by the pressure and vibration sensory receptors in
the fingers. Artificially increasing the rate hardness can act as a haptic
illusion, making the surface seem harder than the stiffness alone would
predict [Lawrence et al. 00].

! !

! !
 ! !

! !

184 8. Stability of Haptic Displays

[Okamura et al. 98] introduced a technique to improve the perception

of contact with virtual objects. High frequency open-loop force transients
corresponding to interaction events in the virtual environment are super-
imposed on a standard virtual wall controller, as indicated in Figure 8.28.
To determine the open loop vibrations to display, high resolution vibration
and position information was gathered while tapping on a variety of ma-
terials. The data was fit to the amplitude A(v), decay constant B, and
frequency ω of a decaying sinusoidal signal Q(t), resulting in each material
having a different vibration signature.

Q(t) = A(v)e−Bt sin(ωt) (8.49)

Typical haptic displays generally do not accurately reproduce high fre-
quency vibration signals. To compensate for these device dynamics, [Oka-
mura et al. 01] improved the vibration models by performing a set of
human perceptual experiments to tune the parameters of the vibration sig-
natures. Based on the results of the experiments, the adjusted parameters
result in more realistic perception of tapping on the three test materials:
rubber, wood, and aluminum. One drawback to these techniques is that
each material type, geometry, and haptic display needs to be individually
characterized prior to use to determine appropriate vibration signatures.
Extending the work on reality-based vibration feedback, [Kuchenbecker
et al. 06] utilized an acceleration matching technique based on the experi-
ence of contacting the real object being rendered in the virtual environment.
To improve the accuracy of the force transients displayed upon contact with
a virtual object, the open-loop acceleration signal is pre-warped by an in-
verted system model to correct for the distortion and dynamics induced by
the haptic display. Chapter 22 treats in more detail measurement-based
haptic rendering.

8.7 Summary
In summary, haptic instability frequently arises from a lack of passivity
when rendering virtual environments. In order to maintain passivity, vir-
tual environment impedance can be reduced to acceptable levels for passiv-
ity, but this depends upon the specific hardware used, and highly complex
virtual environments make this undesirable. To preserve the universal-
ity and accuracy of virtual environments, virtual couplings can be used
to modulate the impedance transmitted between the haptic display and
the virtual environment to ensure passivity. Passivity controllers can in-
crease the nominal impedance of haptic display by counteracting energy
leaks introduced by the sampled-data system. Direct methods of design-

! !

! !
 ! !

! !

BIBLIOGRAPHY 185

ing for passivity work to increase the maximum passive impedance of the
haptic interface, improving performance. Lastly, perceptual methods of
improving performance take advantage of the limits of human perception
to create the illusion of higher performance rendering on existing haptic
display hardware.

Acknowledgements This work was supported in part by Northwestern

University, National Science Foundation IGERT Fellowships through grant
DGE-9987577, and National Science Foundation Grant No. 0413204.

Bibliography
[Abbott and Okamura 05] Jake J. Abbott and Allison M. Okamura. “Ef-
fects of Position Quantization and Sampling Rate on Virtual-Wall Pas-
sivity.” IEEE Transactions on Robotics 21 (2005), 952–964.

[Adams and Hannaford 98] Richard J. Adams and Blake Hannaford. “A

Two-Port Framework for the Design of Unconditionally Stable Hap-
tic Interfaces.” In IEEE/RSJ International Conference on Intelligent
Robots and Systems, 2, 2, pp. 1254–1259. Victoria, BC, Canada: IEEE,
1998.

[Adams and Hannaford 99] Richard J. Adams and Blake Hannaford. “Sta-
ble Haptic Interaction with Virtual Environments.” IEEE Transac-
tions on Robotics and Automation 15 (1999), 465–474.

[Adams et al. 98] Richard J. Adams, Manuel R. Moreyra, and Blake Han-
naford. “Stability and Performance of Haptic Displays: Theory and
Experiments.” In Proceedings ASME International Mechanical Engi-
neering Congress and Exhibition, pp. 227–234. Anaheim, CA: ASME,
1998.

[An and Kwon 06] Jinung An and Dong-Soo Kwon. “Stability and Per-
formance of Haptic Interfaces with Active/Passive Actuators-Theory
and Experiments.” International Journal of Robotics Research 25:11
(2006), 1121–1136.

[Bolanowski et al. 88] S. J. Bolanowski, G. A. Gescheider, R. T. Verrillo,

and C. M. Checkosky. “Four channels mediate the mechanical aspects
of touch.” Journal of the Acoustical Society of America 84:5 (1988),
1680–1694.

! !

! !
 ! !

! !

186 BIBLIOGRAPHY

[Brown and Colgate 98] J. Michael Brown and J. Edward Colgate. “Mini-
mum Mass for Haptic Display Simulations.” In Proceedings ASME In-
ternational Mechanical Engineering Congress and Exhibition, 61, 61,
pp. 85–92. Anaheim, CA: ASME, 1998.

[Brown 95] J. Michael Brown. “A Theoretical and Experimental Investi-

gation into the Factors Affecting the Z-Width of a Haptic Display.”
Master’s thesis, Northwestern University, Evanston, IL, 1995.

[Colgate and Brown 94] J. Edward Colgate and J. Michael Brown. “Fac-
tors Affecting the Z-Width of a Haptic Display.” In IEEE Interna-
tional Conference on Robotics and Automation, 4, 4, pp. 3205–3210.
San Deigo, CA: IEEE, 1994.

[Colgate and Hogan 88] J. Edward Colgate and Neville Hogan. “Robust
Control of Dynamically Interacting Systems.” International Journal
of Control 48:1 (1988), 65–88.

[Colgate and Schenkel 97] J. Edward Colgate and Gerd G. Schenkel. “Pas-
sivity of a Class of Sampled-Data Systems: Application to Haptic In-
terfaces.” Journal of Robotic Systems 14:1 (1997), 37–47.

[Colgate et al. 93] J. Edward Colgate, Michael C. Stanley, and Gerd G.

Schenkel. “Dynamic range of achievable impedances in force reflecting
interfaces.” In Telemanipulator Technology and Space Telerobotics,
2057, 2057, pp. 199–210. Baltimore, MD: SPIE, 1993.

[Colgate et al. 95] J. Edward Colgate, Michael C. Stanley, and J. Michael

Brown. “Issues in the Haptic Display of Tool Use.” In IEEE/RSJ
International Conference on Intelligent Robots and Systems, 3, 3,
pp. 140–145. Pittsburgh, PA: IEEE, 1995.

[Desoer and Vidyasagar 75] C.A. Desoer and M. Vidyasagar. Feedback

Systems: Input-Output Properties. New York, USA: Academic Press,
1975.

[Diolaiti and Niemeyer 06] Nicola Diolaiti and Günter Niemeyer. “Wave
Haptics: Providing Stiff Coupling to Virtual Environments.” In IEEE
Symposium on Haptic Interfaces, 4, 4, pp. 185–192. Alexandria, VA:
IEEE, 2006.

[Diolaiti et al. 06] Nicola Diolaiti, Günter Niemeyer, Federico Barbagli,

and J. Kenneth Salisbury. “Stability of Haptic Rendering: Discretiza-
tion, Quantization, Time-Delay and Coulomb Effects.” IEEE Trans-
actions on Robotics 22 (2006), 256–268.

! !

! !
 ! !

! !

BIBLIOGRAPHY 187

[Gillespie and Cutkosky 96] R. Brent Gillespie and Mark R. Cutkosky.

“Stable User-Specific Haptic Rendering of the Virtual Wall.” In
ASME International Mechanical Engineering Conference and Expo-
sition, DSC, 58, 58, pp. 397–406. Atlanta, GA: ASME, 1996.

[Gosline et al. 06] Andrew H. Gosline, Gianni Campion, and Vincent Hay-
ward. “On the Use of Eddy Current Brakes as Tunable, Fast Turn-On
Viscous Dampers for Haptic Rendering.” In Eurohaptics, pp. 229–234.
Paris, France, 2006.

[Hannaford et al. 01] Blake Hannaford, Jee-Hwan Ryu, and Yoon S. Kim.
“3 - Stable Control of Haptics.” In Touch in Virtual Environments:
Proceedings USC Workshop on Haptic Interfaces, edited by Margret
McLaughlin. Prentice Hall, 2001.

[Janabi-Sharifi et al. 00] Farrokh Janabi-Sharifi, Vincent Hayward, and

Chung-Shin J. Chen. “Discrete-Time Adaptive Windowing for Veloc-
ity Estimation.” IEEE Transactions on Control Systems Technology
8:6 (2000), 1003–1009.

[Karnopp et al. 00] Dean C. Karnopp, Donald L. Margolis, and Ronald C.

Rosenberg. System Dynamics: Modeling and Simulation of Mecha-
tronic Systems, Third edition. New York, USA: A Wiley-Interscience
Publication, 2000.

[Kawai and Yoshikawa 02] Masayuki Kawai and Tsuneo Yoshikawa. “Hap-
tic Display of Movable Virtual Object with Interface Device Capable
of Continuous-Time Impedance Display by Analog Circuit.” In IEEE
International Conference on Robotics and Automation, pp. 229–234.
Washington, DC: IEEE, 2002.

[Kawai and Yoshikawa 04] Masayuki Kawai and Tsuneo Yoshikawa. “Hap-
tic Display with an Interface Device Capable of Continuous-Time
Impedance Display Within a Sampling Period.” IEEE/ASME Trans-
actions on Mechatronics 9:1 (2004), 58–64.

[Kuchenbecker et al. 06] Katharine J. Kuchenbecker, Jonathan Fiene, and

Günter Niemeyer. “Improving contact realism through event-based
haptic feedback.” IEEE Transactions on Visualization and Computer
Graphics 12 (2006), 219–230.

[Lawrence et al. 00] Dale A. Lawrence, Lucy Y. Pao, Anne M. Dougherty,

Mark A. Salada, and Yiannis Pavlou. “Rate-hardness: a new perfor-
mance metric for haptic interfaces.” IEEE Transactions on Robotics
and Automation 16 (2000), 357–371.

! !

! !
 ! !

! !

188 BIBLIOGRAPHY

[Mehling et al. 05] Joshua S. Mehling, J. Edward Colgate, and Michael A.

Peshkin. “Increasing the Impedance Range of a Haptic Display by
Adding Electrical Damping.” In IEEE First World Haptics Conference
and Symposium, pp. 257–262. Pisa, Italy: IEEE, 2005.

[Miller et al. 00] Brian E. Miller, J. Edward Colgate, and Randy A. Free-
man. “Guaranteed Stability of Haptic Systems with Nonlinear Virtual
Environments.” IEEE Transactions on Robotics and Automation 16:6
(2000), 712–719.

[Miller et al. 04] Brian E. Miller, J. Edward Colgate, and Randy A. Free-
man. “On the role of dissipation in haptic systems.” IEEE Transac-
tions on Robotics 20 (2004), 768–771.

[Okamura et al. 98] Allison M. Okamura, Jack T. Dennerlein, and

Robert D. Howe. “Vibration Feedback Models for Virtual Environ-
ments.” In IEEE International Conference on Robotics and Automa-
tion, 3, 3, pp. 2485–2490. Leuven, Belgium: IEEE, 1998.

[Okamura et al. 01] Allison M. Okamura, Mark R. Cutkosky, and Jack T.

Dennerlein. “Improving Reality-Based Models for Vibration Feed-
back.” IEEE/ASME Transactions on Mechatronics 6:3 (2001), 245–
252.

[Ryu et al. 04] Jee-Hwan Ryu, Yoon S. Kim, and Blake Hannaford.
“Sampled- and continuous-time passivity and stability of virtual envi-
ronments.” IEEE Transactions on Robotics 20:4 (2004), 772–776.

[Ryu et al. 05] Jee-Hwan Ryu, Carsten Preusche, Blake Hannaford, and
Gerd Hirzinger. “Time-domain Passivity Control with Reference En-
ergy Following.” IEEE Transactions on Control Systems Technology
13:5 (2005), 737–742.

[Salcudean and Vlaar 97] S. E. Salcudean and T. D. Vlaar. “On the Emu-
lation of Stiff Walls and Static Friction with a Magnetically Levitated
Input-Output Device.” Transactions of the ASME: Journal of Dy-
namics, Measurement and Control 119:1 (1997), 127–132.

[Stramigioli et al. 02] Stefano Stramigioli, Cristian Secchi, Arjan J. van der
Schaft, and Cesare Fantuzzi. “A novel theory for sample data sys-
tems passivity.” In IEEE/RSJ International Conference on Intelligent
Robots and Systems, pp. 1936–1941. Lausanne, Switzerland: IEEE,
2002.

[Zilles and Salisbury 95] C. Zilles and K. Salisbury. “A Constraint-Based

God Object Method for Haptic Display.” In IEEE/RSJ International

! !

! !
 ! !

! !

BIBLIOGRAPHY 189

Conference on Intelligent Robots and Systems, 3, 3, pp. 146–151. Pitts-

burgh, PA: IEEE, 1995.

! !

! !