Command Generation for

Dynamic Systems
William Singhose 1
Warren Seering 2
Copyright 2011

January 17, 2011

Woodruff School of Mechanical Engineering, Georgia Institute of

Technology
Department of Mechanical Engineering, Massachusetts Institute of
Technology

Preface
Many mechanical systems achieve their design objectives by performing movement from an initial configuration or location to a desired one. Planning these movements involves two tasks: 1) specification of a trajectory that the system is commanded to follow, and 2)
design of a control algorithm to ensure that the system follows the desired path. If the system is subjected to significant disturbances, then
the controller will need feedback to compensate for the disturbances.
A vast literature exists on the design of feedback algorithms. Far less
has been written about methods for specifying trajectories. The purpose of this book is to provide the reader with methods for generating
commands that will cause machines to complete desired moves
quickly and accurately.
One might assume that the best trajectory for moving a system from
an initial configuration to any subsequent one would specify that the
system move at its top speed directly from the former to the latter. For
many systems this is not the best approach. Typically, the commanded
position or velocity will be specified as a function of time and will depend on the motion to be achieved, the capacity of the system actuators, and the dynamic response of the system in its environment.
A simple example showing the need to plan a trajectory occurs when
a golfer putts on an undulating green. Expecting the ball to travel in a
straight line to the cup would be naive; thats not what golf balls do.
Good golfers plan both the position and the velocity1 of the ball for its
entire trajectory. The plan takes into account the distance to the hole,
the changing slope of the green, the condition of the grass, and perhaps a strong crosswind. To ensure best performance, the desired trajectories of machines should be determined with the same care.

4/237

It has been proposed that exact specification of the command is not

particularly important so long as the feedback controller is capable of
guiding the system to its desired state. We disagree. For a given system, feedback controller, and objective, there will be one or a family of
trajectories that will cause the objective to be achieved most effectively. It is the design of these trajectories that the methods in this book
will address. And as will be seen, the differences in performance resulting from employment of different commands, even for systems with
feedback control, is often quite substantial, particularly for systems
with complex dynamic responses.
We assume that the readers of this book will have a basic understanding of dynamic system response equivalent to that which would
be achieved through completion of a freshman class in Newtonian
physics. Familiarity with Laplace transforms will be helpful for those
who wish to fully understand Chapters 6 and beyond. Those who read
and understand the material in Chapters 1 through 11 will be prepared
to design command trajectories for a wide range of mechanical
systems.
1

They are coupled, of course.

Chapter 1
Basic Control System
Elements
Systems with oscillatory dynamics present a challenging class of
control problems. Imagine trying to transfer cargo between ships at
sea using a large crane when waves and wind are causing large disturbances to the ships and dangling payloads. Or, think about trying to
reorient a satellite equipped with huge and very flexible solar panels
that behave like aluminum foil. To control such systems accurately, a
sophisticated and robust control system is indispensable. In the case
of the crane, the current industry standard is a well-trained human operator whose skills have been perfected over time. In the case of satellites, a computer makes control decisions based on various on-board
sensor measurements and information provided by ground-based mission controllers. Then, reaction jets and momentum wheels provide
the appropriate control forces and torques to move the system without
causing damage or inducing large amounts of oscillation.
At the heart of any control problem is a physical system that is acted
upon by control effort. This input induces some type of response, as
shown schematically in Figure 1.1. The engineers that design and work
with the physical plant will have some performance requirements for
the system response. For example, the system may need to move very
quickly from one location to another and then settle to within a few
millimeters of the desired location. Or, the machine may have to pick
up a wide range of payloads and move them through complicated
trajectories.

6/237

The performance requirements drive the physical hardware design,

but they also help guide the design of the command generator that
produces the control effort. Every dynamic system has a component
that generates the control effort, as shown schematically in Figure 1.2.
The command generator receives the desired performance specifications and converts them into control effort that induces an appropriate
response. For the majority of machines, the command generator is a
human operator. Using the performance requirements, the human operator decides when the machine needs to move, where it needs to go,
and how fast it should travel. Modern control systems strive to improve upon the performance of human operators by either enhancing
the human operator control efforts, or replacing them entirely with a
computer controller.

Figure 1.1: Fundamental Plant Dynamics.

Figure 1.2: Command Generator and Plant Dynamics.

7/237

Figure 1.3: Bridge Crane.

As an illustration of this fundamental dynamic process, consider an
overhead bridge crane, like the one sketched in Figure 1.3. The payload is hoisted up and down by a suspension cable. The upper end of
the cable is attached to a trolley that travels along a bridge. Furthermore, the bridge on which the trolley travels can also move perpendicular to the trolley motion, thereby enabling three-dimensional positioning of the payload. By pressing the control buttons, the operator
indicates to the electronic motor controllers that the motors should be
turned on to move the crane in the desired direction. If a control button is pressed for a finite time period, then the trolley will move a certain distance and come to rest. The payload, on the other hand, will
usually oscillate about the new position. The planar motion of the payload for a short trolley movement is plotted in Figure 1.4.
Oscillations can cause significant problems when the payload is being transported through a cluttered work environment containing
obstacles and human workers. Furthermore, payload sway complicates final positioning. Figure 1.5 shows an overhead view of a crane
payload position while the trolley is being driven through an obstacle

8/237

field by a novice operator. There is considerable payload sway both

during the transport and at the goal location. In fact, the payload collided with two of the obstacles. This data was obtained via an overhead
camera that tracked the payload motion of a bridge crane at Georgia
Tech [108]. Obviously, crane payloads are difficult to position accurately and require a skilled human operator to generate effective
commands.

Figure 1.4: Bridge Crane Response to a Single Button Press.

9/237

Figure 1.5: Crane Payload Response Through an Obstacle

Field.
In addition to the large payload oscillations shown in Figure 1.5, the
crane structure can experience vibrations at frequencies much higher
than the pendulum mode of the payload. If the motors accelerate the
crane rapidly to full velocity, then these higher modes of vibration may
be excited. For this reason, the motor drives do not force the motors to
instantly change velocity when the control buttons are pressed. Rather, a full-velocity command issued by the human operator is filtered
by the motor drive and made into a smooth function that transitions
from zero velocity to full velocity over a finite period. Such a smoothed
velocity profile from the Georgia Tech crane is shown in Figure 1.6.

Figure 1.6: Velocity Command Smoothed by Motor Drives.

Figure 1.7: Command Generator, Feedforward, and Plant

Dynamics.

10/237

The nature of the command smoothing operation is designed not

only to suppress the higher vibration modes, but also to decrease
thermal loading on the amplifiers. In this sense, the command
smoothing is based on a model of the system vibration dynamics and
thermal properties. This smoothing operation can be labeled as feedforward control and its function can be allocated to an additional control block, as shown in Figure 1.7. The division of the control process
into command generation and feedforward control in Figure 1.7 is
somewhat artificial because they could simply be combined together
and called the command generator. However, in this case the functional decomposition is reasonable because the human operator is
clearly the command generator and the motor drives are a feedforward controller that utilizes a model of the system to produce the actual control effort.
The distinction between command generation and feedforward control actions becomes more meaningful when the system also has feedback control. Typically, modern control systems can be represented by
the block diagram shown in Figure 1.8. Four distinct elements contribute to the overall system capability: the command generator, feedforward control, feedback control, and the physical plant. As the diagram
in Figure 1.8 shows, the command generator block receives the desired
performance specifications. It then generates reference commands for
both the feedforward and feedback control blocks. These two elements
send control effort directly into the plant. The plant then responds to
this combination of control effort.
Design of each of the four system elements provides unique opportunities for improving the overall system performance. For example,
the feedback control can be designed to reject random disturbances,
while the feedforward block cannot. However, the feedforward block
can compensate for unwanted dynamic effects before they show up in
the output response, while the feedback controller cannot.

11/237

Figure 1.8: General Purpose Control System.

Given the complexity of modern controllers, it is important to define
the various aspects of the control system precisely. For clarity, the following nomenclature will be used:
The Command Generator block transforms the desired system performance objectives into reference command signals designed to direct the system toward the specified objectives.
The Feedback Control block compares the reference command signals
from the command generation block with measurements of the system
states and determines a corrective control effort (force, torque,
voltage, etc.), if needed, to act on the plant.
The Feedforward Control block employs an internal model of the system to transform the reference commands from the command generator block into a forcing function that acts directly on the plant, in addition to the forcing function from the feedback control block.1
Perhaps the least studied and most under appreciated element of
control systems is the command generator. Ironically, it is the most
prevalent control element. Many machines do not have feedforward or
feedback controllers, but every machine has a command generator.
Human operators all over the world who guide machines through their
required tasks each day are the most widespread command generators. However, computers are necessary to generate commands for
complicated systems performing rapid and accurate movements.

12/237

Command generation is of fundamental importance. The shape of

the command has a large influence on system performance. Command
design should always be considered along with the other control system elements. This book concentrates on the command generator element of the control system. Dynamic systems with flexibility will form
the core applications because their performance is strongly dependent
on the choice of command.

Figure 1.9: Long Reach Robot Arm.

1.1 Example Systems and Problems

When performance is pushed to the limit in terms of motion, velocity, and throughput, the problem of flexibility usually arises. Flexibility comes from physical deformation of the structure and/or compliance introduced by the feedback control system. For example, implementing proportional and derivative (PD) feedback control is

13/237

analogous to adding a spring and damper to the system. The controller

spring can lead to problematic flexibility in the system. Physical deformation of the structure can occur in the links, cables, and joints.
This leads to problems with positioning accuracy, trajectory following,
settling time, component wear, and stability. It may also induce complicated nonlinear effects.
A system whose flexibility is obviously detrimental to its performance is the long-reach manipulator sketched in Figure 1.9. This robotic
arm was built to test methods for cleaning nuclear waste storage tanks
[53]. The arm needs to enter a tank through an access hole and then
reach long distances to clean the tank walls. These types of robots have
mechanical flexibility in the links, the joints, and possibly the base to
which they are attached. Both the gross motion of the arm and the
cleaning motion of the end effector can induce vibration. This vibration makes it difficult to clean the tanks properly.
The long links that compose the robot are significant sources of its
flexibility. However, all real systems will deflect if they are moved rapidly enough. Consider the moving-bridge coordinate measuring machine (CMM) sketched in Figure 1.10. The machine is composed of
stiff components including a granite base and large-cross-section
structural members. The goal of the CMM is to move a probe
throughout its workspace so that it can contact the surfaces of manufactured parts that are being inspected. In this way, it can accurately
determine the dimensions of the part and ensure that they meet specifications. The position of the probe is measured by optical encoders
that are attached to the granite base and the moving bridge. However,
if a laser interferometer is used to measure the probe motion accurately, then its actual location will differ from that indicated by the encoders. This difference arises because the physical structure deflects
between the encoders and the probe endpoint. Figure 1.11 shows the
micronlevel deflection for a typical move when the CMM approaches a
part rapidly and then slows down just before making contact with the
part surface [116]. If the machine is attempting to measure with

14/237

micron-level accuracy, then the 20- 25 micron amplitude vibration

during the slow approach phase is a significant impediment.

Figure 1.10: Moving-Bridge Coordinate Measuring

Machine.
In addition to mechanical deflection, flexibility introduced by the
control system is also commonplace. Feedback control works by detecting a difference between the desired response and the actual response. This difference is then used to generate a restoring force. In
many cases this force acts like a spring. Increasing the gains may have
the effect of stiffening the system, but it also increases actuator demand, noise problems, and instability problems.

15/237

Figure 1.11: Deflection of Coordinate Measuring Machine.

1.2 Example Command-Shaping

Solution
Creating special reference commands to move flexible systems
without vibration is an old idea [23, 120, 121, 128]. Commands can be
created such that the systems motion will cancel its own vibration.
Some of these techniques require the commands to be pre-computed
using boundary conditions before the move is initiated. Others can be
implemented with command functions being created in real time.
Another significant difference between the various control methods is
robustness to modeling errors. Some techniques require a very good
system model to work effectively, while others only need rough estimates of the system parameters.
An experienced crane operator can often produce the desired payload motion without much vibration by pressing the control buttons
multiple times at the proper time instances. A simple case consisting
of two button pressings is shown in Figure 1.12. In this case, the

16/237

payload is transported to the desired location and settles without residual oscillation. Comparing Figures 1.4 and 1.12, reveals the benefits
of properly generating the reference command. This is the type of effect that the command generator block in Figure 1.8 strives to achieve.
A command generator element was installed on the Georgia Tech
crane in order to help the human operators. Figure 1.13 shows the payload response with this command generator operating, (labeled
Input-Shaped Response). When this command generator aided the
novice operator by properly shaping the commands, the operator was
able to move the payload cleanly through the obstacles and arrive at
the goal without oscillation.
The first suggested use of this technology on cranes was made in the
late 1950s by O.J.M. Smith when he developed Posicast control [120,
121]. However, given the absence of digital controllers at that time, the
method does not appear to have been implemented on a real cranelike structure until Starr reported some excellent results manipulating
suspended payloads in 1985 [124]. Since the inception of this
command-shaping procedure, its lack of robustness to modeling errors and plant variations has been well known [128]. These effects occur in cranes when the payload is hoisted, the payload geometry
changes, the motors exhibit nonlinearity, or the suspension length is
not known with certainty. To address the problem, command-shaping
technology for cranes [20, 33, 43, 48, 78, 92, 110] (and all other machines) has been improved by the development of robust input-shaping methods, which will be described starting in Chapter 7.

17/237

Figure 1.12: Bridge Crane Response to a Double Button

Press.

Figure 1.13: Crane Payload Response Through an Obstacle

Field.

18/237

Recall that in the absence of a feedback control loop, the command

generator and feedforward control blocks can be combined into a
single block.

Chapter 2
Elementary Command
Generation
The introductory chapter demonstrated that the design of the reference command given to a dynamic system can have a significant impact on its performance. The question then arises as to how to generate the optimal command, or at least a very good reference command.
This question cannot be answered until a set of performance specifications is established. These specifications define what the system must
do and how well the tasks must be done. Another requirement for
command generation is a system model. Without knowing how a system responds to commands, there is no way of intelligently generating
an appropriate command function. In general, the more complicated
the system model, the more difficult it is to determine appropriate
commands. Therefore, models used for command generation should
always be as simple as possible and only represent the important dynamic effects. This chapter will examine command generation for very
simple systems. The techniques presented will be limited to elementary command generation methods that only consider system inertia,
viscous friction, acceleration limits, and velocity limits.

2.1 Commands for a Translational

Mass
One of the simplest systems to analyze is a mass that moves freely
only in one direction, as shown in Figure 2.1. To generate force commands, f(t), for the system, we must know what task the system is to

20/237

perform and we must have a dynamic model of how it will respond to

commands. An important part of the model is the equation of motion:

This gives a mathematical relationship between the command and

the system response. If the desired response, xd(t), is known, then it
can be substituted into (2.1) to calculate the necessary input, f(t), that
will make the actual response, x(t), follow the desired response. Note
that the actual response will never perfectly track the desired response
because there is always some uncertainty in the value of the mass.
Furthermore, the actuator used to create the input force cannot create
the required forcing function perfectly at each point in time.

Figure 2.1: Translational Mass.

Figure 2.2: Bang-Bang Function.

21/237

Suppose that we want the mass to move from one rest location to
another rest location. Furthermore, we want the mass to move quickly.
The faster we move the mass, the larger the force we must provide.
Without a specified limit on the maximum possible force, the solution
is undefined. But, it would take the form of a massive impulse of force
to start the motion, followed immediately by an equally massive negative impulse to bring the mass back to rest. To get a realistic solution,
our dynamic model must contain an actuator limit. Let us assume that
the force is bounded by some value umax:

The combination of (2.1) and (2.2) now gives a more reasonable

model of the system. Given the restriction of (2.2), the fastest way to
move the mass is to apply full positive force to accelerate the mass and
then switch to full negative force to slow the system back to rest. This
type of command is called a bangbang (BB) function and is shown in
Figure 2.2. The duration of the negative pulse must equal the duration
of the positive pulse in order for the system to return to rest.
Given that the amplitude of the bang-bang function is known and it
is antisymmetrical, the problem reduces to finding, t2, the time at
which the command switches from positive to negative. Rewriting
(2.1) to isolate the system response, we get:

Integrating twice to obtain the position during the time period of the
command yields:

22/237

Figure 2.3: Rigid-Body Response to a Bang-Bang Input.

If we realize that the mass will reach half the desired move distance,
xd, when the command is half finished (at t2), then we can simply set
the double integral in (2.4) equal to 0.5xd and perform the integration
from t1 to t2. Without loss of generality we can set the start of the command, t1, equal to zero. Over this range, f(t) is simply a constant value
of umax, so we get:

Solving for the switch time of the command gives:

23/237

where is the force-to-mass ratio (acceleration):

Therefore, the duration of the bang-bang command is:

Figure 2.3 shows the velocity and position response of the mass
when subjected to the bang-bang input. Note that even though the
control effort is a bang-bang function that contains discontinuous
jumps, the position response of the system is a smooth parabolic function, called an S-Curve. The curve transitions smoothly from start to
finish and is characterized by zero slope at the beginning and end. Figure 2.4 shows the switch times of the bang-bang command as a function of the desired move distance. The data shown in Figure 2.4 is for
the case when m = 1 and umax = 1. As expected, the command duration, t3, increases with the desired move distance.

24/237

Figure 2.4: Switch Times of Bang-Bang Command.

Figure 2.5: Bang-Coast-Bang Command.

Just as all real systems have a force limitation (acceleration limit),
most real systems also have a velocity limit. When subjected to a bangbang acceleration input, the peak velocity of the mass increases with
the desired move distance. Eventually, for a long enough move distance, the maximum velocity will be reached. After this point, the forcing function must change to a bangcoast- bang (BCB) command, like
that shown in Figure 2.5. In this case, the command is described by
four switch times, instead of the three that describe a bang-bang function. The velocity profile that corresponds to a bang-coast-bang acceleration function has a trapezoidal shape and the S-Curve position
function has a linear segment in the middle.
Note that setting a limit on the maximum velocity, vmax, results in a
longer move duration than when the velocity is unconstrained. The
duration of the acceleration pulse, t2, is also limited by the maximum
velocity. When the maximum velocity is reached, the acceleration input must switch to zero. To determine the pulse duration, we can

25/237

simply integrate (2.1) with respect to time and set the result equal to
the maximum velocity:

Solving for t2, we get:

Given that the command must be anti-symmetrical to bring the

mass back to rest, the duration of the negative command pulse is immediately known. However, the time between the two pulses (t3 t2)
must be calculated so that the mass moves the desired distance. Again,
this can be done quite simply if we realize that the mass will move half
the desired distance when the command is half finished. In this case,
the input function, f(t), that must be integrated in (2.1) consists of two
parts. The first part, from 0 to t2, has a value of umax. The second part,
from t2 to the command midpoint, tm, is zero. Performing the piecewise integration and setting it equal to half the desired move distance
yields:

Substituting t2 from (2.10) into (2.11) and solving for tm, we get:

The command is anti-symmetrical about the midpoint, so:

26/237

Substituting (2.10) and (2.12) into (2.13) yields a very simple solution for the time location of t3, the beginning of the deceleration pulse:

The deceleration pulse has the same duration as the positive pulse,
so:

The bang-coast-bang command is now completely described by the

three switch times given in (2.10), (2.14), and (2.15). Note that if the
velocity limit is never reached, then the command profile degenerates
to the bang-bang command described by (2.6) and (2.8).

Figure 2.6: Translational Mass with Damper to Ground.

Figure 2.7: Bang-Bang Command for Sliding Mass with Viscous Damping to Ground.

27/237

2.2 Sliding Mass with Constraints to

Ground
In practice, most translating masses experience some resistance to
motion. This resistance is often represented in the system model by a
frictional or viscous damper attached between the moving mass and
ground, as shown in Figure 2.6. The dynamic equation for the system
with viscous damping becomes:

where b is the coefficient of viscous friction.

Here again, if the desired system response is known, then (2.16) will
yield the desired input, f(t). For a mass moving without velocity constraints from one rest position to another, and given the bounded
available force, umax, the command will be similar to the one shown in
Figure 2.2. However, the time duration of the negative force pulse will
be less than the time duration of the positive force pulse, as shown in
Figure 2.7. From an energy perspective, this command shape results
because some of the kinetic energy that enters the system when the
force is applied in the direction of the desired motion is dissipated
through damping. Therefore, a shorter duration negative force is required to remove the remaining energy from the system and bring the
mass back to rest.

28/237

Figure 2.8: Translational Mass with Spring and Damper to

Ground.

Figure 2.9: Bang-Bang Command for Systems with Spring

Connected to Ground.
There is a large class of systems for which a useful system model includes a mass connected to ground by both a viscous damper and a
linear spring, as shown in Figure 2.8. The dynamic equation of motion
for this system is:

where b is the viscous damping coefficient and k is the linear spring

constant.
For these types of systems, it is sometimes the case that the desired
system response includes the constraint that the mass should not
overshoot the desired rest position. One strategy for addressing this
constraint is to set the damping coefficient to a large enough value so
that the system damping ratio,
, is greater than or equal to 1.
When this is the case, and the desired response is a minimum-time
move between two rest states, then the command to be employed is
shown in Figure 2.9. Here again, the time t3 t2 will be less than the
time t2 t1, but the force f(t) will not go to zero when the mass arrives

29/237

at the desired rest position because a steady-state force is required to

counteract the force exerted on the mass by the spring.
Design of the system command becomes more complicated for cases
in which the system parameters are such that underdamped oscillations can occur. As an exercise, the reader may want to try to construct
the minimum-time force profile for such a case. Solutions to this problem will be presented in a subsequent chapter.

Chapter 3
Using Elementary Commands
on Flexible Systems
The commands designed in the previous chapter were derived for
the dynamics of a simple translational mass without oscillatory dynamics, but with consideration of acceleration and velocity limits. The
desired performance was rapid rest-torest motion. (The commands
were actually the time-optimal commands subject to the dynamic limitations.) The results are very convenient in that they are simple
closed-form expressions.
The temptation is to use these well-known and easily derived commands on a wide range of dynamic systems. Unfortunately, when
commands derived for rigid bodies are used on systems with flexibility, they will usually induce a significant amount of vibration. To
demonstrate this problem, we can evaluate the use of the bang-bang
commands on the system model shown in Figure 3.1. The model consists of two masses connected by a linear spring with a control force
acting on the first mass. This model is the simplest description of a
system that has both rigid-body translation and flexible deformation.
The equations of motion for the system are:

where xi is the position of the ith mass, mi are the mass values, and k
is the spring constant. In the examples that follow, the spring constant

31/237

is set to 1 and each of the mass values is set equal to 0.5, so that the
total mass is 1.

Figure 3.1: Two-Mass-Spring System Model.

Figure 3.2: Bang-Bang Response of Two-Mass-Spring

System.

32/237

Figure 3.3: Deflection Induced by Bang-Bang Commands.

3.1 Bang-Bang Commands

When the system is driven with the bang-bang command described
by (2.6) and (2.8), the initial positive pulse will act on the first mass
and start the entire system in motion. However, the connecting spring
will deflect, so the system will start to oscillate. The deceleration pulse
will bring the mass center of the system back to rest at a new location,
but the system will, in general, continue to oscillate. The responses of
the two masses are shown in Figure 3.2 for the case when the desired
move distance is 20 units. The first mass reacts to the control effort
faster than the second mass because the spring must compress for the
force to be transmitted to the second mass. After the mass center is
brought to rest by the negative force pulse, the two masses continue to
oscillate out of phase with each other.
The amplitude of residual vibration induced by the bang-bang command varies with the move distance. Figure 3.3 shows the deflection
(x2-x1) that results from bang-bang commands designed for three different move distances. Note that for the case of xd=10, there is zero residual vibration, This special case arises when the vibration induced
by the positive pulse in the bang-bang command is cancelled by the vibration induced by the negative pulse.

33/237

Figure 3.4: Response to Bang-Bang Commands.

Figure 3.5: Bang-Bang Induced Residual Vibration Amplitude vs. Move Distance.
To further demonstrate the dependence on move distance, Figure
3.4 shows the position response of the second mass, x2, for move distances ranging from 2 units to 40 units. Note that the amount of residual vibration depends greatly on how far the system is moved. To show
this explicitly, Figure 3.5 plots the amplitude of residual vibration as a
function of the move distance. The vibration amplitude does not have
a linear dependence on the move distance as one might expect. Furthermore, the vibration is actually zero for special cases such as move

34/237

distances of 10 and 40 units. The results in Figures 3.3 - 3.5 demonstrate that bang-bang commands designed for a rigid system will usually not work well for a system that has flexibility.

3.2 Bang-Coast-Bang Commands

The bang-coast-bang commands designed using (2.10), (2.14), and
(2.15) will cause similar vibration problems. However, the amplitude
of vibration is a function of both the move distance and the velocity
limit. Figure 3.6 shows the response of the two-mass-spring system
when the move distance is fixed at 10 units, but the velocity limit, vm,
is varied between 0.5 and 1 units/sec. As the velocity limit increases,
the system moves faster, but the amount of transient deflection and
residual vibration vary in a complicated manner.

Figure 3.6: Response to Bang-Coast-Bang Commands.

35/237

Figure 3.7: Residual Vibration Amplitude vs. Velocity Limit

of Bang-Coast-Bang Commands.
Figure 3.7 illustrates this vibration phenomenon. Residual vibration
amplitude is plotted as a function of the velocity limit when the move
distance is held fixed at 10 units. Note that when the maximum allowable velocity reaches a critical level (just over 3 units/sec in this case),
there is no remaining coast period and the bang-coast-bang command
converges to a bang-bang command. After this point, increasing the
maximum allowable velocity has no effect on the command. The results shown in Figure 3.7 demonstrate this effect, as the residual vibration goes to zero, and stays at zero, for maximum velocities above the
critical value. Note that the result of zero vibration from a bang-bang
command, for a move distance of 10 units, was predicted by the results
shown in Figure 3.5. For other move distances, the residual vibration
would not be zero.

36/237

Figure 3.8: Vibration Induced by a Bang-Coast-Bang

Command.
The vibration dependence on maximum velocity combines with the
variation due to move distance that was illustrated in Figures 3.4 and
3.5. Figure 3.8 shows the residual vibration amplitude as a function of
the move distance and the maximum velocity over a fairly small region
of the motion parameter space. Even over this small area, the amplitude of vibration varies widely and is a complicated function of these
two motion parameters. The important result of this investigation is
that bang-coast-bang commands induce significant vibration that varies in a complicated pattern when the system has flexibility.
In the above discussion, the flexible dynamics of the system were
held constant. That is, the mass values and the spring constant were
not changed. Given a real system, an engineer will not know these values with certainty, and they will often vary with time, temperature,
etc. When these system parameters vary, the frequency of the vibration will vary and the amount of vibration will change in ways that are

37/237

comparable with the changes that depend on move distance and maximum velocity shown in Figures 3.5 and 3.7.

3.3 Case Study: Cable Length Effect on

Crane Oscillation
To demonstrate the dependence of the dynamic response on the oscillation frequency, let us reconsider the crane example. When a crane
is moved, the payload will oscillate at a frequency approximated by:

Figure 3.9: Residual Vibration Amplitude vs. the Suspension Cable Length of the SRTC Crane.

where g is the acceleration due to gravity and L is the length of the

suspension cable. Therefore, as the crane hoists the payload up and
down, its dynamic response will change.
To investigate the effect of the changing dynamics caused by hoisting, the response of a bridge crane at the Savannah River Technology

38/237

Center (SRTC) was measured [92]. The bridge can travel back and
forth 85 ft., the trolley can travel 43 ft. side to side, and the vertical
hook can hoist 29 ft. up and down. The crane was repeatedly moved a
horizontal distance of 8 feet using a bangcoast- bang command. For
each experiment, the payload height was varied and the residual oscillation amplitude was recorded. Figure 3.9 shows the amplitude of residual vibration as a function of the suspension cable length. As expected, the amplitude varies with the suspension length and has a dependence similar to portions of the curves shown in Figure 3.7.
Given that commands designed for rigid systems can induce large
amounts of problematic vibration, these commands are often not acceptable to use. Even if the effects of the commands are well understood, the results are usually complex functions of both the motion
parameters and the system parameters. Therefore, compensating for
the induced vibration by changing the motion parameters is not realistic. Designing a feedback controller to address the vibration problem
can also be a challenging task because the controller must accommodate the widely varying dynamic response. The most sensible approach
is to properly design the command profiles, as these are ultimately the
source of motioninduced vibration.

Chapter 4
Command Smoothing
The previous chapters demonstrated that bang-bang and pulse commands can induce unwanted vibration in flexible systems. Other commands with rapid transitions, such as step functions, may cause similar vibration problems. The vibration is induced by the rapid change in
force that occurs at each step change in the command profiles. Therefore, the idea of using smooth commands naturally arises because they
will tend to induce less vibration. Virtually all smooth commands
have a finite first derivative. Smooth commands do not have instantaneous transitions like the step functions that form pulses. The relative smoothness of a command is increased if higher order derivatives
are also finite. An example of a smoothed velocity command for a
crane was shown in Figure 1.6.
The smoothness of a command profile depends on the domain used
to describe the command. For example, a bang-bang force command
is equivalent to a triangular velocity command, which in turn is equivalent to a smooth SCurve in position. Refer to Figure 2.3 to see this
equivalence. Therefore, when analyzing the smoothness of a command
the domain must be clearly stated. A command that is smooth in position may induce abrupt changes in force.
Although smooth commands will often lessen vibration problems,
they will also tend to slow the system down. The trade-off between vibration reduction and operating speed is a fundamental issue that
continually arises. Techniques presented later in this book are designed to maximize the benefits from this trade-off by obtaining large
reductions in vibration, while slowing the system down only slightly.

40/237

Figure 4.1 shows an example of a command that has been smoothed.

This command was formed by starting with a bang-coast-bang function designed for a velocity limit of 2 units/s and a 10-unit move distance. The step functions forming the acceleration pulses were then
replaced by ramp functions that transition from zero to full acceleration in 1 s. The resulting command is a jerk-limited command [50, 65,
98]. The response of the two-mass-spring system in Figure 3.1 to this
type of input is shown in Figure 4.2 for the case when the command
moves the system 20 units. If this response is compared to the response from the bang-bang command that was shown in Figure 3.2,
then we see that the system has less residual vibration. However, the
system takes approximately four seconds longer to first reach the desired location.

Figure 4.1: Bang-Coast-Bang Command Smoothed with Linear Transitions.

41/237

Figure 4.2: Response of Two-Mass-Spring System to a

Bang-Coast-Bang Input Smoothed with Ramp Transitions.
Figure 4.3 shows the response to these smoothed commands for a
variety of move distances. It may not be immediately obvious that the
command smoothing has reduced the vibration amplitude; most of the
responses have noticeable vibration. Figure 4.4 explicitly compares the
residual vibration amplitude resulting from bang-coast-bang (BCB)
commands to that resulting from the smoothed bang-coast-bang commands. The two Xs in the figure correspond to the cases of xd = 10
units and xd = 20 units that are labeled on Figure 4.3. The commands
smoothed by a 1 second ramp reduce the worst case vibration by about
15%. The figure also shows the results when 1.5 second ramps are
used. This extra smoothing reduces the vibration even further.
However, as the command is made smoother, its duration increases,
and therefore, the system takes longer to complete a move.

42/237

Figure 4.3: Response to Bang-Coast-Bang Commands

Smoothed with Ramp Transitions.

Figure 4.4: Residual Vibration from Regular and Smoothed

Bang-Coast-Bang Commands.
The simple command-smoothing approach demonstrated above
does not take into account the specific flexible dynamics of the system.
That is, it does not utilize the systems oscillation frequency in its
design. Rather, it suppresses a wide range of frequencies, but does not

43/237

specifically target any known vibration frequencies. To show this effect, Figure 4.5 compares the frequency content of the original and
smoothed bang-coast-bang commands for a desired move distance of
10 units. As the command is smoothed and lengthened with longer
ramps, the frequency content decreases at high frequencies. However,
the frequency content at low frequencies is not greatly effected by the
smoothing. Therefore, in order to suppress low frequency vibration,
the commands must be made very long and, consequently, the system
will move slowly.

Figure 4.5: Fast Fourier Transform Magnitude.

The results shown in Figure 4.5 are only for the specific move distance of 10 units; however, the general trend applies to all move distances. Figure 4.6 shows the frequency content of the commands for a
large number of cases. The curves are grouped into three categories:
unsmoothed (ramp=0 s), smoothed with 0.5 s ramps, and smoothed
with 1.0 s ramps. The frequency content above 0.5 Hz decreases significantly as the ramp time is increased from 0 to 1 s. However, the frequency content below 0.5 Hz remains nearly constant. The effect of
the command smoothing is clearly a low pass filtering effect.

44/237

4.1 Common Smooth Commands

There are numerous types of smooth command profiles. S-Curves,
trigonometric transition functions, and cam polynomials have been,
and still are, used extensively to command motion of flexible systems.
They are used because they give smooth transitions between system
states. However, these methods usually fail to fully exploit the known
properties of the system such as natural frequency and damping ratio
and instead, simply provide a low-pass filtering effect. While low pass
filtering can reduce residual vibration, it is inefficient because it incurs
a large rise-time penalty [95].
Suppose that the rise portion of a position command is an S-Curve
with a one-unit amplitude, and a rise time of Rc. This position profile
is a piecewise function described by:

Figure 4.6: Fast Fourier Transform Magnitude as a Function of Ramp Time and Move Distance.

45/237

Figure 4.7: Percentage Overshoot of a Second-Order System

in Response to an S-Curve Command.

Figure 4.7 shows the percentage overshoot of an undamped secondorder system as a function of its natural frequency when the system is
driven by an S-Curve with a one-second rise time. The system natural
frequencies at which there will be zero residual vibration occur at multiples of 2 Hz for the case shown. That is, the S-Curve rise time is twice
the duration of the first vibration mode that it can suppress. Furthermore, if the system frequency is above 6 Hz, then there is essentially
no overshoot.

46/237

Figure 4.8: Portable Tower Crane.

4.2 Smoothing Crane Commands

In order to verify some of the key results presented here, experiments were performed on a tower crane located at the Tokyo Institute
of Technology [47]. The crane, shown in Figure 4.8, is approximately 2
m high. The outer limit of its reachable workspace forms a 2 m diameter circle. The crane rotates about the vertical axis, its trolley moves
in and out relative to the support tower, and it hoists a payload up and
down. Encoders measure the radial, angular, and hoisting motion. A
downward-looking camera mounted on the trolley measures the payload swing.
Figure 4.9 shows the trolley position when several S-Curves are
used to move the trolley 1 foot (305 mm). To check the fidelity of the
motion, Figure 4.10 shows that the measured velocity closely matches
the desired triangular velocity profile for each case.

47/237

Using the well-known pendulum relationship between the pendulum length, L, the gravitational constant, g, and the oscillation frequency, we know that the period of pendulum oscillation is:

Recall that if the rise time of an S-Curve is twice that of the system
oscillation period, then there should be no residual vibration as a result of the move. This condition is represented by:

Figure 4.9: Trolley Response to S-Curve Commands.

48/237

Figure 4.10: Trolley Velocity Response to S-Curve

Commands.
Therefore, the pendulum length whose oscillation is canceled by an
S-Curve is:

Using this formula, we can predict that a 4-second S-Curve will cancel the oscillation of a pendulum whose length is 994mm. A series of
experiments was performed using the 4-second S-Curve to move payloads with various suspension lengths. The amplitude of residual vibration induced by the S-Curve command, as a function of the suspension length, is shown in Figure 4.11. The S-Curve does cancel the oscillation near its target length of approximately 1m. However, as the suspension length increases, the residual vibration also increases. Furthermore, as the suspension length decreases from 1m, the residual
oscillation also increases to approximately 30mm peak-to-peak. Additional decreases in the suspension length cause the vibration to decrease back towards zero as the pendulum frequency approaches the
second frequency that the S-Curve can cancel (1.0 Hz). Similar

49/237

experiments were conducted for S-Curves with 3-second and 5-second

rise times. The results from the entire battery of tests is shown in Figure 4.12.

Figure 4.11: Peak-to-Peak Residual Vibration from S-Curve.

Figure 4.12: Peak-to-Peak Residual Vibration from Various

S-Curves.

50/237

While smooth commands can move systems with low levels of vibration, they generally cause the system to have a sluggish response. The
command shaping methods in the following chapters utilize knowledge of the flexible dynamics to maximize effectiveness at reducing vibration, while at the same time moving systems rapidly.

Chapter 5
Impulse Commands for
Vibration Reduction
The preceding chapters have demonstrated that the shape of the reference command going into a flexible system can have a significant effect on the system performance. A well-versed engineer will need to
take this realization a step further and actually design command profiles that produce a desired response. The primary performance requirement is often rapid motion with low levels of vibration. This
chapter will introduce techniques for designing commands that suppress unwanted vibration at specific frequencies. The techniques are
based on the assumption that a good model of the plant dynamics exists. This assumption is not always the case in practice, so subsequent
chapters will present methods to address cases in which accurate system models are not available.
As a first step to understanding how to generate commands that
move flexible systems without vibration, it is helpful to start with the
simplest such command. A fundamental building block for all commands is an impulse. This theoretical command is often a good approximation of a short burst of force such as that from a hammer
blow, or from momentarily turning an actuator full on. Applying an
impulse, A1, to a flexible system will cause it to vibrate. The response
of an underdamped system to such an impulse is shown in the top part
of Figure 5.1. If a second impulse, A2, is applied at a later time, as
shown in the middle of Figure 5.1 (Figure 5.1b), then some of the vibration induced by the first impulse may be cancelled. This concept is
demonstrated at the bottom of Figure 5.1. Impulse A1 induces the

52/237

vibration indicated by the dashed line, while A2 induces the dotted response. Combining the two responses using superposition results in
zero residual vibration. The second impulse must be applied at the
correct time and must have the appropriate magnitude for complete
cancellation. Note that this two-impulse sequence is analogous to the
two-pulse crane command that was shown in Figure 1.12.
To derive the amplitudes and time locations of the impulses shown
in Figure 5.1, a mathematical description of the residual vibration that
results from a series of impulses must be utilized. If an underdamped,
second-order system has an undamped natural frequency of n and a
damping ratio of , then the residual vibration that results from a sequence of impulses can be described by [14, 94, 120]:

53/237

54/237

Figure 5.1: Vibration Caused by Two Impulses Can Cancel.

where,

and

Ai and ti are the amplitudes and time locations of the impulses, n is

the number of impulses in the impulse sequence, tn is the time location of the final impulse, and the damped natural frequency is:

Note that (5.1) is expressed in a nondimensional form. It is generated by taking the amplitude of residual vibration induced by an impulse series and then dividing by the vibration amplitude caused by a
single, unity-magnitude impulse. So, if V is calculated to be 0.05 when
the impulses are entered into (5.1), then the impulses will cause a residual vibration whose amplitude is only 5% of the amplitude that
would result from the application of a single unitymagnitude impulse.

55/237

To generate an impulse sequence that causes no residual vibration,

we can set V equal to zero in (5.1) and solve for the impulse amplitudes
and time locations. However, we must place a few more restrictions on
the impulses because there are an infinite number of solutions that
satisfy this condition. To avoid the trivial solution of all zero-valued
impulses and to obtain a normalized result, we can constrain the impulse amplitudes to sum to one:

The impulses could satisfy (5.5) by taking on very large positive and
negative values. These large impulses would saturate the actuators.
One way to obtain a bounded solution is to limit the impulse amplitudes to positive values:

Limiting the impulses to positive values provides a good solution.

However, performance can be pushed even further by allowing a limited number of negative impulses [117]. This approach will be presented in Chapters 11 - 13. In either case, we want to solve the problem
with the shortest sequence of impulses, so that the motion is accomplished as fast as possible.
The problem we want to solve can now be stated explicitly1:
Find a minimum-time sequence of impulses that makes
the vibration given in (5.1) equal to zero, while also satisfying the impulse amplitude constraints in (5.5) and (5.6).
If we restrict ourselves to finding the two-impulse sequence shown
in Figure 5.1 that satisfies the above specifications, then the problem
has four unknowns - the two impulse amplitudes (A1, A2) and the two

56/237

impulse time locations (t1, t2). Without loss of generality, we can set
the time location of the first impulse equal to zero:

The problem is now reduced to finding three unknowns (A1, A2, t2).
In order for (5.1) to equal zero, the expressions in (5.2) and (5.3) must
both equal zero independently because they are each squared in (5.1).
Therefore, the impulses must satisfy:

Substituting (5.7) into (5.8) and (5.9) simplifies the equations to:

In order for (5.11) to be satisfied in a nontrivial manner, the sine

term must equal zero. This occurs when its argument contains a positive time value and equals a multiple of :

In other words:

57/237

where Td is the damped period of vibration. There are an infinite

number of possible values for the location of the second impulse - they
occur at multiples of the half period of vibration. To obtain the
shortest command we use the smallest value for t2, that is:

The impulse time locations can now be described very simply; the
first impulse is at time zero and the second impulse is located at half
the period of vibration. This result can easily be guessed by examining
Figure 5.1.
For this simple two-impulse case, the amplitude constraint given in
(5.5) reduces to:

Using the expression for the damped natural frequency given in

(5.4) and substituting (5.14) and (5.15) into (5.10) gives:

Solving (5.16) for A1 yields:

58/237

where:

Substituting (5.17) back into (5.15), we get:

The sequence of two impulses that leads to zero vibration can now
be stated in matrix form as:

The impulses given in (5.20) are commonly referred to as a Zero

Vibration (ZV) impulse sequence because they result in no vibration.

5.1 Vector Diagrams

The vibration induced by impulses is mathematically described by
sinusoidal functions. Combining several sinusoidal functions together
in order to determine the residual oscillation amplitude and phase
shift can be challenging. Another approach is to represent the induced
vibration as a vector. Then, the vectors arising from several impulses
can be easily combined geometrically to yield a vector that represents
the residual oscillation resulting from the impulses.
Graphical representations of impulse-induced vibration are displayed on vector diagrams. A vector diagram is created by plotting
each representative vector in polar coordinates. The amplitude of the

59/237

vector is set equal to the amplitude of the applied impulse. The angle,
, at which the vector is plotted is:

Figure 5.2: Impulse Sequence and Corresponding Vector

Diagram.

Figure 5.3: Summing Two Vectors to Get Resultant

Vibration.

60/237

Figure 5.4: Vector Diagram of ZV Shaper.

where, n is the oscillation frequency of the system and ti is the time
at which the impulse is applied to the system. Figure 5.2 shows an impulse sequence and its corresponding vector diagram. Note that the
time (angle) of the first impulse is always set to zero.
When a vector diagram is created, the amplitude of the resultant
from adding the vectors together is proportional to the amplitude of
residual vibration of an undamped 2nd-order system driven by a command convolved with the impulse sequence [93, 114]. The angle of the
resultant is the phase of the residual vibration. This result enables us
to determine residual vibration geometrically. The residual vibration is
calculated by geometrically summing the vectors on the vector diagram, as shown in Figure 5.3.
For the case shown in Figure 5.3, the resultant is larger than either
of the two individual impulses. This indicates that the vibration from
the second impulse adds somewhat constructively with the vibration
from the first impulse.

61/237

Figure 5.5: Belt-and-Shoe Demonstration of Vibration

Cancellation.
Figure 5.4 shows the vector diagram of a ZV impulse sequence designed for an undamped system. In this case the second impulse is 180
degrees out of phase with the first impulses; therefore it lies on the
negative horizontal axis. Furthermore, the amplitudes of the two vectors are equal. When the two vectors are summed, their resultant is
zero.

5.2 Simple Zero-Vibration

Demonstration
The vibration cancellation effect derived above can be demonstrated
by anyone with a belt and shoe, and just a little practice. Take a belt

62/237

and loop it around a shoe. Hold on to the end of the belt so that the
shoe dangles below your hand. This situation is shown on the left side
of Figure 5.5. Make sure that the shoe is hanging motionless. Then, using a quick (impulsive) motion, move your hand to the side, from
point A to C, as shown in the figure. The shoe will lag behind the motion of your hand, but it will soon catch up to your hand and swing
past. This motion is indicated by the arc labeled 2 in the figure.
When the shoe reaches the extreme point in its arc, after one half period of oscillation, it will briefly come to rest. At this point, quickly move
your hand sideways, from point C to B. This motion will move your
hand over the top of the shoe and will cancel the oscillation of the beltand-shoe system. You will have moved a very flexible system from
point A to B with minimal residual vibration.
1
This problem statement and solution are similar to one first published by O.J.M. Smith in 1957 [121].

Chapter 6
Real-Time Generation of
Zero-Vibration Commands
Most real systems cannot be moved around usefully with impulses,
so the properties of the zero-vibration impulse sequence given in
(5.20) must be transferred to a usable command. This can be done by
simply convolving the impulse sequence with any desired command.
The convolution product is then used as the reference command for
the system. If the impulse sequence, also known as an input shaper,
causes no vibration, then the convolution product will also cause no
vibration [9, 94].
This command generation process, called input shaping, is demonstrated in Figure 6.1 using a 2-second pulse as the initial command.
Note that the convolution product in this case is a two-pulse command, similar to the one shown in Figure 1.12 that moved the crane
with no residual vibration. For the case shown in Figure 6.1, the
shaper duration, , is longer than the initial command, but in most
cases the impulse sequence will be much shorter than the baseline
command profile. This is especially true when the system is moving
through a complex trajectory and the period of system vibration is
small compared to the duration of the move. When this is the case, the
shaped command will form a continuous function, as shown in Figure
6.2.
Convolving a two-impulse sequence with a continuous function is
very easy. First, scale the continuous function by the amplitude of the
first impulse. Next, form a secondary function by scaling the original
function by the amplitude of the second impulse and shift it to the

64/237

time location of the second impulse, as shown in Figure 6.2. There are
now two scaled replicas of the original function, one of which is shifted
in time. The final step is to add the two functions together. The solid
line in Figure 6.2 shows the convolution product that results from
summing the two scaled replicas of the original function.

Figure 6.1: Input Shaping a Short Pulse.

Figure 6.2: Input Shaping a Generic Command.

The question that naturally arises is, Why does the shaped command cancel vibration? There are several ways to explain the phenomenon. Since the inception of command shaping for vibration suppression, it has been noted that the process is effectively pole-zero
cancellation. In an early paper showing experimental implementation
of a type of command shaping called Posicast control, Tallman and
Smith wrote [128]:
Posicast control can then be described as a process
whereby complex zeros are generated so as to fall upon
the complex poles of the system . . .

65/237

In his Ph.D. thesis, Singer noted that his input shapers had an infinite number of zeros and that the Fourier transform of the input
shapers for undamped systems were exactly zero at the natural frequency of the system (pg. 63)[93]. Bhat and Miu stated the more general case for damped systems when they wrote [9]:
. . . the necessary and sufficient condition for zero residual vibration is that the Laplace transform of the time
bounded control input have zero component at the system poles.
To rigorously demonstrate the vibration-canceling phenomenon, we
can transfer the input shaping process into the Laplace domain. The
key idea is to establish that the Laplace transform of the system response to the shaped input does not have any frequency content at the
poles of the flexible system. If this is the case, then the shaped command cannot cause the system to vibrate; it cannot inject energy into
the flexible modes.
If we call the initial command f(t), the input shaper i(t), and the
shaped command c(t), then the input-shaping process can be described as:

Taking the Laplace transform of (6.1) yields:

where the capitalized functions of s = + j are the Laplace transforms of the corresponding functions of time. The key result of the

66/237

transform is that the convolution operator has been transformed into

multiplication. Therefore, if I(s) is zero at any value of s, then C(s) will
also be zero, regardless of what F(s) may be.1 We can now restrict our
attention to I(s). For a two-impulse sequence like that given in (5.20),
I(s) can be written as:

For the Zero Vibration (ZV) shaper given in (5.20):

Substituting (6.4) and (6.5) into (6.3) and evaluating the expression
when s = + j yields:

In order for I(s) to equal zero, the exponential term on the right side
must equal -1. To see how this can occur, note that the exponential
term is a complex number that can be written as a real component
multiplied by an imaginary component:

This exponential term will equal -1 when the first component on the
right equals 1 and the component with the imaginary term equals -1:

67/237

Setting the real component equal to 1 yields:

This holds true when:

Setting the imaginary component of the exponential term in (6.7)

equal to -1 yields:

This holds true when:

Combining the results given in (6.10) and (6.12) produces the conditions under which I(s) will equal zero:

Note that these values of s correspond exactly to the pole locations

of an underdamped system whose natural frequency is n and whose
damping ratio is . We have now established that the input shaper has
zero frequency content at the poles of the flexible system for which it
was designed. Given that the Laplace transform of the input shaper
multiplies the Laplace transform of the original command to form the
shaped command, as was shown in (6.2), then the shaped command, C
(s), also has zero energy at the flexible poles.

68/237

To complete the argument, consider that the response of an

underdamped second-order system is:

where G(s) is the system transfer function. Given that C (s) is zero at
the flexible poles, then Y (s) must also be zero at the system poles. This
means that the response has no energy at the flexible poles. The system will not vibrate in response to the shaped command. If the system
has multiple modes, then the argument still holds provided that the
input shaper transfer function, I(s), is zero at every pole of the system
transfer function, G(s). Design of multi-mode input shapers is addressed in Chapter 9.

6.1 Real-Time Implementation

One of the strengths of input shaping is that it can be implemented
when the unshaped command is being generated in real time. This occurs in numerous situations, such as when a human operator is driving a crane. Real-time implementation is possible because the input
shaping technique requires only a simple convolution that can be implemented with a few multiplication and addition operations each
cycle of the control loop. This convolution process does not require a
priori knowledge of the unshaped command signal.
A block diagram representation of the input-shaping convolution is
shown in Figure 6.3. The unshaped reference command, f(t), is fed into n gain blocks that correspond to the impulse amplitudes, Ai. The
scaled functions are then sent through time delay blocks, i, that correspond to the impulse time locations. Note that the first shaper impulse, A1, is located at time zero, so it does not have an associated time
delay block. The shaped command signal, c(t), is formed by simply
summing the scaled and time-delayed functions.

69/237

Many motion control boards and DSP chips have built-in algorithms
for performing the real-time convolution that is necessary for input
shaping. If these features are not available, then a very simple algorithm can be added to the control system. The algorithm starts by
creating a buffer for the command - a vector variable of a finite length.
This buffer is used to store the command values for each time step.
For example, the first value in the buffer would be the shaped command at the first time instance; the value for the nth time step would
be stored in the nth buffer location. A graphical representation of such
a buffer is shown in the upper right-hand corner of Figure 6.4. The
upper left-hand portion of the figure shows the unshaped command at
time step 1.

Figure 6.3: Input Shaping Convolution as a Summation of

Time-Delayed Functions.

70/237

Figure 6.4: Real-Time Input Shaping.

In order to fill the command buffer with the input-shaped command, the algorithm acquires the unshaped command at each time
step. Then, the algorithm multiplies the unshaped command by the
amplitude of the first impulse in the input shaper. This value is added
to the current time location in the shaped-command buffer. The amplitude of the second impulse is then multiplied by the unshaped command. However, this value is not sent directly out to the control loop.
Rather, it is added to the buffer slot that corresponds to the time of the
impulse. For example, assuming a 10 Hz sampling rate, if the time location of the second impulse was at 0.4 seconds, then the value from
the second impulse would be added to the buffer four slots ahead of
the current position. This real-time process will generate the shaped
command, as demonstrated in Figure 6.4. The figure indicates the
state of the algorithm at the first time step and at the eighth time step.
To avoid having the index exceed the size of the buffer, a circular buffer is used where the index goes back to the beginning when it reaches
the end of the buffer.

Figure 6.5: Payload Response with a 975mm Suspension

Length.

71/237

6.2 Case Study: Tower Crane

Oscillation
In Chapter 4, it was shown that smooth commands can move a
tower crane with low levels of payload oscillation. Figure 6.5 shows the
experimental response of a crane to a smooth S-curve command constructed to cause no residual payload oscillation. The figure also shows
the payload response to a ZV-shaped step input. The results clearly
demonstrate that the input shaping works very well, as it reduces the
residual payload sway to near zero. Furthermore, the payload motion
was accomplished 2.6 times faster with the ZV-shaped input than with
the S-Curve. The faster motion of the system is made apparent by the
larger transient swing that occurs during the response to the ZVshaped command. In order to move faster, the trolley must have a
higher acceleration. The increased acceleration leads to the larger
transient deflection. However, even with the much higher acceleration, the payload gets to the desired location with virtually no residual
oscillation.

6.3 Case Study: Micro Milling

Milling machines are used throughout the world to manufacture
complex parts. The essential function of a milling machine is to move
a cutting tool precisely through a complex trajectory in space and
time. If the cutting tool deviates from the intended trajectory, then the
quality of the manufactured part will suffer. Most standard mills are
built to be very sturdy so that they do not deflect when subjected to the
forces between the cutting tool and workpiece.
A new generation of mills have been developed to machine very
small parts. These micro-milling machines can create parts that have
features on the scale of 1m. Because the scale is much smaller than
with traditional milling, the stiffness from the cutting interface to

72/237

ground is likely to be significantly less. While the stiffness of most

micro-mills is large enough to counteract the very small machining
forces, the inertial forces from high-speed motion of the tool can cause
the system to vibrate. Therefore, command shaping is an effective
technique for reducing motion-induced vibration.

Figure 6.6: Micro Milling Machine.

The throughput of micro-milling operations has been limited by the
speed of the spindle. However, researchers have managed to push
spindle speeds up to a few hundred thousand RPM [69]. With advancement in spindle speed it is theoretically possible to move the
structure of the mill faster, while maintaining cut quality and cutting
force.
A high-speed micro-mill at the Escuela Tecnica Superior de Ingenieros Industriales (ETSII) in Madrid, Spain [22] is shown in Figure 6.6.
The machine uses three identical precision linear stages driven by DC
motors in the X, Y and Z directions with encoder resolutions of 0.5m.
The X stage carries the Y Stage, while the Z axis controls the height of
the tool. The tool spindle is air and oil lubricated, as well as

73/237

refrigerated, in order to achieve 120,000 RPM. The stage velocities are

limited to
with maximum accelerations of
The micro-mill employs trapezoidal velocity profiles to complete all
of its moves. Of course, it does not track these trajectories exactly. For
a cut utilizing the motion of only one stage, the error from the desired
trajectory can lead to an irregular surface finish and increased force on
the tool because the cutting tool will oscillate in the cut direction,
thereby creating a varying feed-rate. However, if the trajectory is in
multiple directions, then the error will lead to significant geometric
variations, thereby altering the gross dimensions of the part.
The vibration amplitude of the mill can be significantly lowered
when input shaping is utilized. The reduction in vibration translates
directly into improved machining performance. Figure 6.7 shows the
surface dimensions of a machined part. This data was obtained by
photographing the machined surface using a microscope and then extracting the surface height with an edge detecting algorithm [22]. Use
of input shaping reduced the variation along the machined surface
from 36m to 15m, a reduction of 58%.

Figure 6.7: Part Surface With and Without Input Shaping.

74/237

This is true except for the special case when F(s) is infinite at the
value of s under consideration.

Chapter 7
Robust Commands
The previous chapter introduced the concept of shaping the reference command for vibration reduction. The process was demonstrated
using a shaper that contained two impulses. If the initial reference
command is a step input, then the input shaping process would proceed as shown in Figure 7.1. The second step in the shaped staircase
command is intended to cancel the vibration induced by the first step
in the command. Complete vibration cancellation only occurs if the
second step occurs at exactly the right time and has the correct amplitude relative to the first step. Stated in terms of the input shaper, the
amplitudes and time locations of the impulses must be calculated with
exact knowledge of the vibration parameters (n and ).
If the shaped command was issued to a system that was modeled
perfectly, then the system would respond without residual vibration.
This is shown in Figure 7.2 for the case of an undamped oscillator
whose frequency is 1 Hz. The zero-vibration (ZV) shaper for these
parameters can be obtained by substituting n = 2 (1 Hz) and = 0
into (5.20). For this case, the equation yields:

If there are errors in the modeling parameter values (and there always are), then the input shaping process will not result in zero vibration. This effect is demonstrated in Figure 7.3, which shows the response to the shaped command when the model is perfect and when

76/237

the actual system frequency is 10% smaller than expected (0.9 Hz). As
the plot shows, there can be a noticeable amount of vibration for this
relatively small modeling error. This lack of robustness was a major
stumbling block for the command-shaping methods developed in the
1950s [23, 120, 128].

Figure 7.1: Input Shaping a Step Input.

Figure 7.2: Response to Step and Shaped-Step Inputs.

77/237

Figure 7.3: Response to Shaped Inputs.

7.1 Sensitivity Curves

The robustness of an input shaper can be visualized by plotting its
sensitivity curve. This curve shows the amplitude of residual vibration
as a function of the system parameters. One such sensitivity curve, for
the ZV shaper, is shown in Figure 7.4. The horizontal axis is a normalized frequency and the percentage vibration as defined in (5.1) is on
the vertical axis. Note that as the actual frequency, a, deviates from
the modeling frequency, m, the amount of vibration increases
rapidly.
One solution to the problem of an ineffective input shaper is simply
to redesign the shaper using the correct system parameters. For example, the correct ZV shaper for the 0.9 Hz system would be:

78/237

Figure 7.4: Sensitivity Curve for a ZV Input Shaper.

Figure 7.5: Convolving Two ZV Input Shapers.

This shaper provides perfect vibration cancellation of the 0.9 Hz oscillation. Unfortunately, we will never know the system parameters
perfectly, so we cannot assume that redesigning the input shaper is always an option. Furthermore, the parameters are likely to change
somewhat over time or during the motion of the physical plant. What

79/237

is needed is a robust input shaper that works well even when there are
errors in the system model.

7.2 Combining Input Shapers

One simple way to create a robust input shaper is to combine two
ZV input shapers. For example, the ZV shaper given in (7.1) will cancel
1 Hz vibration, while the ZV shaper in (7.2) will eliminate 0.9 Hz vibration. We can combine the qualities of the two shapers by convolving them together, as shown in Figure 7.5. In this case, the resulting input shaper is:

Figure 7.6: Sensitivity Curves for ZV and Robust Input

Shapers.

80/237

The sensitivity curve for this combined shaper is shown in Figure

7.6. Note that the shaper yields zero vibration at both 0.9 and 1.0 Hz.
In between these two modes, the vibration is also very small. In fact, if
we accept that 5% residual vibration is a small amount of vibration,
then the shaper effectively suppresses all frequencies between 0.8 Hz
and 1.1 Hz. Therefore, if we are uncertain about the frequency value,
but we are confident that it lies in the range from 0.8 to 1.1 Hz, then
the performance of the convolved shaper in (7.3) will be robust to the
uncertainty.
The excellent improvement in robustness comes at the price of increased rise time. Note that the duration of the combined shaper is the
sum of the individual shaper durations. Therefore, the shaped command that results from this shaper will be longer than the one resulting from a single-mode ZV shaper. However, this penalty in rise time
is usually a small price to pay for the robust vibration suppression. To
demonstrate this tradeoff, Figure 7.7 shows the response of the oscillator to step commands shaped by the ZV shaper in (7.1) and the robust shaper in (7.3). The plot shows the response as the oscillation frequency, f, is varied from 0.8 Hz to 1.0 Hz. The rise time of the system
is longer when the robust shaper is used, but the vibration suppression
is very good over a wide range of oscillation frequencies. Therefore,
the system tends to settle more quickly when the robust shaper is
used, even though the rise time is slower.

7.3 ZVD and EI Shapers

The first input shaper designed to have robustness to modeling errors was developed by Singer and Seering in the late 1980s [94]. This
shaper was designed by requiring the partial derivative of the residual
vibration, with respect to the frequency, to be equal to zero at the
modeling frequency. Mathematically, this can be stated as:

81/237

Figure 7.7: Response of Oscillator to ZV and Robust Input

Shapers.

Enforcing this constraint has the effect of keeping the vibration near
zero as the actual frequency starts to deviate from the modeling frequency. The resulting Zero Vibration and Derivative (ZVD) shaper is
given by:

where, K is the variable given in (5.18):

82/237

Note that the ZVD shaper has a duration of one vibration period;
this is exactly twice the duration of the ZV shaper. However, for this
small increase in rise time, a substantial amount of robustness is obtained, as shown by the sensitivity curves in Figure 7.8.
Robust shapers are used when there is a fair amount of uncertainty
in the modeling parameters. In these cases, the actual system will almost certainly not coincide with the modeling parameters. Therefore,
forcing the command to produce exactly zero vibration at the modeling parameters is not a particularly useful design constraint. Using
this realization, input shapers with even more robustness can be obtained, without incurring an additional rise-time penalty. The key is to
relax the constraint of zero vibration at the modeled natural frequency. Rather than force the vibration to zero, simply constrain it below some tolerable level. The sensitivity curve for such an Extra-Insensitive (EI) input shaper is shown in Figure 7.8. For undamped systems the EI shaper is given by [113, 114]:

83/237

Figure 7.8: Sensitivity Curves for Various Input Shapers.

where V tol is the tolerable limit on percentage residual vibration.

The equations describing the EI shaper for damped systems are available in the references.
The robustness of an input shaper can be quantitatively evaluated
by measuring the width of its sensitivity curve at the tolerable vibration level. Such measurements are shown for the 5% level in Figure
7.8. In this case, the 5% insensitivity of the ZVD shaper is 0.286, while
the 5% insensitivity of the EI shaper is 0.40. Note that this significant
increase in robustness does not incur a rise time penalty. Equations
(7.5) and (7.7) show that the ZVD and EI shapers are both one vibration period in duration. The improved robustness results from the
realization that the vibration need not be forced to zero at the modeled
frequency, but rather simply limited to a low level.

7.4 Pole-Zero Plots of Robust Shapers

The analysis of input shaping robustness in the frequency domain
was introduced by Singer [93]. He pointed out that the ZVD input
shapers had a wider trough in the amplitude plot of their Fourier
transforms than did the less robust ZV input shapers. Bhat and Miu
provided a more general statement when they stated the sensitivity to
modeling errors could be reduced by decreasing the magnitude of the
Laplace transform in the neighborhood of the system poles [10]. They
further stated that the ZVD input shaping technique of forcing the derivative of the Laplace transform to zero at the modeling parameters
would achieve such a desired result.

84/237

Figure 7.9: Command Shaping as Pole-Zero Cancellation in

the S-plane.
A clearer explanation of this effect was provided by Singh and
Vadali [100] when they noted that the zero derivative constraint added
a second zero on top of the resonant poles in the s-plane. This case is
shown in the centre of Figure 7.9. Placing additional zeros over the
flexible poles corresponds to setting higher order derivatives to zero
and leads to additional improvement in robustness. However, a more
efficient means of obtaining increased robustness is to place the zeros
from the input shaping process in the neighborhood of the flexible
poles, rather than directly at the modeled locations [74, 111, 114]. This
is the EI approach and it is shown on the right side of Figure 7.9.

7.5 Case Study: Robustness to Crane

Length Changes
Robust shapers have been implemented on numerous machines, including several bridge cranes. The crane at the Savannah River Technology Center, whose vibration characteristics were shown in Figure
3.9, was equipped with a robust input shaper [92]. The shaper was
similar to the EI shaper given in (7.7). It was designed with a tolerable
vibration limit of 2 inches. This provided effective vibration suppression as the suspension cable length varied from 8 to 29 ft. The crane is

85/237

rarely hoisted all the way to the ceiling, so oscillation control for very
short cable lengths was not required.
The input shaping control system was experimentally evaluated as a
function of the suspension cable length. Figure 7.10 shows the experimentally measured residual vibration. This percentage vibration was
computed by taking the vibration amplitude that resulted with the input shaping control enabled and dividing by the vibration amplitude
from an identical move with the input shaping disabled (normal operation). The data indicates that the robust shaper is very successful at
reducing the vibration over the entire useful workspace of the crane.

Figure 7.10: Reduction in Vibration of Savannah River

Crane.

Chapter 8
Extra-Robust Input Shapers
Chapter 7 demonstrated that command shaping can be made robust
to uncertainty in the system parameters. This robustness allows input
shaping to work well on many real systems, such as those whose parameters vary over time or change with configuration. Command shaping can also yield good performance on moderately nonlinear systems.
The cost of the robustness is a small increase in the system rise time.
This rise time increase can be directly correlated with the duration of
the input shaper.
In cases where the system parameters are very uncertain, or are expected to change substantially during normal operation, the control
system may require greater robustness than an EI shaper provides.
Consider a robot arm whose dynamic properties change significantly
when it picks up a payload. Under such circumstances, the inertia of
the system increases, and the oscillation frequency decreases. If the
desired application requires an extra-robust shaper, then several options are available. However, increases in robustness will require some
increase in the shaper duration, and consequently the rise time will
increase.

8.1 Multi-Hump EI Shapers

The EI shaper described in the previous chapter contains three-impulses, has a duration equal to one period of vibration, and yields a
sensitivity curve with a hump centered at the nominal frequency. A
natural extension would be to design a shaper with two humps in its
sensitivity curve, like the one shown by the solid line in Figure 8.1. We

87/237

hypothesize that there exists a shaper containing four evenly-spaced

impulses with a duration of one and a half vibration periods (1.5T)
that will form the sensitivity curve shown in Figure 8.1.
To simplify the derivation of the two-hump EI shaper, we will transfer the problem to a vector diagram. (Refer to Section 5.1 for a discussion of vector diagrams.) A vector diagram of the proposed two-hump
EI shaper is shown in Figure 8.2. The first vector lies on the positive
horizontal axis. Each subsequent vector is rotated 180o from the preceding vector. A modeling error in frequency appears on the vector
diagram as a rotation of each vector through an angle ?i = ti, where
is the frequency error. This effect is shown in Figure 8.3. Once the
vectors have been rotated away from their starting positions, their resultant represents the residual vibration that will occur in the presence
of the modeling error represented by . To distinguish between vectors and their amplitudes, vector quantities are denoted with a hat
symbol.

Figure 8.1: Sensitivity Curves for Shapers with a Duration

of 1.5T.

88/237

Figure 8.2: Vector Diagram of Two-Hump EI Shaper.

Figure 8.3: Effect of Frequency Error on Two-Hump EI

Shaper.
By examining Figure 8.1, we can construct the set of constraint
equations that must be satisfied by the two-hump EI shaper. The first
requirement suggested by Figure 8.1 is that the vibration must be zero
when the modeling frequency is exactly equal to the actual frequency.
This means that the vectors shown in Figure 8.2 must sum to zero.
This gives the following relationship for the impulse amplitudes:

The amplitudes A2 and A4 are preceded by negative signs because

the vectors 2 and 4 point in the opposite direction of 1 and 3.
Given that the desired sensitivity curve is symmetrical, the shaper
amplitudes must also be symmetrical. This yields:

89/237

Equations (8.1)-(8.3) are not independent; combining (8.2) and

(8.1) yield (8.3). Therefore, (8.3) is not used in the following
derivation.
Figure 8.1 also indicates that at some frequency lower than the modeling frequency, H1, the vibration must equal Vtol and the derivative
of the vibration curve must equal zero. These constraints translate to:

where,

In these equations, the rotation of the second vector is, ? = t2,

and is the difference between and H1. Note that ? represents
the frequency shift from the modeling frequency to the frequency that
corresponds to the peak of the first hump in the sensitivity curve. The
constraints contain trigonometric terms with arguments of (i 1)?.
This occurs because the first vector, 1, does not rotate in response to

90/237

a modeling error; it always occurs at time zero. However, 2 rotates ?,

3 rotates 2?, etc.; each vector i rotates (i 1)?.
Finally, the impulse amplitudes must sum to one:

The five equations for the two-hump EI shaper (8.1, 8.2, 8.4-8.8)
contain five unknowns (A1,A2,A3,A4, ?) and one design parameter,
Vtol. We can solve for the input shaper amplitudes as a function of V
tol[111]. Combining (8.1), (8.2), and (8.8) yields:

By expanding (8.4) and (8.5), combining terms, and using (8.9), we

obtain:

and

Equation (8.10) can be solved for cos(?):

Plugging this into (8.11) yields:

91/237

where,

Therefore, the two-hump EI shaper for undamped systems is [111]:

where T is the period of the modeling frequency.

Figure 8.1 compares the two-hump EI shaper to the Zero Vibration
Double Derivative (ZVDD) shaper [94]. The ZVDD shaper is obtained
by appending the ZVD constraints to require the second derivative of
the residual vibration with respect to frequency be equal to zero at the
modeling frequency. Similar to the two-hump EI shaper, the ZVDD
shaper has a duration equal to 1.5 periods of the modeled frequency.
For undamped systems, it is given by:

The 5% insensitivity (the width of the sensitivity curve below the 5%

level) of the two-hump EI shaper is 51% larger than that of the ZVDD
shaper.

92/237

The multi-hump sensitivity curve idea can be extended to a threehump EI shaper by assuming the shape of the sensitivity curve, choosing the form and duration of the shaper, and establishing constraint
equations based on the sensitivity curve. The derivation will be omitted because it is similar to that of the two-hump EI shaper. The
undamped, three-hump EI shaper containing 5 impulses and having a
duration of two vibration periods is [111]:

Figure 8.4: Sensitivity Curves for Shapers with a Duration

of 2T.
where,

93/237

Figure 8.4 compares the three-hump EI shaper to the zero-derivative shaper formed by setting the third derivative equal to zero. This
ZVDDD shaper also has a duration of two cycles of vibration. When
Vtol = 5%, the three-hump EI shaper has 54% more robustness than
the ZVDDD shaper.
The multi-hump EI design procedure can be used for damped systems. Three modifications to the above process must be performed.
First, the damped vibration equation and the derivative of the damped
vibration equation must be used in place of the undamped equations
in (8.4) and (8.5). Second, the assumption of a symmetrical input
shaper must be discarded. These discarded equations are replaced by
constraints describing the sensitivity curve on both sides of the modeling frequency. The above undamped procedure only constrained one
half of the sensitivity curve, because the symmetry of the input shaper
ensured symmetry of the sensitivity curve. Third, the set of constraint
equations must be solved numerically [111].
The amplitudes and time locations of the two-hump EI (Vtol = 5%)
shaper and the three-hump EI (Vtol = 5%) shaper are given in Table
8.1. The curve fits for the two-hump EI shaper have maximum errors
in the impulse times and amplitudes of less than 0.5% over the range
0 0.3. The curve fits for the three-hump EI shaper are accurate to
within 0.4% over the range 0 0.2.

8.2 Specified-Insensitivity Shapers

The multi-hump EI shapers and the zero-derivative shapers provide
several discrete choices for tailoring input shaper robustness to a particular application.
Table 8.1: Damped Multi-Hump EI Shapers.

94/237

However, suppose that an intermediate level of robustness is

needed. In this case, we need a design method that can create an input
shaper with any desired level of robustness.
A straightforward method to achieve any level of robustness is to
simply suppress the vibration at several frequencies within the desired
frequency suppression range. This process, known as frequency
sampling [91] or parameter sampling [115], is demonstrated in Figure
8.5. In the case shown, the vibration is limited at six frequencies over a
range of 20% of the model frequency, m. (This 20% range corresponds to an insensitivity of 0.4.) Because this approach allows the designer to specify the suppressed frequency range, the resulting shapers
are called Specified Insensitivity (SI) shapers [115].
The number of constrained frequencies within the suppression
range is set by the designer. Using fewer points reduces the number of
constraint equations and may make the problem easier to solve numerically. However, using a smaller number of points allows the vibration curve to further exceed the vibration limit between points, as
shown near the center of the hump in Figure 8.5. In practice, the number of constrained frequencies can be very low (such as the 6 shown in
Figure 8.5) because the vibration curve cannot rapidly change its
slope. Furthermore, exceeding the vibration limit by a small amount is
generally inconsequential. Also note that even though the SI problem
formulation may require more constraint equations than the multihump EI approach, the equations are simpler in that there are no constraints on the derivatives of the sensitivity curve. The derivative equations, such as the one in (8.5) are complex and much more difficult to
satisfy than the vibration equations.

95/237

Figure 8.5: Specified Insensitivity Constraints.

When given the task of designing an input shaper, a controls engineer will be faced with a trade-off between robustness and rise time
(shaper duration). Figure 8.6 provides information about the time cost
of obtaining a desired level of 5% insensitivity. (Recall that insensitivity, I, is the nondimensional width of the sensitivity curve that is below
the tolerable vibration limit - 5% vibration in this case.) It is unlikely
that a specific value of I will be required. Rather, an acceptable range,
say 0.4 I 0.6, will likely be under consideration. The question then
arises as to which value of I should be used to design the shaper. If I =
0.4 is chosen, then the shaper will have a duration of T - one period of
the modeling frequency. If I = 0.6 is chosen, then the insensitivity increases 50%, but the duration of the shaper increases to 1.37T. The
controls engineer must balance this tradeoff in a way best suited for
the particular application.
There are several interesting properties of the SI shaper. The duration of the shaper and, therefore, the rise time increase with I. The
duration does not steadily increase; rather there are locations where
the slope changes rapidly. These points correspond to ZV and EI
shapers. The first rapid change, at approximately I = 0.06, corresponds to the ZV shaper. The second change, at approximately I = 0.4,

96/237

corresponds to the EI shaper. The third change, at approximately I =

0.72, corresponds to the two-hump EI shaper. The ZVD shaper does
not correspond to any SI shapers because it does not maximize insensitivity for its duration.
The optimal value of I will depend on several factors including the
relative importance of rise time and robustness. To aid in choosing I,
we can calculate the ratio of insensitivity to shaper duration. Figure
8.7 shows the ratio of (5% insensitivity)/(shaper duration) as a function of I. The points at which the SI shaper equals the ZV, EI, and twohump EI shapers are local maximum points. This result indicates that
if the desired insensitivity range includes one of these points, then
those shapers are good choices.

Figure 8.6: Duration of SI Shaper for Undamped Systems.

97/237

Figure 8.7: Ratio of Insensitivity to Shaper Duration.

As the required robustness is increased, the impulse amplitudes
change in a complicated manner. Figure 8.8 shows the amplitude of
the first SI shaper impulse, A1, as a function of the insensitivity, I, and
the damping ratio, , for the case of Vtol = 5%. The figure is divided into 0, 1, and 2-hump regions by thick lines. Thin lines further divide the
data into regions where A1 remains constant or changes as I varies. If I
is held constant and is increased, then A1 increases. On the other
hand, if is held constant and I is varied, then A1 changes in a complicated manner. For low values of I, A1 is equal to the first impulse of
a ZV shaper. As I increases, A1 decreases in value until it reaches the
value of the first impulse in an EI shaper. Then, A1 remains constant at
this value over a large range of I. With further increases in I, A1 decreases again to the value corresponding to a two-hump EI shaper.

98/237

Figure 8.8: Amplitude of the First Impulse in the SI shaper.

8.3 Multiple Parameter Specified

Insensitivity
Input shaping robustness is not restricted to errors in the frequency
value. Shapers can be designed for robustness to changes in other system parameters, such as damping. Fortunately, most input shapers
tend to be insensitive to errors in the damping ratio estimation.
However, to ensure good performance when the damping ratio is expected to change significantly, robustness in damping can be achieved
in the same manner as the robustness to frequency errors. Figure 8.9
shows a three-dimensional sensitivity curve for an SI shaper that was
designed to suppress vibration between 0.7 Hz and 1.3 Hz and also
over the range of damping ratios between 0 and 0.2. Vibration equations for various damping values were used as constraint equations to
limit the vibration over the desired range of damping ratios.

99/237

8.4 Input Shaping & Traditional

Digital Filters
Input shapers are finite impulse response (FIR) filters [70, 79, 129,
147]. Engineers skilled in the design and use of traditional digital FIR
filters will notice many similarities between digital filters and input
shapers. This raises two interesting questions, What is unique about
input shaping?, and Why does it have advantages over traditional
FIR filters?
The first step in answering the questions is to realize that the main
applications for input shaping involve minimizing the oscillations of
mechanical systems. In mechanical oscillation control, the engineer
wants to remove harmful excitation frequencies from the reference
command signal. However, in traditional filter design there is always
the requirement that some frequencies must pass through the filter
without substantial attenuation. Therefore, the design procedure with
input shaping differs from traditional digital FIR filtering in at least
one significant way; no frequency passband is used. Because there are
no passband requirements, the design of input shapers is not only
simpler, but the resulting shaper will have a shorter duration than a
corresponding FIR filter that must pass certain frequencies.

100/237

Figure 8.9: Three-dimensional Sensitivity Curve.

To further demonstrate this important result, consider that traditional FIR filters can be considered as one of five basic types [70]:
Lowpass filter: Pass low frequency vibrations, attenuate high
frequencies.
Highpass filter: Pass high frequency vibrations, attenuate low
frequencies.
Bandpass filter: Pass a range of frequencies, attenuate frequencies
above and below this range.
Bandstop filter: Attenuate a range of frequencies, pass frequencies
above and below this range.
All-pass filter: Pass all frequencies, introduce predictable phase shifts
at certain frequencies.
The frequency domain performance specifications for the five types
of FIR filters are compared to the performance specifications for an

101/237

input shaper in Figure 8.10. The gray areas are desired regions for the
filter frequency response. Each type of filter has some desired passband, but the input shaper does not.
The type of traditional FIR filter most closely related to input shaping is the bandstop filter, or notch filter, shown in Figure 8.10d.
However, these filters are designed so that frequencies outside of the
stop band are passed with only minor attenuation. Input shaping has
no such performance requirement. Therefore, the set of performance
specifications for input shaping is less extensive than for traditional
filtering. In fact, as shown in previous chapters, many input shapers
do not have a range of stop frequencies, as shown in Figure 8.10f. Instead, limits on the frequency response amplitude are applied at only a
single frequency, or a set of individual frequencies. As a result of the
simplicity of the performance specifications, many types of input
shapers have been determined in closed-form, whereas closed-form
solutions of traditional FIR filters are not available.

102/237

Figure 8.10: Frequency-Domain Performance

Specifications.

103/237

In addition to the fundamental difference of pass band specifications, input shaping has certain characteristics that differ from the
general characteristics of traditional FIR filtering. Input shaping techniques can readily accommodate damped frequencies because mechanical systems always have some damping. On the other hand, traditional filtering methods are often based on frequency domain techniques that assume undamped dynamics. The number of filter coefficients is generally chosen with filtering, but not with input shaping.
Finally, the duration of an input shaper is almost always minimized.
On the other hand, filters are usually designed by choosing a filter duration and then minimizing the stop-band amplitude or the pass-band
ripple.
To summarize, input shaping is advantageous when compared to
traditional FIR filtering because it has a low number of impulses, short
duration (filter length), and large robustness to modeling errors (stop
band width).

Figure 8.11: Schematic Diagram of MACE Experiment.

104/237

Figure 8.12: MACE Experiment Onboard the Space

Shuttle.Photo courtesy of NASA.

8.5 Case Study: MACE Mission

The performance of multi-hump EI shapers was tested in space using the Middeck Active Control Experiment (MACE) [66, 111, 133]. As
shown in Figure 8.11, the MACE hardware was designed to represent a
typical satellite with multiple pointing mechanisms. This experimental
apparatus first flew on board the Space Shuttle Endeavor in March
1995, as shown in Figure 8.12. It was then relaunched in September of
2000 to the International Space Station. A goal of the experimental
program was to develop control algorithms that would allow both gimbals to simultaneously make accurate motions. This requires each
gimbal to be robust to disturbances caused by the motion of the other
gimbal.
In one set of experiments, a white noise signal was fed into the actuator of the primary gimbal. The resulting structural motion was then
recorded at the secondary gimbal. The secondary gimbal response was
analyzed to determine the frequency response of the structure under
such conditions. Figure 8.13 shows the experimentally-obtained frequency response. With the white noise excitation, the frequency response indicated a resonant frequency near 2 Hz. When the excitation

105/237

signal was shaped by a two-hump EI shaper, the amplitude of the system response at the resonance frequency was sharply reduced. The
two-hump notch in the frequency response near 2 Hz is readily apparent in Figure 8.13.

Figure 8.13: MACE Frequency Responses.

106/237

Figure 8.14: MACE Step Responses.

Other experiments recorded the step response of the structure. Figure 8.14 shows the shaped and unshaped responses of the primary
gimbal when it was moved approximately 3 degrees. The shaper eliminates virtually all residual vibration. The very low frequency drift in
the position was caused by the umbilical that connected the free-floating hardware to the space shuttle.

Table 8.1: Damped Multi-Hump EI Shapers.

Shaper

TwoHumpEI

M0

M1

M2

M3

t2

0.49890

0.16270

-0.54262

6.1618

t3

0.99748

0.18382

-1.58270

8.17120

t4

1.49920

-0.09297

-0.28338

1.85710

A1

0.16054

0.76699

2.26560

-1.2275

A2

0.33911

0.45081

-2.58080

1.7365

A3

0.34089

-0.61533

-0.68765

0.4226

A4

0.15997

-0.60246

1.00280

-0.931

108/237

ThreeHumpEI

t2

0.49974

0.23834

0.44559

12.472

t3

0.99849

0.29808

-2.36460

23.399

t4

1.49870

0.10306

-2.01390

17.032

t5

1.99960

-0.28231

0.61536

5.4045

A1

0.11275

0.76632

3.29160

-1.4438

A2

0.23698

0.61164

-2.57850

4.8522

A3

0.30008

-0.19062

-2.14560

0.1374

A4

0.23775

-0.73297

0.46885

-2.086

A5

0.11244

-0.45439

0.96382

-1.460

Chapter 9
Multi-Mode Input Shaping
Every real system has several flexible modes. Fortunately, the amplitude of vibration associated with higher frequency modes is inconsequential most of the time. The largest contribution to vibration amplitude arises from the primary flexible mode, which is usually the
mode with the lowest frequency. However, when systems are pushed
to their performance limit, additional modes can be excited to a level
that degrades performance. Furthermore, some systems are not naturally dominated by a single mode, rather two or three modes play important roles in the dynamic response. When multiple modes make
significant contributions to the dynamic response, command shaping
can be very useful.

9.1 Convolved Shapers

Several methods have been developed for generating input shapers
to suppress multiple modes of vibration [30, 71, 82, 99, 106]. A simple
method for constructing a multi-mode input shaper is to design an input shaper independently for each problematic mode of vibration.
Then, the individual input shapers can simply be convolved together
to produce a multi-mode shaper. This straightforward process is
demonstrated in Figure 9.1. The duration of the multi-mode shaper is
equal to the sum of the durations of the individual shapers. The robust
shaper in (7.3) was a product of this design process.
When shapers are convolved together, the vibration suppressing
properties of the individual shapers are passed on to the multi-mode
shaper. The sensitivity curve of the multi-mode shaper is formed by

110/237

multiplying together the sensitivity curves for the single-mode

shapers. Figure 9.2 shows this process for two ZV shapers designed to
suppress 1 Hz and 2 Hz modes. All of the frequencies suppressed by
the 1 Hz shaper (1 Hz, 3Hz, ...) and by the 2 Hz shaper (2 Hz, 6 Hz, ...)
are suppressed by the convolved shaper.

Figure 9.1: Generating a Multi-Mode Input Shaper.

111/237

Figure 9.2: Combining Sensitivity Curves of Convolved

Shapers.

112/237

If the shapers being convolved together contain impulses with only

positive amplitudes, then each single-mode shaper sensitivity curve is
less than or equal to 100% vibration at all frequencies. Therefore,
when the two curves are multiplied together, the vibration amplitude
will be reduced or stay the same at every frequency. This causes the
combined shaper to have better robustness properties than either of
the individual shapers. For example, suppose a system has undamped
modes at 1 Hz and 2.5 Hz, then ZVD shapers for each mode are:

Figure 9.3: Sensitivity Curves for Two-Mode ZVD Shapers.

113/237

Convolving the shapers given in (9.1) and (9.2), results in:

The sensitivity curves for the two single-mode shapers given in (9.1)
and (9.2) and the convolved two-mode shaper are shown in Figure 9.3.
The convolved shaper is very robust to modeling errors or variations
of the mode at 2.5 Hz. This large insensitivity occurs because the
shaper designed to eliminate 1 Hz vibration also suppresses 3 Hz vibration. The combined effect allows the second mode to range
between 2.1 Hz and 3.3 Hz without exceeding a 5% residual vibration
level.

9.2 Simultaneous Shapers

Multi-mode convolved shapers are designed by satisfying vibration
constraints on each mode independently and then combining the resulting input shapers. A direct solution for the input shaper design can
be obtained by simultaneously satisfying the constraint equations for
all modes. For example, the zerovibration equations for two distinct
frequencies, 1 and 2, that have associated damping ratios of 1 and
2 are:

114/237

These equations can be solved simultaneously (impulses amplitude

limitations must also be specified) to determine the impulse amplitudes and time locations of the shortest duration two-mode ZV
shaper. The solution will be different than when two ZV shapers are
convolved together.
A drawback to the direct design approach is that it is more difficult
to complete than convolving two single-mode shapers obtained from
closed-form equations. In general, simultaneous shapers must be designed using an optimization program. In spite of this drawback, there
are some advantages to the direct solution. Most importantly, a simultaneous shaper will always produce a faster rise time than a convolved
shaper.
The simultaneous solution for two (or more) distinct frequencies
with ZVD, EI, or SI robustness properties can be obtained by adding
the robustness constraints to those listed above for the zero vibration
shaper. Returning to the example system with two undamped modes
at 1 Hz and 2.5 Hz, the simultaneous ZVD shaper is:

Figure 9.3 compares the sensitivity curves of the simultaneous solution in (9.5) and the convolved solution in (9.3). While the robustness
to errors in the low mode are approximately the same, the convolved
shaper is much more robust to modeling errors of the second mode.
On the other hand, the direct solution is 18.4% shorter in duration and
contains four fewer impulses. Having fewer impulses means that the
real-time execution of the shaping convolution will be faster. Having a
shorter input shaper duration means that the system rise time will be
faster. Figure 9.4 compares the time duration of the convolved and
simultaneous two-mode ZVD shapers as a function of mode ratio,
2/1. The convolved ZVD shaper is up to 33% longer than the

115/237

simultaneous ZVD shaper, and it averages 12% longer for the range
shown in the figure.

9.3 Case Study: Double-Pendulum

Cranes
Cranes often swing with a single-pendulum oscillation mode.
However, under certain conditions double-pendulum dynamics can
arise. For example, these effects occur when a lumped payload is suspended a significant length below the crane hook, as shown in Figure
9.5. In this case, the payload forms a secondary pendulum that rotates
about the connection point to the hook. This secondary mode becomes
significant when the hook has a large mass as compared with the payload [119]. Double-pendulum effects can also occur when the payload
is an extended body with significant inertia, as illustrated in Figure 9.6
[58].

Figure 9.4: Duration of Convolved and Simultaneous

Shapers.

116/237

Figure 9.5: Double-Pendulum Crane: Point Masses

Figure 9.6: Double-Pendulum Crane: Distributed Payload

Consider the double-pendulum crane shown in Figure 9.5. The
crane is moved by applying a force, u(t), to the trolley. A cable, of
length L1, hangs below the trolley and supports a hook, of mass m1, to
which the payload is attached using some form of rigging. The rigging
and payload are modeled as a second cable, of length L2, and point
mass, m2. The linearized frequencies of the double-pendulum dynamics are [11]:

117/237

where,

Figure 9.7: Double-Pendulum Crane Frequencies.

R is the ratio of the payload mass to the hook mass. The two oscillation frequencies depend on the two cable lengths and the mass ratio. It
is of interest to investigate how the frequencies change as a function of
the system parameters. Such information could be used to design the
rigging to support the payload, or it could be used to design an
effective input-shaping control scheme.

118/237

Consider a subset of crane manipulation tasks where the crane is

used to move payloads that remain near the ground. That is, hoisting
up and down is kept to a minimum and the sum of the two cable
lengths stays roughly constant. Figure 9.7 shows the two oscillation
frequencies as a function of the rigging length, L2, when the total
length (cable length plus rigging length) is held constant at 8.5 m and
the mass ratio is approximately 0.51. The low frequency changes very
little and corresponds closely to the frequency of a single pendulum
with a length of 8.5 m. On the other hand, the second mode has a
strong dependence on the rigging length.
To demonstrate multi-mode input shaper design for double-pendulum dynamics, input shaping was implemented on the portable bridge
crane shown in Figure 9.8 [46]. The crane was equipped with a twomass double pendulum similar to that shown in Figure 9.5. Figure 9.9
shows the experimentally measured responses of the portable bridge
crane when it was moved approximately 1 foot using the standard on/
off velocity commands generated when a human operator pushes the
control buttons. For the case shown, the mass ratio of the double pendulum was set to 2 and the length ratio, L1/L2, was 1. For this case, the
first and second modes of vibration occur at 0.55 Hz and 1.7 Hz. Figure 9.9 also shows the case when ZV shaping is utilized to suppress the
low mode. While the low mode is virtually eliminated, the second
mode is still problematic.
In order to adequately suppress the double-pendulum dynamics,
two-mode shapers must be utilized. Either the convolved or simultaneous approach will yield effective shapers. However, in order to maximize the response speed, simultaneous shapers must be designed.
Furthermore, to tailor the shaper to the system dynamics, the
Specified-Insensitivity (SI) shaping method discussed in Section 8.2
should be utilized. The SI method can be extended to two-mode systems by placing vibration constraints over frequency ranges near both
of the expected frequencies. Figure 9.10 demonstrates this approach
when modes near 1 Hz and 2.7 Hz are suppressed.

119/237

Figure 9.8: Portable Bridge Crane.

Figure 9.9: Experimental Response, R=2.

A two-mode SI shaper was designed for the portable crane for a case
when the mass ratio is 2. The shaper was designed to accommodate a
5% variation in the low mode (0.55 Hz) and a 10% variation in the

120/237

high mode (1.72 Hz). The resulting input shaper contains five
impulses:

Figure 9.10: Frequency Sampling Over Two Ranges.

121/237

Figure 9.11: Sensitivity Curves for the Two-Mode SI

Shapers.
The sensitivity curve for this shaper is the solid line in Figure 9.11.
Figure 9.12 shows the experimental response when this double-pendulum SI shaper is utilized. As predicted by the sensitivity curve, both
the first and second modes are effectively suppressed by the input
shaping process.
For low mass ratios, the second mode will be more important and
the need for two-mode shaping will increase. When the mass ratio is
lowered to 0.5, the two frequencies shift to approximately 0.58 Hz and
1.13 Hz. Note that the low mode is virtually unchanged, but the second
mode has shifted considerably. For these values, the two-mode SI
shaper is:

Figure 9.12: Experimental Response with Double-Pendulum

SI Shaper, R=2.

122/237

Figure 9.13: Experimental Response with Double-Pendulum

SI Shaper, R=0.5.
The sensitivity curve for this shaper is shown by the dashed line in
Figure 9.11. The oscillation suppression characteristics near the low
mode are nearly identical to the SI shaper for R=2; however, the
second-mode suppression range has shifted to a lower frequency
range to match the decreased second-mode frequency of the R=0.5
case. The experimental responses shown in Figure 9.13 demonstrates
that the two-mode SI shaper works well when R=0.5.
1
These values match common operating conditions of a bridge crane
at the Savannah River Technology Center (SRTC) [92].

Chapter 10
Shaping Operator Commands
The most common and useful position for an input shaper is outside
of any feedback loops, as was shown in the generic control diagram of
Figure 1.8. However, it is possible to insert an input shaper inside a
feedback loop, as shown in Figure 10.1. In many cases, this configuration does not provide advantages, and can create significant problems.
The major disadvantage is that system stability can be degraded by the
partial time delays caused by inserting the shaper into the feedback
loop. However, there are a few applications where placing a shaper inside of a feedback loop can be advantageous 1) when the feedback controller is a human operating a flexible machine and 2) when the system has hard nonlinearities. This chapter will examine the case when
human crane operators utilize input shaping.

10.1 Operators of Single-Pendulum

Cranes
Human-controlled positioning systems are used throughout the
world to perform challenging manipulation tasks. Auxiliary control
schemes can be added to human-operated systems to improve performance. For example, even a skilled crane operator can have difficulty maneuvering a payload without inducing large amounts of sway.
Therefore, a secondary control scheme may be added to ensure lowsway motions.
When a human operator attempts to maneuver payloads using an
overhead bridge crane, the oscillations make it difficult to manipulate
the payload quickly and with good positioning accuracy. If the

124/237

workspace is cluttered with obstacles, then the oscillations can create

significant safety issues, especially when the payload or obstacles are
of a hazardous or fragile nature. This problem was demonstrated by
the bridge crane payload response displayed in Figure 1.13.

Figure 10.1: Input Shaper Inside a Feedback Loop.

Figure 10.2: Test Course.

125/237

Figure 10.3: Completion Times.

The figure also shows the advantage provided by input shaping. In
this case, the human operator is acting as a feedback controller using
visual feedback to adjust the desired path in real time. The operatorgenerated commands are then modified by the input shaper, which is
inside the feedback loop formed by the human operator.
Given that input shaping changes the dynamic nature of the crane, it
is important to investigate how operator performance and system stability are affected by the addition of input shaping. Volunteer operators ran the 10-ton Georgia Tech bridge crane through the obstacle
course shown in Figure 10.2 [35]. The goal was to move the crane from
the start region to the end region quickly, but without running into
obstacles. The target area was a 22-inch diameter circle. In order to
observe changes in operator behavior, the course had distinct path
choices for reaching the target. These long and short paths are shown
schematically in Figure 10.2. The shorter path is more difficult because it has more narrow bends and turns as compared to the longer
path.
Figure 10.3 show the task completion times. The dark bars represent
the unshaped runs, whereas the light bars represent the runs with input shaping enabled. The average time to complete the course with

126/237

input shaping was 51 seconds. The average times for the unshaped
runs was 135 seconds. Input shaping reduced the completion time by
more than 60%.

Operator Learning
Although input shaping cancels out payload swing, skilled operators
can also employ manual swing control techniques to reduce payload
sway. On the other hand, most novice crane operators rely on a passive approach for swing cancellation. They limit themselves to slow and
simple movements in an attempt to avoid large oscillations. When
large oscillations do arise, the operators usually elect to take less challenging paths to the target, or they make numerous small motions to
proceed slowly through a dangerous region. This tactic works, but results in long task completion times. The large swings can be actively
damped if the human operator can appropriately manipulate the crane
motions. Acquiring the skill to achieve an effective manual swing control requires a great deal of practice. A study was conducted to determine the effect of crane operator learning with and without input shaping [35]. Some operators drove the crane only 3 times, whereas others
drove the crane up to 9 different times. The average testing frequency
was 2.2 trials per week.

127/237

Figure 10.4: Task Completion Times for Subject 1.

Figure 10.5: Task Completion Times without Input Shaping.

Figure 10.6: Task Completion Times with Input Shaping.

Figure 10.4 shows the task completion time for a typical subject.
After only a few trials, the operator became more skilled at running
the crane without input shaping. On the other hand, the operator
learning did not help much in the cases when input shaping was

128/237

enabled. Even though the operator became more skilled over time, the
performance without input shaping was still considerably poorer than
when input shaping was utilized. Figure 10.5 shows the progression of
completion times without input shaping for all of the operators. There
is a clear downward trend in completion time as the number of trials
increases. Figure 10.6 shows the corresponding information when input shaping was enabled. In this case, the operator learning does not
have a big effect on task completion time. However, the completion
times with shaping are considerably shorter than without input shaping, even after substantial operator learning on the unshaped crane.
The results indicate that significant learning is not needed when input
shaping is utilized. With input shaping enabled, crane operators immediately become safer and more efficient in their manipulation tasks.

10.2 Operators of Double-Pendulum

Cranes
When cranes move large or complicated payloads, the manipulation
task can become more difficult than when the payload acts like a
single-pendulum. One important case of these complicated dynamics
is the double-pendulum effect discussed in Section 9.3 [109]. To investigate the effect of input shaping on operators of these more complicated cranes, additional studies were conducted with the 10-ton
crane at Georgia Tech. Fifty test subjects drove a double pendulum
crane formed by attaching a payload to the hook with a rigging cable
[40]. The initial length of the suspension cable was approximately 3.8
m and the rigging length was approximately 2.3 m. The operators
changed the suspension length when they performed hoisting operations. The mass of the hook is approximately 50 kg. Three different
payloads were used during testing (4.5 kg, 13.6 kg, 25 kg).
For the experiments described in this section, only the 4.5 and 13.6
kg payloads were used. An SI shaper was designed for the average of
the two payloads (9 kg). The frequencies for this payload and a 3.8 m

129/237

suspension cable length are approximately 0.25 Hz and 0.35 Hz.

Given the uncertainty in the crane parameters and the changes in suspension length during hoisting, the input shaper needs to be moderately robust. Therefore, the shaper was designed to accommodate a
5% variation around the low-mode frequency and a 15% variation
for the high-mode. The SI shaper for these design constraints is:

Three different types of manipulation tasks were investigated. The

first test was a point-to-point motion. The crane operator moved the
payload from a START area (a 0.5m2 square) to an END region (a
0.75m diameter circle) that was two meters away. Sixteen operators
were tested with this simple task. The second test included an obstacle
in the desired path. The 1.2 m wide by 1.2 m tall obstacle was placed
one meter from the START square, forcing the operator to maneuver
around it. Typical responses of the crane hook during this test are
shown in Figure 10.7. Twenty operators were tested with this more
challenging task. The third test required the operator to hoist the payload over the obstacle. Fourteen operators were tested with the hoisting task.
The time to complete the tasks, the number of obstacle collisions,
and the number of times the operator pushed the control buttons were
recorded. A significant difference between the three cases exists because moving around the obstacle induces two-dimensional oscillation, rather than the planar motion induced by the straight-line test.
Hoisting of the payload during transverse motion induces a different
dynamic response because the system frequencies vary with suspension length. Each human operator drove the crane through one of the
three types of motions: point-to-point, obstacle avoidance, or hoisting.

130/237

The human operator drove the crane with and without input shaping,
and with 4.5 kg and 13.6 kg payloads.

Figure 10.7: Typical Hook Responses During Manipulation

Tests.

Figure 10.8: Point-to-Point Completion Time (13.6 kg).

131/237

The task completion times for the sixteen operators performing

point-topoint motion with the 13.6 kg payload are shown in Figure
10.8. The average unshaped task completion time was 113.8 sec, but
the average time using input shaping was only 12.9 sec. The task completion times for the obstacle avoidance tests using the 13.6 payload
are shown in Figure 10.9. The average completion time without input
shaping was 152.5 sec. With SI shaping enabled, the average time of
completion was reduced to 29.6 sec. The hoisting task completion
times using the 13.6 kg payload are shown in Figure 10.10. The unshaped average completion time was 114.3 sec. With input shaping enabled, the average completion time was reduced to 41.5 sec.
Without input shaping enabled, the operators had difficulty completing their tasks. As a result, they had to work harder by repeatedly
pushing the control button in an effort to reduce the cranes oscillations. To quantify this measure of effort required by the operator, the
number of times the operator pushed a control button was recorded.
Figure 10.11 shows the number of button pushes for each operator as
they performed the point-to-point motion with the 4.5 kg payload.
Considering all of the tests, the average number of button pushes for
the shaped motion was 2.3 button pushes. However, for the same task
and without input shaping, the average number of button pushes increased to 21. This represents an increase of almost 900% in the operator effort for the pointto- point task.

132/237

Figure 10.9: Obstacle Avoidance Completion Time (13.6 kg).

Figure 10.10: Hoisting Completion Time (13.6 kg).

133/237

Figure 10.11: Point-to-Point Operator Effort, Pendent (4.5

kg).
Similarly large reductions in operator effort were obtained with input shaping in the obstacle avoidance and hoisting tests. Without input shaping, the operators used an average of 36 button pushes for the
obstacle avoidance task, but only 4.5 with SI shaping. Thus, there was
a 794% increase in the number of button pushes without input shaping enabled. For the hoisting task, there was an average of 16.8 pushes
without shaping and 3.8 button pushes with shaping enabled; a 450%
increase when shaping was not enabled.

Operator Learning on Double-Pendulum Cranes

In order to study operator learning on double-pendulum cranes, a
setup similar to the one in the previous section was used, but a 25 kg
payload. Two different types of manipulation tasks were investigated,
obstacle avoidance and hoisting over obstacles. Eight operators drove
the crane six times each, over a period of eight days [39]. The objective
was to observe how the operators learned to drive the crane. The operators were part of a crane-driving contest where the goal was to

134/237

complete the manipulation task as fast as possible without colliding

the payload into the obstacle. For each collision, the operator received
a time penalty of 5 sec in order to discourage collisions with the
obstacle. About half of the operators had driven the crane in a previous study and therefore, had some level of experience in crane
operation.
For these operator learning studies, an SI shaper was designed to
suppress a range of expected frequencies around 0.21 Hz and 0.29 Hz.
These ranges were found by measuring the frequencies of the hook
and payload when the payload was near the floor and when it was
hoisted over the obstacle. The amplitudes and times of this SI shaper
are:

Obstacle Avoidance
Figure 10.12 shows the average completion time for each trial in the
obstacle avoidance tests. The bars for each trial represent the standard
deviation. The data clearly show that the completion times with input
shaping are substantially lower than the completion times without input shaping. The downward trends indicate that the operators were
learning to control the double-pendulum oscillations. However, the
downward trend in completion times is not as apparent for operation
with shaping enabled. This is because the operators could drive the
crane with shaping near its theoretical limit during their first trial.
The average number of button pushes for each trial is shown in Figure 10.13. For both shaped and unshaped operations, the number of
button pushes decreased as the operators learned to control the
double-pendulum crane. The button pushes for the unshaped case
levels off at about 30 pushes, while the SI shaped case levels off at 4

135/237

button pushes. This clearly shows that the amount of effort required of
the operator is much less when using the input shaper.

Figure 10.12: Obstacle Avoidance Average Completion

Time.

Figure 10.13: Obstacle Avoidance Average Effort.

136/237

Figure 10.14: Obstacle Avoidance Average Collisions.

The average number of collisions for each trial is shown in Figure
10.14. The collisions for the unshaped case decreases to 0.125 collisions per operator until the last trial where it spikes up about 0.8 collisions. This increase occurred because the operators knew it was their
last chance to obtain a fast completion time. Therefore, they drove
more aggressively. However, with shaping enabled, there were no collisions with the obstacle.
The completion times, operator effort, and collisions (Figures 10.12
through 10.14) all have decreasing trends as the number of trials increases. This indicates that the operators were learning how to control
the double-pendulum oscillations and completing the tasks faster, but
also using less effort. As the operators drove the crane more, they also
became safer at the task manipulation as measured by the decreasing
number of collisions. However, operation with input shaping enabled
was far superior to unshaped operation, even after substantial learning. With input shaping, there were no collisions, shorter completion
times, and less required effort.

137/237

Hoisting Over Obstacles

The average hoisting completion time with and without input shaping is shown in Figure 10.15. For both the shaped and unshaped operation, there is a slight downward trend in task completion time as the
number of trials increases. This indicates some small amount of operator learning. However, the hoisting completion times during Trial 1
with input shaping enabled are far shorter than after six trials without
input shaping.
Figure 10.16 shows the average number of button pushes for the
hoisting task. For both shaped and unshaped operations, the number
of button pushes decreases as the operators learned to control the
crane better. The button pushes for the unshaped case levels off at
around 15 pushes while the SI shaped case leveled off at 4.5 button
pushes. This again shows that the amount of effort required of the operator is much less when using input shaping.
The average number of collisions for each trial is shown in Figure
10.17 for the hoisting task. With input shaping, the number of collisions per operator varied from 0.5 collisions at the first trial to 0.125
collisions at the sixth trial. On the other hand, the number of collisions
per operator went from 0.25 collisions at the first trial down to zero at
the third when input shaping was enabled.

138/237

Figure 10.15: Hoisting Average Completion Time.

Figure 10.16: Hoisting Average Operator Effort.

Figure 10.17: Hoisting Average Collisions.

Given the results in Figures 10.15-10.17, it is apparent that learning
occurred without input shaping. However, the data also reveal that
there was substantial difficulty in controlling the double-pendulum oscillations even during the last trials. With input shaping enabled, the
operators could drive the crane very skillfully during their first trials,

139/237

and only minor learning occurred as they perfected their control

technique.

Chapter 11
Negative Input Shapers
In order to maximize throughput and productivity, machines must
be moved with as much force as their actuators are designed to produce. To generate shaped commands that drive actuators near their
limits, input shapers must contain negative impulses. This may seem
counterintuitive because negative impulses in an input shaper have
the effect of decreasing the command value - possibly even driving the
system in the opposite direction. However, using negative impulses allows larger positive impulses to be used. The net result is a shaped
command that is more aggressive than one generated by using only
positive impulses in the shaper.
To demonstrate this effect, consider the shaped commands shown
in Figure 11.1. The top of the figure shows a step input shaped with a
positive ZV shaper. The rise time of the command equals the duration
of the shaper, 1 = T/2. Figure 11.1b shows the step input being shaped
by a ZV input shaper that contains a small negative impulse. To compensate for the negative impulse, the two positive impulses must increase in amplitude so that the impulse amplitudes sum to 1. (Recall
the amplitude summation requirement in (5.5).) In order to cancel the
vibration of the first (positive) and second (negative) impulses, the
third (positive) impulse must be brought closer in time to the first impulse. The net result of these adjustments is that the duration, 2, of
the negative shaper shown in Figure 11.1b is less than that of the positive ZV shaper shown in Figure 11.1a. Consequently, the shaped command rise time is faster.
Figure 11.1c shows the step input being shaped with a ZV shaper
that contains a negative impulse that has the same magnitude as the

141/237

positive impulses. All three of the impulse magnitudes must equal one
(1, -1, 1) so that the amplitude sum equals one. This shaper results in
an even faster shaped command than the shaper shown in Figure
11.1b. Note that the shaped command has an initial value that is the
same as the original step input. As the magnitude of the negative impulse increases, the shaper duration decreases. However, Figure 11.1d
shows that using a negative impulse that is too large will cause the
magnitude of the shaped command to exceed the magnitude of the original command. This can lead to actuator saturation and poor vibration suppression.

142/237

Figure 11.1: Input Shaping with Negative Impulses.

143/237

11.1 Unity Magnitude Shapers

One simple and effective way to limit the magnitude of negative impulses is to switch the impulse amplitudes between 1 and -1 [117]:

These unity-magnitude (UM) shapers can be convolved with a wide

variety of unshaped inputs without causing actuator saturation. Figure
11.1c showed a step input shaped with a UM-ZV shaper. The duration
of the UM-ZV shaper, 3 = T/3, is only one-third of the vibration period, as compared to one-half period for the positive ZV shaper shown
in Figure 11.1a. The time saved by using the negative impulse is not
without penalty. Shapers containing negative impulses require larger
actuator effort. They also have a tendency to excite unmodeled high
modes [117]. This issue will be examined in Section 11.4, and solutions
to potential high-mode excitation problems will be presented in Section 11.5.
If zero vibration constraints are combined with the unity-magnitude
amplitude constraint, (11.1), then the impulse time locations of the resulting shaper are rather complex functions of . However, when = 0,
the problem simplifies, and we can derive an analytic solution for the
UM-ZV shaper. The minimum number of impulses required to satisfy
the UM amplitude and ZV vibration constraints is three, so the impulse amplitudes are:

By setting the residual vibration expression in (5.1) equal to zero, we

get two constraint equations because both the sine and cosine terms
must be zero. Using the amplitudes in (11.2), the two equations that
ensure zero vibration are:

144/237

Squaring (11.3) and using sin2 () = 1 cos2 (), we obtain:

Because t2 t3, we know from (11.5) that:

Using (11.6) and (11.4) to solve for t2 yields:

Similarly, using (11.6) and (11.4) to solve for t3 yields:

Therefore, the UM-ZV shaper for undamped systems can be described as:

For damped systems, an analytic solution of the impulse times for

UM-ZV shapers is not available. However, third-order polynomial

145/237

curve fits to solutions obtained with a numerical optimization program were generated. The curve fits to t2 and t3 are shown in Table
11.1 as a function of the damping ratio. The curve fits give time locations within 0.5% of their actual values over the range 0 0.3.
Table 11.1: UM-ZV Shaper for Damped Systems
Table 11.2: UM-ZVD Shapers for Damped Systems

11.2 Robust UM Shapers

In order to increase the robustness of UM shapers, the amplitude
constraints given in (11.1) can be combined with ZVD, EI, or SI constraints. To satisfy the ZVD or EI constraints, the shaper must contain
five impulses. Therefore, the amplitudes are:

In order to satisfy extra-robust requirements, such as two-hump EI

or large insensitivity SI constraints, the shaper will require 7, or possibly, 9 impulses.
The time location of each impulse scales linearly with the vibration
period being suppressed, but they are complex functions of . Curve
fits to the impulse times for the ZVD shapers are shown in Table 11.2.
Curve fits for EI, Two-Hump EI, and Three-Hump EI unity-magnitude
shapers are shown in Table 11.3. The impulse times for the EI shapers
are calculated for the case when Vtol = 5%. The time savings provided
by the negative impulses can be quickly assessed by examining the
dominant constant term (M0) for the final shaper impulse. The durations of the UM-ZVD and UM-EI shapers are approximately 73% of
the positive ZVD and EI shapers. The time duration of the UM-TwoHump EI shaper is about 79% of its positive-impulse counterpart.

146/237

11.3 Vector Diagrams of Negative

Shapers
Negative impulses in a shaper can be represented very easily on a
vector diagram. To take into account the negative value, the vector
representing the impulse is directed inward, toward the origin, rather
than outward. Figure 11.2a shows the vector diagram representation of
a UM-ZV shaper. By translating the third impulse vector to the end of
the first vector, as shown in Figure 11.2b, it is easy to see that the sum
of the three vectors is zero. This indicates that the shaper will produce
zero residual vibration.
Table 11.3: UM-EI Shapers for Damped Systems, Vtol = 5%

Figure 11.2: Vector Diagram Representation of a UM-ZV

Shaper.

11.4 High-Mode Excitation

Negative input shapers produce aggressive commands that drive a
system fast. As a result, they tend to excite high frequencies more than

147/237

positive shapers. To illustrate the effect of high-mode excitation, the

sensitivity curve for a shaper can be plotted over a range of high frequencies. The value of the curve at any high frequency indicates the
degree to which that frequency will be excited. Figure 11.3 compares
the sensitivity curves for the positive ZV and UM-ZV shapers for frequencies up to 15 times the modeling frequency for which the shaper is
designed. Note that the curve for the positive ZV shaper never exceeds
100%. This indicates that the shaper will never cause high frequencies
to be excited more than they are without input shaping. On the other
hand, the sensitivity curve for the UM-ZV shaper reaches a maximum
value of 300% at frequencies 3, 9, and 15 times greater than the modeling frequency. This means that if the system happens to have unmodeled modes at these frequencies, then they will be excited to 3
times their level without input shaping. Note that about half of the
high frequency range is excited to beyond 100% by the UM-ZV shaper.

Figure 11.3: High-Mode Sensitivity Curves

148/237

The amplitude of high-mode vibration is often small compared to

the amplitude of the lowest mode. For example, structural beams that
undergo transverse deflection have an infinite number of flexible
modes. However, the low mode dominates the amplitude of response.
Also, for a given damping ratio, amplitudes of higher frequencies decay more quickly than those of lower frequencies. Therefore, eliminating the low mode, even while exciting the higher modes, will often result in substantial overall vibration reduction.
Consider a beam that is rotated by applying a torque to one end.
When accelerated, the beam will deflect and start vibrating. When the
beam is decelerated to rest at the desired location, additional beam vibration will be caused. The net residual vibration contains an infinite
number of modes whose amplitudes depend on the beam properties,
as well as the acceleration and deceleration torque profiles. However,
the vast majority of the beam tip deflection will arise from the low
mode. To demonstrate this, a lumped mass was attached to the end of
a steel beam that was mounted to a rotary table [107]. The beam response to a bang-bang angular acceleration command is shown by the
dotted line in Figure 11.4. An FFT of the residual vibration showed a
dominant low mode at approximately 2 Hz, and a second mode close
to 8 Hz. The rest of the bending modes were undetectable in the
response.

149/237

Figure 11.4: Rotating Beam Experimental Results.

The 8 Hz mode was neglected and a Vtol = 5% One-Hump EI shaper
containing negative impulses was designed to rapidly rotate the beam.
The sensitivity curve for the shaper predicted that the 8 Hz mode
would be excited to approximately 10 times the level resulting from a
single step in acceleration. Figure 11.4 compares the shaped response
of the beam with the response to a bangbang command. The EI shaper
reduced the vibration substantially - down to about 16% of the bangbang vibration level. While the greatly excited 8Hz mode is visible in
the residual vibration, its amplitude is only about the same as the residual vibration of the dominant 2 Hz mode that remains after the input shaping is applied. The results demonstrate that, in some cases,
the amplitude of a highly excited secondary mode is relatively unimportant. On the other hand, recall that the double-pendulum crane in
Section 9.3 had significant amplitude in its second mode.

150/237

11.5 Controlling High-Mode Excitation

Although excitation of unmodeled high modes with negative input
shapers is often an inconsequential effect, for certain systems it can
limit performance. For high-mode excitation to occur, the system
must have a significant resonant mode at a frequency where the input
shaper sensitivity curve exceeds 100%.
In cases where high-mode excitation is performance limiting, there
are several ways to address the problem:
Give up the time savings gained by using a negative shaper and use a
positive shaper.
Add restrictions to limit vibration at the problematic high frequencies
and solve the augmented set of constraints to get a new shaper.
Add a digital low-pass filter.
Use a reference input with low high-frequency content.

Figure 11.5: Sensitivity Curves for UM-EI Shapers.

151/237

The first option is the easiest and most appropriate when the
throughput of the system is not a high priority. Option 2 is the highest
performance solution because an input shaper can be customized to a
specific system with a minimal time penalty. Unfortunately, the lookup method provided by the curve-fit tables must be abandoned. Options 3 and 4 can still utilize the tables, however, there can be large increases in the computational requirements during run-time and the
rise time will suffer.
Option 2, the process of restricting only a few problematic high frequencies, is best demonstrated by returning to the flexible beam example. The UM-EI shaper greatly reduced the 2 Hz mode. However,
assume that the 8 Hz mode causes the residual vibration to exceed the
tolerable amplitude threshold. This high-mode component of the vibration can be reduced by adding an equation to the set of inputshaper design constraints to limit the vibration at 8 Hz. This approach
is illustrated by the sensitivity curves in Figure 11.5. The sensitivity
curve fro the UM-EI shaper shows that the 2 Hz mode will be robustly
suppressed. However, the 8 Hz mode is amplified by the shaping action. By adding one constraint equation that limits the 8 Hz vibration
to a low level, a new shaper is obtained that has the sensitivity curve
shown by the dotted line in Figure 11.5. When the shaper with highmode limitation was implemented on the flexible beam, the 8 Hz
mode was virtually eliminated, as shown in Figure 11.6.
If more than one high mode is problematic, then simply add a constraint equation for each mode of vibration and solve the augmented
set of constraints. For each mode that is constrained, a small amount
of time will be added to the shaper duration. However, a negative
shaper with high-mode constraints will continue to be shorter than a
positive shaper for the low mode until constraints have been placed on
a large number of high modes.

152/237

Figure 11.6: Beam Response with 8Hz-Limiting UM-EI

Shaper.
The technique of restricting the vibration at only a few high modes
is advantageous because:
The shaper is customized to a specific system and, therefore, it does
not over-constrain the system response.
The computational requirements during run-time are only slightly
increased.
The drawback of this approach is that it requires the one-time solution of a set of simultaneous, transcendental constraint equations; the
curve-fit tables cannot be used.
Instead of restricting a few high modes, all high frequencies can be
reduced by adding a low-pass filter. The low-pass filter is used in conjunction with a negative shaper of the designers choice from the
curve-fit tables. A time delay is added because the rise time of the
command is increased by the duration of the input shaper plus the

153/237

duration of the low-pass filter. The computational requirements can

increase significantly because implementation of the modified
command-shaping algorithm requires N more multiplication and addition operations than the original shaper, where N is the low-pass filter length number.
The main benefits of augmenting negative input shapers with a lowpass filter are:
It requires little specific information about the problematic high frequencies; only a pass band and a stop band are needed.
It eliminates a large range of high frequencies.
It uses well established filter design tools in combination with the
solutions from the curve-fit tables.
This approach has two major drawbacks:
There is a longer rise-time delay associated with this process than
when we restrict just a few high modes.
The run-time computation can be prohibitive.

Figure 11.7: Schematic Diagram of Disk Drive Head Tester.

The fourth option for dealing with high-mode excitation is using an
input function that does not contain energy at high frequencies. For
example, smoothed commands such as those discussed in Chapter 4

154/237

could be utilized. This would have much the same effect as a low-pass
filter and may be simpler to implement. However, the system response
will be sluggish compared to when input-shaping is used. More intelligently designed smooth commands target only the problematic frequency and act as a notch filter [2, 19, 60, 61, 127, 145]. This allows the
system to move faster. However, the procedures for designing such
smooth commands are often computationally intensive. Furthermore,
it may be necessary to calculate a command for each unique motion
that the system must perform.

11.6 Case Study: Test Stand for Hard

Disk Drive Reading Heads
Millions of disk drives are manufactured every year. Each drive has
multiple heads that read and write the information on the disk platters. Before these reading heads are assembled into a disk drive, they
must be tested to ensure that they are functioning properly. One type
of head-testing machine is shown in Figure 11.7. During the testing
process, reading heads are placed in the machine and the XY stages
perform a gross motion to position the head near a calibrated hard
disk. The XY stages are mounted on air bearings to minimize friction.
Once the head has been positioned near the calibration disk, the head
is tested by having it read the information on the disk. The small
moves from track-to-track are accomplished by a spring-loaded piezo
actuator. When the piezo moves the reading head to a new track, the
force backdrives the air bearing stages and oscillations occur. The machine must wait for the oscillations to dissipate before information can
be read from the disk.
Using the standard controller, lightly damped oscillations at 75 and
1700 Hz result from typical moves, as shown by the solid line in Figure
11.8. Because the low mode dominates the dynamic response, a singlemode positive ZVD shaper was used to shape the command signal.
Figure 11.8 compares the shaped and unshaped responses for a 190

155/237

in move. The vertical scale for the shaped response on the right side
of the figure has been shifted upwards so that the residual vibration in
each case is clearly visible. Even though the shaped response has a
longer rise time, the operational speed of the machine is increased because much less time is spent waiting for vibration to dissipate.

Figure 11.8: Shaped and Unshaped Head Tester Responses.

156/237

Figure 11.9: Comparison of Positive ZVD and UM-EI

Responses.
The throughput can be increased even further by using negative
shapers. Figure 11.9 compares the response for a 110 in move when
positive ZVD and UM-EI shapers are used. The level of residual vibration is approximately the same, however, the negative shaper reaches
the desired location approximately 25% faster.

Table 11.1: UM-ZV Shaper for Damped Systems

Ai

ti

M0

M1

M2

M3

t1

-1

t2

0.16724

0.27242

0.20345

t3

0.33323

0.00533

0.17914

0.20125

Table 11.2: UM-ZVD Shapers for Damped Systems

Ai

ti

M0

M1

M2

M3

t1

-1

t2

0.08945

0.28411

0.23013

0.16401

t3

0.36613

-0.08833

0.24048

0.17001

-1

t4

0.64277

0.29103

0.23262

0.43784

t5

0.73228

0.00992

0.49385

0.38633

Table 11.3: UM-EI Shapers for Damped Systems, Vtol = 5%

Shaper

Ai

ti

M0

M1

M2

M3

UM-EI

1
-1
1
-1
1

t1
t2
t3
t4
t5

0
0.09374
0.36798
0.64256
0.73664

0
0.31903
-0.05894
0.28595
0.00162

0
0.13582
0.13641
0.26334
0.52749

0
0.6
0.6
0.2
0.1

TwoHump
UM-EI

1
-1
1
-1
1
-1
1

t1
t2
t3
t4
t5
t6
t7

0
0.05970
0.40067
0.59292
0.78516
1.12640
1.18640

0
0.31360
-0.08570
0.38625
-0.08828
0.20919
-0.02993

0
0.31759
0.14685
0.34296
0.54174
0.44217
0.79859

0
1.58
1.6
1.2
1.38
0.3
0.1

ThreeHump
UM-EI

1
-1
1
-1
1
-1
1
-1
1

t1
t2
t3
t4
t5
t6
t7
t8
t9

0
0.04275
0.42418
0.56353
0.83047
1.09760
1.23710
1.61890
1.66190

0
0.31845
-0.05725
0.48068
-0.09785
0.38825
-0.08706
0.09964
-0.09711

0
0.46272
0.04989
0.38047
0.34048
0.35290
0.81706
0.42780
0.80045

0
3.3
3.9
4.2
4.4
2.9
2.8
1.31
1.0

Chapter 12
Specified-Negative-Amplitude
and Duration Shapers
Given the tradeoffs that exist with negative input shapers, it is desirable to develop complimentary alternatives to unity magnitude (UM)
shapers. For example, it is useful to have an input shaper whose negative impulse amplitudes can be set to any value. This allows a designer
to precisely trade off rise time and high-mode excitation. Furthermore, it can also be useful to specify the duration of an input shaper.
This allows the designer to set the rise time to any desired value. This
chapter describes how to accomplish both of these goals using negative input shapers.

12.1 Specified-Negative-Amplitude
Shapers
A technique that specifies the amplitude of negative impulses can
produce shapers that span the performance gap between positive
shapers and UM shapers. Such shapers are shown schematically in
Figure 12.1. The key feature of specifiednegative- amplitude (SNA)
shapers is that the magnitude of the negative impulses, b, can be set to
any value. When 0 < b < 1, these shapers have longer durations than
UM shapers. If b > 1, then SNA shapers are faster than UM shapers,
but they will often cause actuator saturation. So the most useful range
is 0 < b < 1.
For undamped systems, the SNA-ZV shaper is given in closed form
by [104]:

161/237

Figure 12.1: Forms of SNA Shapers.

Figure 12.2: Duration of SNA-ZV Shaper.

where, is the modeling frequency and,

162/237

Note that b is a positive value that indicates the negative impulse

magnitude.
Robust SNA shapers must satisfy more design constraints. Therefore, they must contain additional impulses, as shown on the right of
Figure 12.1. In addition to specifying the amplitude of the negative impulses, the amplitude of the middle impulse, c, must also be specified.
The robust shapers cannot be determined in closed form, so a numerical optimization must be performed to minimize the shaper duration.
This is a relatively easy optimization that can be performed using
standard tools, such as the MATLAB Optimization Toolbox.

12.2 SNA Shaper Properties

It is important to understand how the shaper properties vary with
the magnitude of the impulses. It can be seen from examining t3 in
(12.1) that as b increases (the negative impulse becomes larger), the
duration of the SNA-ZV shaper decreases. Figure 12.2 shows the duration of the SNA-ZV shaper over a range of negative amplitudes and
damping ratios. When the negative impulse is nonexistent (b = 0), the
shaper is the same as the positive ZV shaper. When b = 1, the second
impulse is -1 and the shaper is equivalent to the UM-ZV shaper given
in (11.9).
Figure 12.3 shows the duration of the more robust SNA-ZVD shaper
as a function of the magnitude of the negative impulses and the
middle impulse for an undamped system. The ZVD shaper duration is
minimized at some finite value of the middle impulse (for a given
amplitude of the negative impulse). For a given value of b, the value of
c that minimizes the shaper duration is:

163/237

Figure 12.3: Duration of SNA-ZVD Shaper.

Figure 12.4: Impulse Times of SNA-ZVD (b = 1).

164/237

The middle impulse cannot be smaller than c 0.5. At this value, the
shaper is equivalent to the positive ZVD shaper. This result can be
seen in Figure 12.4, where the impulse time locations are plotted as a
function of the middle impulse amplitude when b = 1. The curve for t5
(the shaper duration) is the same curve shown in Figure 12.3 for the
case of b = 1. The two negative impulses located at t2 and t4 approach
the positive impulses at t1 = 0 and t5 as c decreases. The positive and
negative impulses eventually coincide when c = 0.5 and form the
three-impulse positive ZVD shaper. Also noted on the figure is the case
(c = 1) when the shaper is equivalent to the UM-ZVD shaper.

Figure 12.5: SNA-ZV Average High-Mode Excitation.

165/237

Figure 12.6: SNA-ZVD Average High-Mode Excitation.

SNA High-Mode Excitation

One way to characterize the effect on high modes is to determine the
average value of the sensitivity curve over a range of high frequencies.
In this chapter, average high-mode excitation (AHE) will refer to the
average value of the shapers sensitivity curve from 2 to 10 times the
modeling frequency. Figure 12.5 shows the AHE for the SNA-ZV
shaper as a function of the negative amplitude b and the system damping ratio, . As b gets larger, the high-mode excitation increases as
expected.
Figure 12.6 shows the high-mode excitation for the SNA-ZVD
shapers. As the middle amplitudes increase, the AHE tends to increase. However, it is clear there are some minima that would be
prudent choices when selecting the middle amplitude. So, rather than
simply using the middle impulse that will minimize the shaper duration, an engineer may want to use a slightly different value for c in order to decrease high-mode excitation.

166/237

Figure 12.7: Long Reach Manipulator with Suspended

Payload.

12.3 SNA Shaping on a Long-Reach

Robot
It is possible for an SNA shaper to have a much faster rise time than
a positive shaper and yet not significantly excite unmodeled high
modes. To experimentally verify this important result, tests were conducted on a large robot arm used to move a suspended payload, as illustrated in Figure 12.7. While the endpoint of the robot arm was
moved in the horizontal direction, the position of the payload was captured using a video camera. The UM-ZV, SNA-ZV (b = 0.5), and Positive ZV shapers were used to shape the robots reference commands.
The shapers were designed for the payloads pendulum frequency of
approximately 0.357 Hz. The impulse amplitudes and time locations
for the shapers are:

167/237

Note that the SNA shaper provides a 22% time savings and the UM
shaper provides a 33% time savings, relative to the positive ZV shaper.

Figure 12.8: Experimental Results from Long Reach

Manipulator.
The dominant flexible mode arising from the robot arm bending dynamics was approximately 1.8 Hz. This high frequency is

168/237

approximately five times the pendulum frequency. Given that its contribution to the overall vibratory response is likely to be small, this
high mode was not taken into account when designing the input
shapers in (12.4)-(12.6). The payload response using the three shapers
is shown in Figure 12.8. The unshaped response is not shown because
it would be off the scale (22 in. peak-to-peak). All three shapers eliminate most of the residual vibration of the pendulum mode.
The ZV shaper does not excite the higher frequencies of the robotic
arm and moves the hanging pendulum with a minimal amount of residual vibration. On the other hand, the UM excites the higher mode
of the robot and this shows up in the payload response. However, it
does produce the fastest rise time. The payload of the robot reaches
the desired location much faster with the SNA shaper than with the
positive ZV shaper, and the robot bending mode is not excited. These
results demonstrate that the flexibility of the SNA shaper provides a
means for achieving excellent performance that is not possible with
either the positive ZV or UM-ZV shapers.
Note that the second mode could also have been reduced by utilizing
the techniques in Chapter 9 to create a multi-mode shaper. For example, the UM-ZV shaper given in (12.4) could be convolved with the
positive ZV shaper designed to suppress the 1.8 Hz robot mode given
by:

The resulting shaper has a longer duration (0.9337+0.2778=1.2115

s) than the SNA shaper and it also has more impulses (6). To shorten
the shaper, one could consider convolving the UM-ZV shaper in (12.4)
with a UM-ZV shaper designed for the 1.8 Hz mode. However, when
two UM shapers are convolved, the resulting shaper can cause actuator saturation. Therefore, multi-mode negative shapers should not be

169/237

formed through convolution of single-mode shapers, rather they

should be designed using simultaneous solution of all constraints. This
procedure is covered in Chapter 13.

12.4 Specified-Duration UM Shapers

Up until this point, input shapers have been designed by satisfying
various combinations of residual vibration, robustness, and impulse
amplitude constraints while minimizing the shaper duration. When
the shapers are used to eliminate vibration, the rise time of the shaped
command increases by the duration of the shaper. If a human operator
requests the system to stop immediately, then the system decelerates
for a period equal to the shaper duration. Deceleration periods are
common on many mechanical systems, such as cranes, even without
input shaping. For example, crane motor velocities are often ramped
down so as to decrease residual sway and decrease mechanical wear.
Without shaping, these deceleration periods are constant and the operator becomes accustomed to the lag. Using previously discussed input shaping schemes, the deceleration period would vary with the system frequency and the operator may have to adjust to the varying deceleration period.
In order to create a control system that has a fixed deceleration
period, specified-duration (SD) shapers are developed in this section.
In this approach, vibration-limiting constraints are used in conjunction with the unity magnitude amplitude constraints and a specified
shaper duration:

With the shaper duration fixed at t, it no longer makes sense to

minimize the shaper duration. A reasonable alternative is to maximize
the robustness of the shaper. That is, maximize the range of frequencies over which the residual vibration is kept below Vtol.

170/237

To demonstrate the specified-duration (SD) shaper design process,

suppose that we are designing shapers for a crane with a maximum
cable length of 20 ft. The vibration constraint equations include a limitation at the low frequency, l, corresponding to a 20 ft. suspension
length. An additional vibration constraint is applied at a high frequency, h, to be determined. The frequencies between l and h are
suppressed by N additional vibration-limitation equations. An optimization is then performed to obtain the shaper that maximizes h.
Suppose that h, as determined by this method, corresponds to a
cable length of 8 ft., but that the system may operate with a cable
length shorter than 8 ft. That is, a single shaper cannot suppress vibration over the entire range of cable lengths. In this case, the optimization is performed again, but this time h plays the role of l. If the
new value, h2, resulting from the second optimization is above the
frequency corresponding to shortest suspension length the crane is
likely to use, then the shaper design process is complete. The shaper
resulting from the first optimization is used to move the crane when
the cable is between 20 ft. and 8 ft., while the second shaper is used
when the cable is less than 8 ft. in length.
The specified-duration shaper design algorithm can be summarized
as:
Specify the shaper duration, t.
Select a desired limit on the percentage residual vibration amplitude,
Vtol.
Require the vibration to be below Vtol at the low end of the possible
frequency range, l.
Perform an optimization that maximizes the frequency range over
which the residual amplitude can be kept below Vtol. The output of the
optimization is an input shaper and the maximum suppressed frequency, h.
If the frequency suppression range of the SD shaper covers the entire workspace, then the design problem is finished. Otherwise, start at

171/237

step 3 and replace l with h to get an additional shaper that covers

the next segment of the desired frequency suppression range. The
product of the algorithm is one or more specified-duration shapers
that can be used to suppress vibration throughout the workspace.

Specified-Duration Shaper Use on a Bridge

Crane
The above procedure for designing specified-duration shapers was
performed for a bridge crane at Savannah River Technology Center
(SRTC). A video camera was placed near the floor to record the residual oscillation of the payload. The damping ratio of the SRTC crane is
approximately 0.004 when the hook is near the floor. If the suspension cable length is very short (5 ft. or less), then the damping ratio
goes up considerably, to approximately 0.1. Because the damping was
so small throughout almost the entire workspace, the damping ratio
was approximated as zero for the input shaper design.
The approximate deceleration period for the existing crane control
system was 3 sec. So, t was set equal to 3.0 seconds and the above
procedure was used to design three UM-SD shapers to span the range
of cable lengths from 29 ft. to 9 ft. (The crane rarely operates with
cable lengths less than 9 ft. and cannot operate at less than 5 ft.) Figure 12.9 shows the measured percentage residual for each of the UMSD shapers. The percentage residual was obtained by dividing the
shaped residual amplitude by the unshaped residual amplitude for the
same motion. If the operator switches between the three shapers at 24
ft. and 19 ft., then the residual vibration can be kept to below 5% of the
unshaped level (for suspension cable lengths ranging from 9 to 29
feet).
Figure 12.9 compares the measured vibration resulting from the
UM-SD shapers to the theoretical residual amplitudes. Given the uncertainty in the video measurement system, there is good agreement
between theoretical and measured values. Notice that the largest deviation between theory and measured values occurs at short cable

172/237

lengths. This makes sense, as the measurement system has progressively worse resolution as cable length is decreased and the payload
moves away from the camera. Furthermore, at short suspension
lengths the actual damping ratio deviates more from the modeled
value of zero.
To compare SD shapers to earlier techniques, three UM-ZV shapers
were designed to suppress vibration throughout the crane workspace.
The shapers were designed to have zero residual vibration at cable
lengths of approximately 24 ft, 16 ft, and 11 ft. Figure 12.10 shows the
measured percentage residual vibration as a function of cable length.
If the operator switches between the three shapers at 20 ft and 13 ft,
then the residual vibration can be kept to below 20% of the unshaped
level (for cable lengths between 9 and 29 feet).

Figure 12.9: Comparison of Theoretical and Measured UMSD Residual Amplitude.

173/237

Figure 12.10: Measured Residual Amplitude Using UM-ZV

Shapers.
Comparing Figures 12.9 and 12.10, we see that the UM-SD shapers
provide much better sway reduction than the UM-ZV shapers
throughout the workspace. Unlike the specified-duration shapers, the
deceleration time is different with each UM-ZV shaper. The operator
must become accustomed to the different deceleration periods. An advantage of the UM-ZV shapers is that the deceleration period is shorter. The maximum shaper duration is 1.8 seconds, as compared with
3.0 seconds for all of the UM-SD shapers shown in Figure 12.9.

Chapter 13
Multi-Mode Negative Shapers
Chapter 9 presented two approaches for creating input shapers to
suppress multiple modes. One method created shapers for each individual mode and then convolved them. Unfortunately, this approach
cannot be used reliably to combine two or more negative shapers. The
resulting shaper will cancel the desired modes, but it may cause actuator saturation.
Recall that single-mode negative shapers have impulse amplitudes
that alternate between positive and negative values. This ensures that
the shaped command does not exceed its limit. A problem can occur
when a convolved shaper formed from multiple negative shapers contains two or more impulses in sequence that have large amplitudes of
the same sign. This can produce a shaped command that exceeds the
value of the unshaped command. Such a large-valued command could
cause actuator saturation. This effect is illustrated in Figure 13.1 by a
convolution of UM and SNA shapers. The two-mode negative shaper
has a sequence of three positive impulse [a, b, a]. When convolved
with a step input, the resulting shaped command will have a larger
maximum amplitude than the original unshaped step command, as
shown in Figure 13.2.
The brief periods when the shaped command requests more actuator effect than is available can be tolerated on many systems. However,
it will lead to some degradation of the shaping vibration reduction.
Furthermore, this overcurrenting may damage the actuator, so it is
usually prudent to avoid convolving two negative shapers together.
Note that a negative shaper can be convolved with a positive shaper to
produce a multi-mode shaper that works much more reliably;

175/237

however, the shaper will have a longer duration that one formed from
two negative shapers.

Figure 13.1: Convolving Two Negative Input Shapers.

Figure 13.2: A Convolved Negative Input Shaper Can Amplify the Peak Command Amplitude.
In order to produce a reliable multi-mode negative shaper, we must
use the simultaneous solution technique presented in Chapter 9. That
is, we form a set of constraint equations that include negative amplitude constraints and vibration limitations on all of the frequencies

176/237

targeted for suppression. Then, the complete set of constraints is satisfied simultaneously while minimizing the shaper duration. Recall that
this process was used in Section 11.4 to limit a problematic high frequency. In the general case, multiple frequency ranges can be suppressed by simultaneously forming specified insensitivity (SI) constraints over each of the desired suppression ranges and combining
them with negative amplitude constraints, such as the unity magnitude (UM) or specified-negative amplitude (SNA) constraints.

13.1 Case Study: CMM Deflection

Recall the moving bridge coordinate measuring machine (CMM)
that was shown in Figure 1.10. The CMM consists of a workspace in
which parts are fixtured, a sensor for detecting the part surfaces, a
mechanical assembly for moving the part sensor around the workspace, and a computer for calculating the part dimensions based on
the sensor measurements. CMMs are available in numerous sizes and
styles. Some are desktop mechanisms moved manually, while others
are computer driven and are large enough to measure car bodies.
The CMM of Figure 1.10 is shown with a touch-trigger probe part
sensor. This sensor uses a ruby-tipped stylus to sense the part. When
the stylus is brought into contact with a part surface, the deflection of
the stylus triggers the computer to record the position indicated by the
x, y, and z encoders. By probing the part on appropriate surfaces and
recording their locations, the dimensions of the part can be calculated.
However, if the CMM structure deflects between the encoder and the
ruby tip, then the measurement will be inaccurate.
For the CMM under consideration, the measurement cycle consists
of four phases. First, the CMM performs a gross motion to move the
probe to the vicinity of the part feature that is to be measured. Second,
the probe is allowed to come to rest. Third, the probe is reaccelerated
to a small constant velocity in the direction of the part. This portion of
the measurement is called the probe approach. Finally, the probe

177/237

contacts the part and the computer records the location of the contact.
The position of a touch-trigger probe during a typical measurement is
shown in Figure 13.3.

Figure 13.3: CMM Measurement Process.

If a CMM is to provide useful information, then its accuracy and repeatability must be greater than the tolerance specifications for the
part. The most important limitation on CMM performance depends on
the design of the CMM and the operating conditions; however, structural deflection between the encoders and touch-trigger probe is always a limitation because it introduces an error in the measurement.
Figure 13.4 graphically demonstrates how deflection in a CMM
structure can adversely effect measurement quality. On the left side of
Figure 13.4 the part width is determined from two measurements.
First, the probe is moved into contact with the left side of the part. At
the moment the contact is made, the position indicated by the encoders is recorded. Next, the probe is moved to the opposite side of the

178/237

part and brought into contact with the right face. The two encoder positions are then subtracted to obtain the part dimension. In the case
shown on the left side of the figure the part is measured accurately because the encoders indicate the true position of the probe. For the case
shown on the right side of Figure 13.4, the measurement is inaccurate
because the structural deflection makes the encoder readings differ
from the true location of the contact points. If the structure is vibrating with an amplitude of during the probe approach (3), then the
calculated dimension can have an error of 2.
Recall that the structural deflection of a CMM was shown in Figure
1.11. By performing an FFT on the deflection response during the approach region, a lightly-damped dominant mode was identified at 8
Hz, and some lightly damped
secondary modes occurred in the 14-18 Hz range. The low mode is
clearly dominant, but the higher modes are not negligible. Therefore, a
two-mode UM shaper was designed by using ZVD constraints for 8 Hz
and 16 Hz. The robustness of the ZVD constraints provided satisfactory suppression over the 14-18 Hz range.

Figure 13.4: Effect of Structural Deflection.

179/237

Figure 13.5: Comparison of Deflections During Shaped and

Unshaped Measurement Cycles.
When the input shaper is applied to the reference command of the
CMM, the deflection of the structure is reduced, both during the gross
motion and, most importantly, during the slow approach toward the
part. This effect is shown in Figure 13.5. The input shaping reduced
the measurement probe deflection from approximately 25 m to about
6 m during the critical measurement phase.

180/237

Figure 13.6: Measurements During a 50-Point Repeatability Test.

Repeatability Tests
The process of input shaping not only improves the measurement
accuracy, it also greatly improves the repeatability of the measurements. Repeatability is a measure of the ability of the instrument to
produce the same indication when sequentially sensing the same
quantity under similar measurement conditions [67]. A repeatability
test for a CMM is conducted by repeatedly measuring the location of a
fixtured part. After each measurement cycle, the probe returns to its
starting position. The minimum measured value is subtracted from
the maximum measurement to obtain the repeatability of the CMM.
Most CMMs have a repeatability of just a few microns.
Figure 13.6 shows 50 individual measurements of a repeatability
test using a 1 mm probe approach distance and a 20 mm gross motion.
The repeatability range for the unshaped measurements is: 25.2735
mm - 25.2687 mm, which equals 4.8 m. The mean value and standard deviation of the unshaped measurements are 25.2711 mm and

181/237

0.00123, respectively. Figure 13.6 also shows 50 measurements obtained with input shaping enabled. For the shaped measurements the
range is 3.3 m, the mean is 25.2670 mm, and the standard deviation
is 0.00067. For this set of measurement parameters input shaping improves the repeatability by 1.5 m and the standard deviation by 45%.
The improved repeatability with input shaping is a straightforward
result given the deflection curves of Figure 13.5. The curves show that
the unshaped measurements have considerably more variation in the
deflection during the probe approach region. During the course of the
repeatability test, the part is contacted at slightly different locations on
the deflection curve. The decrease in deflection variability with input
shaping translates directly into a decrease in measurement variability
(an improvement in repeatability).
To ensure that the benefit of input shaping is not limited to this particular set of parameters, repeatability tests were conducted with several probe approach and gross motion distances [116]. Figure 13.7
shows the repeatability as a function of probe approach distance for
measurements with a 20 mm gross motion. Each data point in Figure
13.7 represents the range of measurements obtained from a 50-point
repeatability test with the given probe approach distance and a 20 mm
gross motion. In general, repeatability improves with increasing probe
approach distance. Figure 13.7 also demonstrates that input shaping
improves repeatability over a wide range of probe approach distances.

182/237

Figure 13.7: Shaped and Unshaped Repeatability for Several Probe Approach Distances.

Chapter 14
Designing Input Shapers and
PD Feedback Concurrently
Input shaping can be used in conjunction with any type of feedback
controller. A block diagram for the case with an input shaper and a
Proportional-Derivative (PD) feedback controller is shown in Figure
14.1. A standard approach to using both input shaping and feedback
controllers is to first calculate the feedback control gains and then
design an input shaper for the resulting frequencies of the overall
closed-loop system. However, this procedure fails to maximize the system performance. Improved performance can be achieved if the feedback gains and input shaper are determined concurrently.

14.1 Command-Enhanced Feedback

The approach used to simultaneously determine the input shaper
and feedback gains is called Command-Enhanced Feedback (CEF)
design [34]. In this section, the position control of a single mass is
analyzed in order to demonstrate the CEF method. The mass position
is maintained with PD control, and friction to ground is neglected. It is
assumed that the command Yd in Figure 14.1 will consist of step
changes in position of varying magnitude.
The plant transfer function is:

184/237

where m is the mass under consideration.

Figure 14.1: Input Shaper and PD Feedback Controller.

The closed-loop transfer function is:

where KP and KD are the proportional and derivative gains. Note

that the closed-loop transfer function is equivalent to that of a massspring-damper system with position input.
Design of a CEF controller requires formulation of constraint equations for use in an optimization routine. The optimization solves for
the feedback gains and the input shaper impulse amplitudes and time
locations. Three constraints are used to limit the maximum values of:
Percent Overshoot, Actuator Effort, and Residual Vibration.

Constraints on Percent Overshoot

The response of the mass to a step input of magnitude L introduced
at time t0 can be expressed as:

185/237

To determine the peak overshoot, (14.3) is first converted to the

time domain:

where,

Now, the response to the sequence of steps that results from convolving a step function with an input shaper can be determined. The
amplitude of step Lj in the step sequence can be expressed as:

where Aj is the amplitude of impulse j in the input shaper, and L is

the amplitude of the unshaped step input.
The response of the system after step k is:

186/237

where tj is the time location of impulse j, and k n, the number of

impulses in the input shaper. Using trigonometric identities, the response in (14.8), for time greater than tk, can be written as:

where,

and

The maximum value of the response given in (14.9) occurs at a time:

The peak response, Yp, can be found by evaluating (14.9) at time tp:

187/237

Now that the peak response has been determined, the peak percent
overshoot can be expressed as:

Note that the peak percent overshoot depends not on the step size,
but rather on the impulse amplitudes and time locations in the input
shaper. Consequently, the overshoot can be constrained with the input
shaper.
To ensure that the tolerable overshoot value, Mp,tol, is never exceeded by the shaped signal, (14.14) is constrained below this tolerable
value after the step corresponding to each impulse. In other words, k
is increased from 1 to n in (14.14), and for each value of k the following
constraint is enforced:

Constraints on Actuator Effort

The expression relating the actuator effort, U, to the reference command, R, is:

Assuming that R is a step change in position of magnitude L introduced at time t0, we can rewrite (14.16) as:

188/237

If the actuator effort is to be limited, then the magnitude of U in the

time domain must be limited. Converting U to the time domain gives:

where d and a are given in (14.6), and

The first term in (14.18) represents an impulse of magnitude LKD.

Because this term only affects the magnitude of u(t) at a single instant
in time, its effect will be insignificant for any real actuator. The effect
of the impulse on the input magnitude is therefore neglected, and
(14.18) is rewritten as:

which is valid for time greater than t0, the time at which the step input is introduced.
Again using trigonometric identities, the plant input resulting from
the shaped step command can be expressed as:

where,

189/237

and Atot and are defined in (14.10). But, the value of in (14.10)
is calculated using the phase shift from (14.20), rather than ? from
(14.5).
To ensure that u(t) does not exceed the maximum allowable actuator effort, Utol, the following constraint is enforced after each step in
the command:

Constraints on Residual Vibration Amplitude

Using a process similar to the previous sections, the equation for the
residual vibration due to a sequence of impulses is:

where,

Atot and y are given by (14.10) and (14.11), again using ? from
(14.20), and

In order to limit the residual vibration, the following constraint is

enforced:

190/237

To summarize the Command Enhanced Feedback design process,

values for the feedback gains KP , KD, and the input shaper impulse
amplitudes and time locations are determined such that (14.15),
(14.24), and (14.28) are satisfied while minimizing the time of the final
shaper impulse. This is a fairly simple optimization problem; the
MATLAB optimization toolbox was used to obtain the solutions discussed in the following sections. Finally, if robustness to parameter
variations is desired, then multiple versions of the constraints can be
applied over a range of parameters, as was shown in previous
chapters.
The CEF design process generates the control system parameters so
that the performance constraints are satisfied for a given move distance L. If L is set equal to the largest expected move distance, then
the resulting controller parameters will yield good performance for all
expected move distances. Therefore, the control system designer
should not need to recalculate the controller parameters for different
move distances.

14.2 Comparison of CEF and PD

Control
To demonstrate the effectiveness of the CEF method, it is compared
to standard PD control without input shaping. Each controller was designed for a system having a nominal mass of 1 kg, maximum actuator
effort of 200 N, and a maximum expected step size of 1 m. To aid in
the controller evaluation, the following definition is used:

where Utol is the actuator limit, m is the nominal mass of the system, and L is the magnitude of the step change in position. Physically,
the ratio /L provides a quantitative measure of the ability of the

191/237

system to respond to a step input. In general, a system with a large ratio of /L will respond faster than a system with a small value of /L.
The parameters for the example system yield a ratio value of 200.
In order to make a reasonable comparison, both control schemes
are required to have a time constant of 0.116s. The time constant of
the system, defined as ( n)1, gives an indication of the systems
ability to reject disturbances. Table 14.1 summarizes the important
parameters of the two control methods.
The input shaper to be used in conjunction with the gains listed in
Table 14.1 for the CEF method is given by:

The input shaper defined by (14.30) was designed without robustness to mass variations; however, any number of robustness methods
could be utilized in the CEF control approach [94, 114, 115].
Three criteria are used to compare the effectiveness of the CEF approach to PD control:
Overshoot
Settling Time
Sensitivity to Disturbances

Overshoot
In order to evaluate overshoot properties, (14.2) is solved for Y when
the system is subjected to a step input of magnitude L:

where in this case,

192/237

An expression for the percent overshoot of the system as a function

of the damping ratio, , can be obtained after (14.31) is converted to
the time domain. This expression is given by:

Table 14.1: Parameters of PD and CEF Control Methods.

Figure 14.2: Percent Overshoot with PD Control.

where,

193/237

The percent overshoot for the system under PD control versus

damping ratio is shown in Figure 14.2. The graph shows that, even
with a damping ratio of 1, the best achievable overshoot is 13.5%.
Recall that the overshoot can be arbitrarily specified with the CEF
method. For the purposes of this comparison, a peak overshoot of 5%
percent was chosen for the CEF controller. A larger overshoot would
result in a faster rise time, but accurate positioning and a fast settling
time were assumed to be very important.

Settling Time
For this comparison, the system is defined to be settled when the residual vibration decays below 5% of the step magnitude. The fastest
settling time that can be attained with only PD control is plotted as a
function of damping ratio in Figure 14.3 for an /L value of 200. The
following equation was used to generate the curve in Figure 14.3:

The fastest settling time with PD control alone is 0.395s. This settling time requires a damping ratio of 0.75, which will yield an overshoot of 19.4%.
The final impulse time defines an upper bound on the settling time
under CEF control because the input shaper forces the residual vibration to below 5% at the time of the final impulse.1 Figure 14.3 shows
the best achievable settling time with CEF control as a function of
damping ratio for /L = 200. The minimum settling time with CEF
control is 0.153s, about 2.6 times faster than with PD control alone.

194/237

Although Figure 14.3 shows only the case for /L = 200, the same effects occur for other values of /L.

Figure 14.3: Best Achievable Settling Time with PD and CEF

Control (/L=200).

Sensitivity
Disturbances that directly affect the plant input are shown as D in
Figure 14.1. Measurement noise, or other similar error sources, could
also be introduced to the plant response, and they are represented by
W in Figure 14.1. The sensitivity of a system to these disturbances is
another good indicator of system performance. Low sensitivity values
are desired, as this indicates that disturbances will have little effect on
both the output of the system and the magnitude of the error signal (E
in Figure 14.1). The sensitivity of the response to disturbances of type
D is:

195/237

This transfer function for both control methods is shown in Figure

14.4. The CEF control method is less sensitive to low frequency disturbances and exhibits approximately the same sensitivity to high
frequencies.
The sensitivity transfer function relating disturbances of type W to
system response is equivalent to the closed-loop transfer function
defined by (14.2). This closed-loop transfer function is shown in Figure 14.5. Both control methods have similar performance at high frequencies, which are usually of greatest concern with disturbances of
type W. At low frequencies, the CEF controller produces a curve with a
slightly larger bandwidth.
Another measure of system sensitivity relates the magnitude of the
error signal, E, to the reference command, R. This sensitivity should
be very low, as a small error signal typically allows for larger gains and
improved performance. Furthermore, the error sensitivity is an indicator of the general robustness of the closed-loop system. The error
sensitivity transfer function is:

Figure 14.4: Effect of D(s) on Response (Y (s)/D(s)).

196/237

Figure 14.5: Effect of W(s) on Response (Y (s)/W(s)).

Figure 14.6 indicates that the CEF control system will result in error
signals of smaller magnitude than the PD controller for low frequency
inputs, while at high frequencies the two controllers produce error signals of similar magnitude.
This section has compared CEF control to optimized PD control
without input shaping. The reader should note that the controller with
input shaping is much better in nearly all respects. Another comparison can be made between CEF control and the optimized PD control
with input shaping added after the PD gains have been chosen. PD
control with input shaping achieves performance levels closer to the
CEF control; however, the higher proportional gain of the CEF control
allows it to move the system faster. This is another key result - concurrently designing the feedback gains and input shaper, (CEF

197/237

algorithm), provides better performance than the traditional method

of designing the feedback controller first and then designing the input
shaper based on the closed-loop performance.

Figure 14.6: System Sensitivity (E(s)/R(s)).

Time response Simulations

The response of the mass to a 1 m step input under both types of
control is shown in Figure 14.7, along with the reference commands.
The benefits of using the CEF design method, in terms of overshoot
and settling time, are apparent. Because the maximum overshoot for
the CEF method is set to 5%, the settling time of the system under CEF
control is equal to the rise time, approximately 0.153s.
In addition to the response of the closed-loop system to reference
commands, the disturbance rejection properties of each control
scheme also provide a good basis for comparison. The effect of an impulse disturbance of type D introduced at 1s is shown in Figure 14.8.
Because both controllers have the same time constant, they are equally
capable of rejecting type D disturbances. However, the higher proportional gain of the CEF control allows it to start suppressing the

198/237

disturbance sooner than the PD control. Therefore, although settling

time from a disturbance is the same, the CEF control results in a smaller excursion when a disturbance occurs.
Type W disturbances will affect the response of the mass in the
manner shown by Figure 14.9. The W disturbance used to generate
Figure 14.9 was a sinusoidal input of amplitude 0.05 and frequency
115 rad/s. As shown in Figure 14.5, the steady-state responses of the
two closed-loop systems to the W disturbance are identical at a frequency of 115 rad/s.

Figure 14.7: Mass Position vs. Time with PD Feedback.

199/237

Figure 14.8: Response to Disturbance of Type D(s).

Notice from Table 14.1 that the process solves for the proportional
and derivative gains divided by the nominal system mass. This implies
that, if the nominal value of mass varies from the expected value, then
the new values of KP and KD can be calculated without repeating the
optimization process. Also, if the nominal mass value varies, then the
maximum step size, L, should be changed accordingly to achieve the
same value of /L used in the design. If these two steps are taken, then
the same control settings can be used under a wide range of different
operating conditions.
The results shown in this chapter clearly demonstrate that PD control with input shaping outperforms PD control alone. However, there
is still the question of how CEF control compares to the sequential
design process where the feedback gains are chosen first and then the
input shaper is designed to accommodate the resulting dynamics. The
relative improvement with the concurrent design approach cannot be
stated definitively because the answer depends on how the PD gains
are chosen in the sequential input shaping approach. If, by some
chance, the PD gains were chosen to be the exact values determined

200/237

through the CEF approach, then the sequential method could yield
equivalent performance (provided that the optimal shaper was then
designed in the subsequent step). However, the sequential approach
does not have any mechanism for finding the optimal gain values, other than through a trial-and-error process that includes designing the
complimentary input shaper and evaluating the resulting
performance.

Figure 14.9: Response to Disturbance of Type W(s) .

The sequential approach would usually choose the gains so that the
unshaped response would be fairly well damped. In this case, the proportional gain would be smaller or the derivative gain would be larger
than in the CEF case, and the CEF controller would produce a faster
response. Therefore, the traditional sequential input shaping design
approach would, at best, equal the CEF control, but for practical reasons it is unlikely to perform as well.

201/237

The final impulse time is an upper bound on settling time because

the response can enter the settling region, and remain within it, before
the time of the final impulse. Such a case will be shown in Figure 14.7.

Table 14.1: Parameters of PD and CEF Control Methods.

PD

CEF

KP/m

132.2

321.5

KD/m

17.24

17.24

0.75

0.48

, sec.

0.116

0.116

Overshoot, %

19.4

ts, sec.

0.395

0.209

Chapter 15
Historical Overview of
Command Shaping
15.1 Early Work
Over 70 years ago command shaping appeared in the literature discussing cam profile design [45, 89, 130, 145] and command smoothing
[2, 19, 59, 60, 127, 131]. The early command-shaping methods did not
target specific system frequencies. Rather, they limited input energy to
a range of frequencies below the resonant frequencies of the system
being controlled. As this work evolved, various authors presented
methods for constructing command functions, typically trigonometric
or polynomial functions, that caused limited residual vibration [2]. To
minimize move and settling time, step and pulse command functions
were approximated as sums of sinusoids that were carefully selected to
minimize energy at the system resonant frequencies [60].
Systematic research on cam profile design can be found as far back
as the 1930s [130] and significant advancements in vibration reduction for cam systems occurred in the 1960s and 70s [45, 144, 145]. As
digital computers became more powerful, specification of cam profiles
came to be seen as an optimization problem. It is generally assumed
that these cam designs are appropriate for use in those systems which
are intended to operate at a constant or nearly constant speed [145].
Away from the design speed, system performance is often degraded
significantly.
An early method for constructing command waveforms for a specific
system model was Signal Component Control [23]. The objective of

204/237

this method was to construct a useful command signal that would

minimize the error between the command and the system response. In
1965, an interesting study of oscillation control was conducted for
large dock cranes [1]. Working in the phase plane, the authors derived
a command trajectory for limiting the cranes bucket or grab sway,
while moving it rapidly from the ship to a storage bin on land. Now,
more than 40 years later, cranes are just beginning to make use of
command-shaping methods in significant numbers [28, 31, 41, 122,
123].

15.2 Posicast Control

The first rigorous developments of command-shaping methods appeared in the 1950s. At that time, digital computers were a rarity and
implementing commandshaping methods with analog computers was
very challenging. Some of the earliest work on systematic command
shaping was performed by OJM Smith [120, 121, 128]. Smith was the
first to provide a theoretical description of command shaping that
could be easily followed. His method, which he named Posicast control, applies a baseline reference command and a delayed copy of the
command [68]. The delayed portion of the command cancels the vibration induced by the portion of the reference command that is not
delayed.
During the 1960s, 70s, and 80s, several researchers continued to
pursue work on Posicast control [16, 62, 68, 90, 124, 125]. One of the
main challenges was simply to find ways to implement the method on
actual machines. To quote one author [62]:
The principle of posicast control of resonant systems is
so elegant that more use should be made of it. This paper
presents practical feedback realisations of this control
idea...

205/237

Generating the time-delayed portion of the command was particularly difficult. Researchers tried to produce the delayed components
with mechanical delay elements, transmission delay lines, and analog
circuitry [90, 128]. One method determined a Fourier series expansion
of the command, truncated the series, and then reconstituted the comparable command [90]. Another approach changed the system set
point in predetermined ways when the system passed through a point
of zero velocity [62].

15.3 Robustness
In the 1980s, use of microprocessors became widespread. This enabled practical development and use of more sophisticated commandshaping methods. A key element of these methods was robustness to
modeling errors, uncertainties, and nonlinearities. Command-shaping
methods require some information about the system dynamics. This
information will always have some degree of inaccuracy. So, for command shaping to be successful on real applications, it must have an
adequate level of robustness.
In 1990, Singer and Seering presented a command-shaping method
that produced commands that limited residual vibration, even in the
presence of significant modeling errors [93, 94]. The method achieved
robust performance by limiting residual vibration at frequencies near
the modeled system frequency. The constraint was implemented by
first constructing an equation that describes the amount of residual vibration as a function of a series of impulses. The derivative of this
equation with respect to frequency was then set to zero at the modeled
value of the systems natural frequency. The effect of this constraint is
to keep the vibration at a low level, even when the actual frequency deviates somewhat from the modeled value.
The idea of robust command shaping was a significant step forward,
as it greatly expanded the design possibilities, as well as the applications for command shaping. In a very short time, several other

206/237

research groups adopted the idea and were making extensions and experimental verifications of Singer and Seerings input-shaping method
[8, 36, 83, 137, 146].
Even though it was not robust, the early work in command shaping
sought to understand the operating principles so that they could be
better exploited [8, 9]. In his book [120], OJM Smith demonstrated
that the posicast method placed zeros over the complex poles representing a systems natural frequency, thereby canceling their vibratory
effects. Singer and Seerings derivative constraints place additional
zeros over the flexible poles of the plant [74, 100]. This concept of
adding additional zeros at, or near, the flexible poles would prove to be
a powerful concept for designing command-shaping methods.
Soon after the development of the zero vibration and derivative
(ZVD) input shaper [94], many other researchers sought to extend the
robustness idea. The extensions can largely be categorized as: 1) Builtin robustness, or 2) Adaptive robustness. Built-in robustness seeks to
make the input shaper inherently robust, as for example the ZVD
shaper. A fundamental tradeoff in this approach is that obtaining
more robustness can lead to an increase in shaper duration. A key
challenge is to obtain significant robustness while limiting the shaperduration penalty.
To address this challenge, another approach was proposed in 1990,
wherein the constraint of zero vibration at the modeling frequency was
replaced with a constraint that merely limited the vibration to a small
value [41, 77, 102, 112, 114]. This approach evolved from the understanding that vibration can never be completely eliminated. Furthermore, methods that attempt to achieve a perfectly vibration-free response usually pay a significant penalty for the last few percent of vibration reduction. This more practically-based method was called the
Extra-Insensitive (EI) approach because it provides extra robustness
without increasing the shaper duration - the EI shaper has the same
duration as the ZVD shaper.

207/237

Specified-Insensitivity (SI) input shaping was then developed to

suppress any desired range of frequencies [91, 109, 115]. This is an extension of the extra-insensitive shaper for which the concept of allowing a tolerable level of vibration was extended for an arbitrarily large
range of frequency suppression. The SI design method can be augmented by weighting the importance of the frequencies within the suppression range [44, 57, 72]. For example, the lower frequencies within the
suppression zone could be suppressed more, because they are likely to
produce larger amplitude residual vibration than higher frequencies.
Rather than construct an input shaper that has inherent robustness
properties, many researchers developed methods to use a non-robust
shaper, but adapt its impulse amplitudes and time locations to the
changing dynamic properties of the system. This approach can provide
a faster rise time because the nonrobust shaper is shorter in duration
than a comparable robust shaper. However, it requires additional development complexity, time, and expense.
Adaptive input shaping uses feedback measurements of the system
states to continually change the input shaper to improve its effectiveness. For example, the ZV shaper can be continually changed during
operation by updating the frequency that is used to calculate the
shaper impulses. A significant cost of adaptive input shaping is that
sensors must be added to the control system. Furthermore, these
sensors must give indications of how the oscillation frequencies
change. Key challenges in adaptive shaping are updating the shaper
impulses rapidly and achieving stable behavior.
One of the earliest adaptive input-shaping methods used a
frequency-domain identification scheme to estimate the vibration frequencies and then update the spacing between the shaper impulses
[135137]. The challenge with this approach is to perform the identification in real-time without placing too large of a computational burden on the control computer. Numerous experimental results were obtained to demonstrate the effectiveness of this method [36, 38, 146].
The approach can be modified to use other types of frequency

208/237

identification methods, for example the Empirical Transfer Function

Estimate approach [37].
Given the time delay required to calculate the dominant mode from
sensor data, one adaptive approach measured the second vibration
mode and then used a known relationship between the first and
second modes to calculate the primary mode [42]. This method can be
advantageous because the second mode can be calculated from a
shorter set of sensor measurements; therefore, it can be obtained
faster than the first mode. The system identification necessary for effective adaptive command shaping can also be accomplished in the
time domain [13, 76], or with algebraic non-asymptotic identification
[80].
In the indirect adaptive approaches discussed above, the system
parameters are identified first and then the input shaper is designed.
Direct adaptation algorithms can generate input shapers without explicitly calculating the system frequency or damping ratio [75, 85, 86].
In many cases, this approach can have better convergence characteristics than indirect approaches. Given the importance of the adaptation algorithm to the success of adaptive command shaping controllers, it is prudent to compare the various methods and match their
capabilities with the intended application [12, 13, 54, 86].
In addition to the real-time computational burden, another significant challenge for some applications is the effect of measurement noise
and random disturbances. Noise and disturbances can erroneously indicate that the dynamic properties are changing. This could lead to an
incorrect change in the input shaper. This issue was studied and a
method to optimize solutions for systems with noise was developed
[84].
If systems have significant residual vibration amplitude at more
than one resonant frequency, then the shaping needs to suppress multiple modes [30, 106]. Much of the work done on multi-mode command shaping has been directed at specific applications because the
number of modes and their relative amplitudes and frequency ratios

209/237

can significantly affect the design procedure that should be used.

Papers have been published on multi-mode command shaping for specific applications such as robots [25, 26, 5456, 146], spacecraft [46,
29, 88, 96, 126, 132, 134] and cranes [27, 41, 57, 109, 122, 140].
Some methods for multi-mode command shaping have sought to
optimize performance by satisfying auxiliary constraints to improve
certain aspects of the system performance. For example, an approach
was developed to eliminate multiple modes of residual vibration while
using a minimum number of impulses in the input shaper [99]. Using
only a small number of impulses decreases the computational requirements during real-time implementation. Methods have also been developed to optimize the shaping process when multiple actuators are
used to drive the system [49, 50, 71, 139]. Additional methods have
been developed to optimally design command shaping to work in conjunction with feedback control [34, 64] and damping elements [21,
52].

15.4 Input Shaping for Time-Optimal

Control
Time-optimal control for flexible systems is a special type of command shaping that creates commands that move a system as fast as
possible from one state to another. This endeavor dates back to the
1960s in the work of Pontryagin [81], but many other researchers followed on with contributions to this field [3, 7, 15, 24]. The approach
requires a good model of the system dynamics, as well as knowledge of
the actuator limits. The commands are then designed to use the maximum actuator effort to move the system as fast as possible.
While the mathematical constructs for analyzing time-optimal commands for flexible systems were well developed, the actual calculation
of such commands and their practical implementation lagged far behind. Such commands could only be generated for very simple systems
and they suffered from poor robustness to modeling errors [138].

210/237

These limitations were greatly reduced by the robustness concept. Wie

and Liu were perhaps the first to see the possibility of using the zeroderivative robustness constraint of [94] in a timeoptimal formulation
that greatly reduces the sensitivity of time-optimal commands [51, 73,
141, 142].
Early on, researchers realized that there was an equivalence
between timeoptimal commands and input shaping when the input
shaper contains impulses of [1, -2, 2, -2, ..., -2, 1] and is convolved with
a step input whose magnitude is equal to the maximum actuator force
[107]. Rather than use the convolution process, Singh and Vadali
demonstrated the time-optimal input-shaping problem formulation
and solution in the frequency domain [101]. The extrainsensitive and
specified-insensitive constraints were also formulated and solved for
the special case of time-optimal control [17, 107].
Given that the resulting time-optimal command uses full actuator
effort throughout the motion, these commands can be used for on-off
actuators such as reaction jets on spacecraft. Unfortunately, the timeoptimal on-off commands are not efficient; they use a lot of actuator
effort (fuel). This fact motivated several research groups to search for
methods to make the commands more fuel efficient. Some of the
methods start with an inherently fuel-efficient command profile and
solve for the times at which the command should switch on and off
[105]. Other methods use a weighting function between move speed
and fuel usage [18, 63, 97, 143], or simply allow the command designer to specify the amount of fuel that is to be used for any particular
move [118]. All of these methods can be formulated by changing the
impulse times of the input shaper.
When using the on-off shaping methods above, researchers discovered that the internal forces and transient deflection could be quite
large because they are not directly controlled, and the full actuator effort is being used. Therefore, the system could be damaged during the
motion by large internal deflection forces. This can be especially damaging for spacecraft, so researchers have modified the command-

211/237

shaping process by adding additional constraints to limit transient deflection during the motion [32, 42, 87, 103].

15.5 Summary
Methods that generate specially-shaped reference commands to
move flexible systems are an important part of control theory and application - nearly 1,000 papers on the subject have been published.
The major advantage of such commands is that they do not require
sensor measurements, although sensors can be used in adaptive
command-shaping methods to improve performance. Another
strength of command shaping is that it acts to suppress vibration in a
preemptive way that is faster than anything possible with feedback
control. Feedback control must wait for an error to arise and be sensed
before suppression can be initiated. Command shaping uses a dynamic model to anticipate the occurrence of vibration, so it can effectively
start to act as soon as the system motion is initiated.
A key point in the history of command shaping was the development
of robust methods. These algorithms produce commands that can
work well even when large modeling errors exist. Furthermore, the dynamic model required by command shaping is usually quite simple,
consisting of estimates of the systems natural frequencies and damping ratios. Command shaping is very versatile in that many types of
auxiliary constraints, such as actuator limits, fuel usage, and transient
deflection limits, can be integrated into the design of the commands.
These beneficial properties have enabled engineers to harness the benefits of command shaping on millions of machines worldwide.

Appendix A: Notes on
Feedback Control
The introductory chapter discussed a general system that consisted
of a physical plant, command generator, feedback, and feedforward
control. While this is a general system description, many systems do
not contain all four elements. For example, the systems discussed in
Chapters 2 and 3 only contained the physical plant and the command
generator. These types of systems may lack final positioning accuracy
and can be greatly affected by disturbances. Therefore, feedback control is often added to address such problems. If this is the case, then
the system has the structure shown in Figure A.1. Note that the system
must retain the command generator; without this element there is no
means to create the reference signal and affect the state of the system.
This appendix does not provide a fundamental understanding of
feedback control systems. There are numerous books devoted entirely
to the subject. It is assumed that the reader already has good familiarity with feedback control. The goal here is to discuss feedback control
in the context of how it relates to command generation.
Important properties of feedback-controlled systems can be illustrated by considering the simple example of a mass under proportional and derivative (PD) control, as shown in Figure A.2. Transfer functions for the mass and PD controller are shown explicitly in the figure,
but the command generator has been left in generic form. The performance of the system depends on the dynamic effects between the
performance specifications going into the command generator and the
system response, Y. Block diagram reduction can be used to simplify
the closed-loop portion of the system, as shown in Figure A.3.

213/237

Figure A.1: System with Feedback Control.

Figure A.2: Rigid Mass Under PD Feedback Control.

Figure A.3: Reduced Representation of Mass Under PD

Control.
Note that the system has the same structure as those considered in
the previous chapters. The desired performance specifications are given to the command generator. These are then converted into a reference signal that drives the system. In this case, a feedback controller
has been added to the physical plant and this combination forms the
dynamic system driven by the reference command. The purpose of the
command generator is to select an effective reference command so
that the dynamic response achieves the desired performance

214/237

specifications. A command generator does not need to know whether

or not the system it is driving has a feedback controller, it simply
needs some knowledge of the overall dynamic response of the system
for which it is generating commands. If we examine the dynamic behavior of the system shown in Figure A.3, we will find that it has many
similarities to other vibratory systems shown throughout this book.
Furthermore, the reference commands that are issued to the system
will have similar effects as those discussed previously.
Consider the case when the desired motion is a rapid displacement
from one location to another location. A reasonable initial choice for
the reference signal would be a rapidly changing function like a step
input, a steep ramp, or an aggressive S-curve. Note from Figure A.2,
that the actuator effort, U, that is applied to the mass incorporates the
derivative of the reference signal, R. Therefore, if a step function were
used as the reference command, then the feedback controller would
request an impulse in actuator effort when the step command was first
issued. It is impossible for a real actuator to generate an impulse, so
the actuator would saturate. In a real-world application, this would
probably not cause a significant problem. The actuator would produce
as much force as it could until the feedback controller issued a request
for less effort. This would occur at the next time step through the control loop, as the derivative of the step function goes to zero at t = 0+.1

215/237

Figure A.4: Step Response of a Mass Under PD Feedback

Control.
Figure A.4 shows the response of the mass under PD control when a
step function is used as the reference command. For the case shown,
the mass has a value of 1, each of the control gains, Kp and Kd, are set
to 1, and the maximum actuator force is also limited to a value of 1.
The system has a sizable overshoot, so we may want to redesign the
system. The physical plant (mass) could be modified, the reference
command could be changed, or the control gains could be adjusted to
improve performance.
In this appendix, we will briefly examine the effect of changing the
controller gains. Figure A.5 shows the response as the proportional
gain is decreased from 1 to 0.5. The overshoot is reduced by decreasing
the gain, but the rise time is also increased. The overshoot could also
be reduced by increasing the derivative gain. Figure A.6 shows the

216/237

effect of this approach. Once again, the decreased overshoot comes at

the expense of slower rise time.
Note that the effect of the proportional controller gain is analogous
to attaching a spring to the mass and then pulling on the spring to
move the mass. On the other hand, the derivative gain acts like a
damper. Therefore, installing the feedback controller is similar to
changing the physical structure. However, because the feedback controller is a software routine, it may prove to be cheaper and more practical than making substantial changes to the physical structure. It is
also more adaptable because the system dynamics can be changed by
simply changing the gain values in the software. Feedback control
does have the drawback of requiring the installation of sensors and it
always poses the risk of inducing instability.
As mentioned above, the role of the command generator is not affected by the presence of a feedback controller. Command generators
are needed whether or not a feedback controller is installed on a system. However, the detailed structure and the parameters used in the
command generator are likely to change when a feedback controller is
installed. This is because the feedback controller will change the dynamic response of the system. A good control system will use a welldesigned command generator, and also a properly designed feedback
controller when necessary.

217/237

Figure A.5: Effect of Kp on the Dynamic Response of a Mass

Under PD Feedback Control.

218/237

Figure A.6: Effect of Kd on the Dynamic Response of a Mass

Under PD Feedback Control.
1

However, when performing numerical simulations of the system,

we must be careful to limit the maximum actuator effort to a finite
level so that unreasonable results are not obtained.

Appendix B: Derivation of
Residual Vibration
Amplitude Equation
The amplitude of residual vibration of an underdamped second-order system resulting from a series of impulses can be derived from the
response from a single impulse applied at time t0:

where, A0 is the amplitude of the impulse.

The response to a sequence of m impulses is simply the superposition of the response given in (B.1) with the appropriate values of impulse amplitudes and time locations entered into the equation. Therefore, the response to an impulse sequence, after the time of the last
impulse, tm, is given by:

Given (B.2), an expression for the amplitude of residual vibration

can be determined by using the trigonometric identity:

220/237

where, A is given by:

By comparing the trigonometric identity of (B.3) to the residual vibration given in (B.2), the Bi terms can be identified as:

To calculate the residual vibration amplitude, we evaluate (B.4) at

the time of the last impulse, tm. Substituting (B.5) into (B.4) and
bringing the constant portion of the coefficients out of the square root
term gives:

where C and S are given by:

To express the amplitude in a non-dimensional manner, (B.6) can

be divided by the amplitude of residual vibration from a single

221/237

impulse of unity magnitude. The amplitude of such an impulse response is:

The division of equation (B.6) by equation (B.9) yields a normalized

value, the percentage residual vibration at the time of the last impulse,
tm :

Bibliography
[1] C. F. Alsop, G. A. Forster, and F. R. Holmes. Ore unloader automation - a feasibility study. Tokyo Symposium on Systems Engineering for Control System Design, pgs. 295305. Tokyo, Japan, 1965.
[2] D. M. Aspinwall. Acceleration profiles for minimizing residual response. ASME Journal of Dynamic Systems, Measurement, and
Control, 102:36, March 1980.
[3] M. Athens and P. L. Falb. Optimal Control. McGraw-Hill, New
York, 1966.
[4] Arun Banerjee, Nelson Pedreiro, and William Singhose. Vibration
reduction for flexible spacecraft following momentum dumping
with/without slewing. AIAA J. of Guidance, Control, and Dynamics,
24(May- June):417428, 2001.
[5] Arun K. Banerjee. Dynamics and control of the wisp shuttle-antennae system. J. of Astronautical Sciences, 41(1):7390, 1993.
[6] Arun K. Banerjee and William E. Singhose. Command shaping in
tracking control of a two-link flexible robot. AIAA J. of Guidance,
Control, and Dynamics, 21(Nov.-Dec.):10121015, 1998.
[7] J. Ben-Asher, J. A. Burns, and E. M. Cliff. Time-optimal slewing of
flexible spacecraft. J. of Guidance, Control, and Dynamic,
15:36067, 1992.
[8] Sudarshan P. Bhat and Denny K. Miu. Precise point-to-point positioning control of flexible structures. ASME Winter Annual Meeting. San Francisco, CA, 1989.
[9] Sudarshan P. Bhat and Denny K. Miu. Precise point-to-point positioning control of flexible structures. ASME J. of Dynamic Systems,
Measurement, and Control, 112(4):667674, 1990.
[10] Sudarshan P. Bhat, M. Tanaka, and Denny K. Miu. Experiments
on point-to-point position control of a flexible beam using laplace

223/237

transform technique-part 1: Open-loop. ASME J. of Dynamic Systems, Measurement, and Control, 113:4327, 1991.
[11] Robert Blevins. Formulas for Natural Frequency and Mode
Shape. Van Nostrand Reinhold Co., New York, NY, 1979.
[12] Marc Bodson. Experimental comparison of two input shaping
methods for the control of resonant systems. IFAC World Congress.
San Francisco, CA, 1996.
[13] Marc Bodson. An adaptive algorithm for the tuning of two input
shaping methods. Automatica, 34(6):771776, 1998.
[14] Ray E. Bolz and George L. Tuve. CRC Handbook of Tables for Applied Engineering Science. CRC Press, Inc., Boca Raton, FL, 1973.
[15] Arthur Bryson and Yu-Chi Ho. Applied Optimal Control. Hemisphere Publishing, 1975.
[16] G. Cook. An application of half-cycle posicast. IEEE Trans. on
Automatic Control, 11(3):556559, 1966.
[17] A. Dhanda and G. F. Franklin. Optimal control formulations of vibration reduction problems. IEEE Trans. on Automatic Control,
55:37894, 2010.
[18] B. J. Driessen. On-off minimum-time control with limited fuel usage: Near global optima via linear programming. Optimal Control
Applications and Methods, 27(3):161168, 2006.
[19] R. L. Farrenkopf. Optimal open-loop maneuver profiles for flexible spacecraft. J. of Guidance and Control, 22:491498, 1979.
[20] J. T. Feddema. Digital filter control of remotely operated flexible
robotic structures. American Control Conf., volume 3, pgs.
27102715. ACC, San Francisco, CA, 1993.
[21] J. Fortgang and W. Singhose. Concurrent design of vibration absorbers and input shapers. J. of Dynamic Systems, Measurement,
and Controls, 127:329335, 2005.
[22] Joel Fortgang, Juan de Juanes Marquez, and William Singhose.
Application of input shaping on micro-mills. Japan USA Symposium on Flexible Automation. Denver, Colorado, 2004.

224/237

[23] Donald J. Gimpel and John F. Calvert. Signal component control.

AIEE Transactions, (Nov.):339343, 1952.
[24] Henry Hermes and Joseph P. Lasalle. Functional Analysis and
Time Optimal Control, volume 56 of Mathematics in Science and
Engineering. Academic Press, New York, 1969.
[25] Kenneth L. Hillsley and Stephen Yurkovich. Vibration control of a
twolink flexible robot arm. IEEE Int. Conf. on Robotics and Automation, pgs. 212126. Sacramento, CA, 1991.
[26] Kenneth L. Hillsley and Stephen Yurkovich. Vibration control of a
twolink flexible robot arm. J. of Dynamics and Control, 3:261280,
1993.
[27] Kyung-Taue Hong and Keum-Shik Hong. Input shaping and vsc
of container cranes. IEEE International Conference on Control Applications, pgs. 15701575. Taipei, Taiwan, 2004.
[28] Seong-Wook Hong, Gyu-Hyun Bae, and Byung-Gyu Kim. Development of miniature tower crane and payload position tracking system using webcam for education. IASTED Int. Conference on Robotics and Applications. Cambridge, MA, 2010.
[29] Qinglei Hu. Input shaping and variable structure control for simultaneous precision positioning and vibration reduction of flexible
spacecraft with saturation compensation. Journal of Sound and
Vibration, 318(n 1-2):18 35, 2008.
[30] James Hyde and Warren Seering. Using input command preshaping to suppress multiple mode vibration. IEEE Int. Conf. on
Robotics and Automation, pgs. 26042609. Sacramento, CA, 1991.
[31] Kimmo Hytnen. Method for controlling a crane. U.S. Patent
7,484,632, Feb. 3 2009.
[32] Amine Kamel, Friedrich Lange, and Gerd Hirzinger. New aspects
of input shaping control to damp oscillations of a compliant force
sensor. IEEE Int. Conference on Robotics and Automation, pgs.
262935. 2008.

225/237

[33] Michael Kenison and William Singhose. Input shaper design for
doublependulum planar gantry cranes. IEEE Conference on Control
Applications, pgs. 53944. Hawaii, 1999.
[34] Michael Kenison and William Singhose. Concurrent design of input shaping and proportional plus derivative feedback control.
ASME J. of Dynamic Systems, Measurement, and Control,
124(3):398405, 2002.
[35] Attir Khalid, John Huey, William Singhose, and Jason Lawrence.
Human operator performance testing using an input-shaped bridge
crane. ASME J. of Dynamic Systems, Measurement, and Control,
128(4):835841, 2006.
[36] F. Khorrami, S. Jain, and A. Tzes. Experiments of rigid-body
based controllers with input preshaping for a two-link flexible manipulator. American Control Conference, volume 3, pgs.
29572961. Chicago, IL, 1992.
[37] F. Khorrami, S. Jain, and A. Tzes. Adaptive nonlinear control and
input preshaping for flexible-link manipulators. American Control
Conference, pgs. 27052709. San Francisco, CA, 1993.
[38] F. Khorrami, S. Jain, and A. Tzes. Experiments on rigid-body
based controllers with input preshaping for a two-link flexible manipulator. IEEE Transactions on Robotic Automation, volume 10,
pgs. 5565. 1994.
[39] Dooroo Kim and William Singhose. Human operator learning on
doublependulum bridge cranes. ASME IMECE. Seattle, WA, 2007.
[40] Dooroo Kim and William Singhose. Studies of human operators
manipulating double-pendulum bridge cranes. European Control
Conference. Kos, Greece, 2007.
[41] Dooroo Kim and William Singhose. Performance studies of human operators driving double-pendulum bridge cranes. Control
Engineering Practice, 18:567576, June 2010.
[42] H. Kojima and W. Singhose. Adaptive deflection limiting control
for slewing flexible space structures. AIAA J. of Guidance, Control,
and Dynamics, 30:6167, 2007.

226/237

[43] R. L. Kress, J. F. Jansen, and M. W. Noakes. Experimental implementation of a robust damped-oscillation control algorithm on a full
sized, two-dof, ac induction motor-driven crane. 5th International
Symposium on Robotics and Manufacturing, pgs. 58592. Maui,
HA, 1994.
[44] Ravi Kumar and Tarunraj Singh. Design of input shapers using
modal cost for multi-mode systems. Automatica, 46(3):598604,
2010.
[45] H. Kwakernaak and J. Smith. Minimum vibration cam profiles. J.
of Mechanical Engineering Science, 10:219227, 1968.
[46] Jason Lawrence and William Singhose. Design of minicrane for
education and research. 6th Int. Conference on Research and Education in Mechatronics. Annecy, France, 2005.
[47] Jason Lawrence, William Singhose, Rolff Weiss, Adrian Erb, and
Urs Glauser. An internet-driven tower crane for dynamics and controls education. IFAC Symp. on Control Education. Madrid, Spain,
2006.
[48] Derek Lewis, Gordon G. Parker, Brian Driessen, and Rush D.
Robinett. Command shaping control of an operator-in-the-loop
boom crane. American Control Conference, pgs. 26437. Philadelphia, PA, 1998.
[49] Sungyung Lim and Jonathon How. Input command shaping techniques for robust, high-performance control of flexible structures.
AIAA Guidance, Navigation, and Control Conf. San Diego, CA,
1996.
[50] Sungyung Lim, H.D. Stevens, and J.P. How. Input shaping design
for multi-input flexible systems. J. of Dynamic Systems, Measurement and Control, 121(3):4437, 1999.
[51] Q. Liu and B. Wie. Robust time-optimal control of uncertain flexible spacecraft. J. of Guidance, Control, and Dynamics,
15:597604, 1992.

227/237

[52] D. P. Magee, D. W. Cannon, and W. J. Book. Combined command

shaping and inertial damping for flexure control,. American Control Conf., volume 3, pgs. 13301334. 1997.
[53] David P. Magee and Wayne J. Book. Eliminating multiple modes
of vibration in a flexible manipulator. Proceedings of the IEEE International Conference on Robotics and Automation, volume 2,
pgs. 474479. IEEE, Atlanta, GA, 1993.
[54] David P. Magee and Wayne J. Book. Implementing modified
command filtering to eliminate multiple modes of vibration. Proceedings of the American Controls Conference, volume 2, pgs.
474479. IEEE, San Francisco, CA, 1993.
[55] David P. Magee and Wayne J. Book. Filtering schilling manipulator commands to prevent flexible structure vibration. Proceedings of
the American Controls Conference, volume 2, pgs. 474479. IEEE,
Baltimore, MD, 1994.
[56] David P. Magee and Wayne J. Book. Filtering micro-manipulator
wrist commands to prevent flexible base motion. Proceedings of the
American Controls Conference, volume 2, pgs. 474479. IEEE,
Seattle, WA, 1995.
[57] Raymond Manning, Jeffrey Clement, Dooroo Kim, and William
Singhose. Dynamics and control of bridge cranes transporting
distributed-mass payloads. ASME J. Dynamic Systems, Measurement, and Control, 132:0145053, Jan. 2010.
[58] Raymond Manning, Dooroo Kim, and William Singhose. Reduction of distributed payload bridge crane oscillations. WSEAS International Conference on Automatic Control, Modelling, and Simulation. Istanbul, Turkey, 2008.
[59] Peter Meckl and Warren Seering. Experimental evaluation of
shaped inputs to reduce vibration for a cartesian robot. J. of Dynamic Systems, Measurement, and Controls, 112:159165, June
1990.

228/237

[60] Peter H. Meckl and Warren P. Seering. Minimizing residual vibration for point-to-point motion. Journal of Vibration, Acoustics,
Stress and Reliability in Design, 107:378382, October 1985.
[61] Peter H. Meckl and Warren P. Seering. Reducing residual vibration in systems with time varying resonances. IEEE Int. Conf. on
Robotics and Automation, pgs. 16901695. Raleigh, NC, 1987.
[62] D. H. Mee. A feedback implementation of posicast control using
sampling circuits. Proceedings of the Institute of Radio and Electronics Engineering, Jan-Feb:1115, 1974.
[63] J. L. Meyer and L. Silverberg. Fuel optimal propulsive maneuver
of an experimental structure exhibiting spacelike dynamics. Journal
of Guidance, Control, and Dynamics, 19(1):1419, 1996.
[64] Marco Muenchhof and Tarunraj Singh. Concurrent feedforward/
feedback controller design using time-delay filters. AIAA Guidance,
Navigation and Control Conference. Monterey, CA, 2002.
[65] Marco Muenchhof and Tarunraj Singh. Jerk limited time optimal
control of flexible structures. J. of Dynamic Systems, Measurement,
and Controls, 125(1):13942, 2003.
[66] Rory Ninneman and Keith Denoyer. Middeck active control experiment reflight (mace ii): Lessons learned and reflight status. Proceedings of SPIE - The International Society for Optical Engineering, volume 3991, pgs. 131137. Newport Beach, CA, 2000.
[67] American Society of Mechanical Engineers. Methods for performance evaluation of coordinate measuring machines. Tech. report,
ASME B89.1.12M, 1990.
[68] Katsuhiko Ogata. Modern Control Engineering, pg. 282.
Prentice-Hall, Inc., 1970.
[69] Yuichi Okazaki, Toshimichi Mori, and Noboru Norita. Desk-top
nc milling machine with 200 krpm spindle. ASPE 2001 Annual
Meeting, pgs. 192 195. Crystal City, VA, 2001.
[70] A.V. Oppenheim and R.W. Schafer. Digital Signal Processing.
Prentice Hall, Inc., Englewood Cliffs, NJ, 1975.

229/237

[71] L. Y. Pao. Multi-input shaping design for vibration reduction.

Automatica, 35:8189, 1999.
[72] L. Y. Pao, T. N. Chang, and E. Hou. Input shaper designs for minimizing the expected level of residual vibration in flexible structures.
American Control Conference, pgs. 35423546. Albuquerque, NM,
1997.
[73] Lucy Y. Pao and William E. Singhose. Verifying robust time-optimal commands for multi-mode flexible spacecraft. AIAA J. of
Guidance, Control, and Dynamics, 20(4):831833, 1997.
[74] Lucy Y. Pao and William E. Singhose. Robust minimum time control of flexible structures. Automatica, 34(2):229236, 1998.
[75] J. Park and S. Rhim. Extraction of optimal time-delay in adaptive
command shaping filter for flexible manipulator control. J. of Inst.
of Control, Robotics and System, 14(564-572), 2008.
[76] Juyi Park and Pyung Hun Chang. Learning input shaping technique for not-lti systems. American Control Conference. Philadelphia, PA, 1998.
[77] U. H. Park, J. W. Lee, B. D. Lim, and Y. G. Sung. Design and sensitivity analysis of an input shaping filter in the z-plane. J. of Sound
and Vibration, 243:157171, 2001.
[78] Gordon G. Parker, Kenneth Groom, Johnny Hurtado, Rush D.
Robinett, and Frank Leban. Experimental verification of a command shaping boom crane control system. American Control Conference, pgs. 8690. San Diego, CA, 1999.
[79] T.W. Parks and C.S. Burrus. Digital Filter Design. John Wiley &
Sons, Inc., New York, 1987.
[80] E. Pereira, J.R. Trapero, I.M. Daz, and V. Feliu. Adaptive input
shaping for manoeuvring flexible structures using analgebraic identification technique. Automatica, 45(4):10461051, 2009.
[81] L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze, and E. F.
Mishchenko. The Mathematical Theory of Optimal Processes. John
Wiley & Sons, New York, 1962.

230/237

[82] B. W. Jr. Rappole, N.C. Singer, and W. P. Seering. Multiple-mode

impulse shaping sequences for reducing residual vibrations. 23rd
Biennial Mechanisms Conference, pgs. 1116. Minneapolis, MN,
1994.
[83] Kuldip S. Rattan and Vincente Feliu. Feedforward control of flexible manipulators. IEEE International Conference on Robotics and
Automation, volume 1, pgs. 788793. Nice, France, 1992.
[84] S. Rhim and W. Book. Noise effect on time-domain adaptive command shaping methods for flexible manipulator control. IEEE
Trans. of Control Systems Technology, 9(1):8492, 2001.
[85] S. Rhim and W. Book. Adaptive time-delay command shaping filter for flexible manipulator control. IEEE/ASME Trans. on Mechatronics, 9:619 626, 2004.
[86] S. Rhim andW. J. Book. Adaptive command shaping using adaptive filter approach in time domain. American Control Conference,
pgs. 8185. San Diego, CA, 1999.
[87] Michael Robertson and William Singhose. Specified-deflection
command shapers for second-order position input systems. ASME
J. of Dynamic Systems, Measurement and Control,
129(6):856859, 2007.
[88] Michael Robertson, Andrew Timm, and William Singhose. Evaluation of command generation techniques for tethered satellite retrieval. AAS/AIAA Space Flight Mechanics Conference. Copper
Mountain, CO, 2005.
[89] H. Rothbart. Cams: Design, Dynamics, and Accuracy. John Wiley & Sons, Inc., 1956.
[90] V. C. Shields and G. Cook. Application of an approximate time
delay to a posicast control system. Int. Journal of Control,
14(4):649657, 1971.
[91] Neil Singer and Warren Seering. An extension of command shaping methods for controlling residual vibration using frequency
sampling. IEEE International Conference on Robotics and Automation, volume 1, pgs. 800805. IEEE, Nice, France, 1992.

231/237

[92] Neil Singer, William Singhose, and Eric Kriikku. An input shaping controller enabling cranes to move without sway. ANS 7th Topical Meeting on Robotics and Remote Systems, volume 1, pgs.
22531. Augusta, GA, 1997.
[93] Neil C. Singer. Residual Vibration Reduction in Computer Controlled Machines. Ph.D. thesis, Massachusetts Institute of Technology, 1989.
[94] Neil C. Singer and Warren P. Seering. Preshaping command inputs to reduce system vibration. J. of Dynamic Sys., Measurement,
and Control, 112:7682, 1990.
[95] Neil C. Singer, William E. Singhose, and Warren P. Seering. Comparison of filtering methods for reducing residual vibration.
European Journal of Control, (5):208218, 1999.
[96] T. Singh and S. R. Vadali. Input-shaped control of three-dimensional maneuvers of flexible spacecraft. J. of Guidance, Control, and
Dynamics, 16(6):10618, 1993.
[97] Tarun Singh. Fuel/time optimal control of the benchmark problem. J. of Guidance, Control, and Dynamics, 18(6):122531, 1995.
[98] Tarunraj Singh. Jerk limited input shapers. J. of Dynamic Systems, Measurement, and Controls, 126(1):21519, 2004.
[99] Tarunraj Singh and G. R. Heppler. Shaped input control of a system with multiple modes. Journal of Dynamic Systems, Measurement, and Control, 115:341347, September 1993.
[100] Tarunraj Singh and S. R. Vadali. Robust time-delay control.
ASME Journal of Dynamic Systems, Measurement, and Control,
115:3036, 1993.
[101] Tarunraj Singh and S. R. Vadali. Robust time-optimal control: A
frequency domain approach. J. of Guidance, Control and Dynamics, 17:346 353, 1994.
[102] William Singhose. A vector diagram approach to shaping inputs
for vibration reduction. Tech. Report MIT Artificial Intelligence Lab
Memo No. 1223, MIT, 1990.

232/237

[103] William Singhose, Arun Banerjee, and Warren Seering. Slewing

flexible spacecraft with deflection-limiting input shaping. AIAA J. of
Guidance, Control, and Dynamics, 20(2):291298, 1997.
[104] William Singhose, Erika Biediger, Ye-Hwa Chen, and Bart Mills.
Reference command shaping using specified-negative-amplitude input shapers for vibration reduction. ASME J. of Dynamic Systems,
Measurement, and Controls, 126(March):210214, 2004.
[105] William Singhose, Kristen Bohlke, and Warren Seering. Fuel-efficient pulse command profiles for flexible spacecraft. AIAA J. of
Guidance, Control, and Dynamics, 19(4):954960, 1996.
[106] William Singhose, Ethan Crain, and Warren Seering. Convolved
and simultaneous two-mode input shapers. IEE Control Theory and
Applications, 144(Nov.):515520, 1997.
[107] William Singhose, Steve Derezinski, and Neil Singer. Extra-insensitive input shapers for controlling flexible spacecraft. AIAA J. of
Guidance, Control, and Dynamics, 19(2):38591, 1996.
[108] William Singhose, John Huey, Jason Lawrence, and David
Frakes. Input shaping curriculum: Integrating interactive simulations and experimental setups. 11th Mediterranean Conference on
Control Automation, pgs. IV1105. Rhodes, Greece, 2003.
[109] William Singhose, Dooroo Kim, and Michael Kenison. Input
shaping control of double-pendulum bridge crane oscillations.
ASME J. of Dynamic Systems, Measurement, and Control,
130(034504), May 2008.
[110] William Singhose, Lisa Porter, Michael Kenison, and Eric
Kriikku. Effects of hoisting on the input shaping control of gantry
cranes. Control Engineering Practice, 8(10):11591165, 2000.
[111] William Singhose, Lisa Porter, Timothy Tuttle, and Neil Singer.
Vibration reduction using multi-hump input shapers. ASME J. of
Dynamic Systems, Measurement, and Control, 119(June):320326,
1997.
[112] William Singhose and Warren Seering. Generating vibration-reducing inputs with vector diagrams. IFToMM Eighth World

233/237

Congress on the Theory of Machines and Mechanisms, volume 1,

pgs. 315318. Prague, Czechoslovakia, 1991.
[113] William Singhose, Warren Seering, and Neil Singer. Shaping inputs to reduce vibration: A vector diagram approach. IEEE Int.
Conf. on Robotics and Automation, volume 2, pgs. 922927. IEEE,
Cincinnati, OH, 1990.
[114] William Singhose, Warren Seering, and Neil Singer. Residual vibration reduction using vector diagrams to generate shaped inputs.
ASME J. of Mechanical Design, 116(June):654659, 1994.
[115] William Singhose, Warren Seering, and Neil Singer. Input shaping for vibration reduction with specified insensitivity to modeling
errors. Japan- USA Sym. on Flexible Automation, volume 1, pgs.
30713. Boston, MA, 1996.
[116] William Singhose, Neil Singer, and Warren Seering. Improving
repeatability of coordinate measuring machines with shaped command signals. Precision Engineering, 18(April):138146, 1996.
[117] William Singhose, Neil Singer, and Warren Seering. Time-optimal negative input shapers. ASME J. of Dynamic Systems, Measurement, and Control, 119(June):198205, 1997.
[118] William Singhose, Tarun Singh, and Warren Seering. On-off control with specified fuel usage. J. Dynamic Systems, Measurement,
and Control, 121(2):206 212, 1999.
[119] William Singhose and Samuel Towell. Double-pendelum gantry
crane dynamics and control. IEEE Conf. on Control Applications.
Trieste, Italy, 1998.
[120] O. J. M. Smith. Feedback Control Systems. McGraw-Hill Book
Co., Inc., New York, 1958.
[121] Otto J. M. Smith. Posicast control of damped oscillatory systems.
Proceedings of the IRE, 45(September):12491255, 1957.
[122] Khalid Sorensen and William Singhose. A multi-operationalmode antisway and positioning control for an industrial 30-ton
bridge crane. IFAC World Congress. Seoul, Korea, 2008.

234/237

[123] Khalid Sorensen, William Singhose, and Stephen Dickerson. A

controller enabling precise positioning and sway reduction in bridge
and gantry cranes. Control Engineering Practice, 15(7):825837,
2007.
[124] G. P. Starr. Swing-free transport of suspended objects with a
pathcontrolled robot manipulator. J. of Dynamic Systems, Measurement and Control, 107:97100, 1985.
[125] D.R. Strip. Swing-free transport of suspended objects: a general
treatment. IEEE Trans. Robotics and Automation, 5(2):234 6,
1989. ISSN 1042-296X. Materials handling;suspended objects;control strategies;swing-free movement;bridge crane;.
[126] Y. G. Sung and J. P. Wander. Applications of vibration reduction
of flexible space structure using input shaping technique. Southeastern IEEE Conference, pgs. 333336. Athens, OH, 1993.
[127] C. J. Swigert. Shaped torque techniques. Journal of Guidance
and Control, 3(5):460467, 1980.
[128] G. H. Tallman and O. J. M. Smith. Analog study of dead-beat
posicast control. IRE Transactions on Automatic Control,
(March):1421, 1958.
[129] F. Taylor. Digital Filter Design Handbook. Marcel Dekker, Inc,
New York, 1983.
[130] D. Tesar and G. K. Mathew. The Dynamic Synthesis, Analysis,
and Design of Modeling Cam Systems. Lexington Books, 1935.
[131] J. D. Turner and J. L. Junkins. Optimal large-angle single-axis
rotational maneuvers of flexible spacecraft. Journal of Guidance
and Control, 3:578 585, 1980.
[132] T. D. Tuttle and W. P. Seering. Vibration reduction in flexible
space structures using input shaping on mace: Mission results. IFAC
World Congress, pgs. 5560. San Francisco, CA, 1996.
[133] Timothy Tuttle and Warren Seering. Experimental verification of
vibration reduction in flexible spacecraft using input shaping. J. of
Guidance, Control, and Dynamics, 20(4):658664, 1997. Command Shaping.

235/237

[134] Timothy D. Tuttle andWarren P. Seering. Vibration reduction in

0-g using input shaping on the mit middeck active control experiment. American Control Conf., volume 2, pgs. 919923. ACC,
Seattle, WA, 1995.
[135] Anthony Tzes and Stephen Yurkovich. An adaptive input shaping
control scheme for vibration suppression in slewing flexible structures. IEEE Transactions on Control Systems Technology,
1(June):114121, 1993.
[136] Anthony P. Tzes, Matthew J. Englehart, and Stephen Yurkovich.
Input preshaping with frequency domain information for flexiblelink manipulator control. AIAA Guidance, Navigation and Control
Conference, volume 2, pgs. 11671175. Boston, MA, 1989.
[137] Athony P. Tzes and Stephen Yurkovich. Adaptive precompensators for flexible-link manipulator control. 28th Conference on Decision and Control, pgs. 208388. Tampa, FL, 1989.
[138] Wallace VanderVelde and Jun He. Design of space structure
control systems using on-off thrusters. Journal Guidance, Control
and Dynamics, 6(1):5360, 1983.
[139] J. Vaughan and W. Singhose. Reducing vibration and providing
robustness with multi-input shapers. American Control Conference.
2009.
[140] Joshua Vaughan, Dooroo Kim, and William Singhose. Control of
tower cranes with double-pendulum dynamics. IEEE Trans. on
Control Systems Technology, 18(6):1345 1358, 2010.
[141] B. Wie, R. Sinha, and Q. Liu. Robust time-optimal control of uncertain structural dynamic systems. J. of Guidance, Control, and
Dynamics, 15:9803, 1993.
[142] Bong Wie and Qiang Liu. Comparison between robustified feedforward and feedback for achieving parameter robustness. Journal
of Guidance, Control, and Dynamics, 15(4):935943, 1992.
[143] Bong Wie, Ravi Sinha, John Sunkel, and Ken Cox. Robust fueland timeoptimal control of uncertain flexible space structures. AIAA

236/237

Guidance, Navigation, and Control Conference, volume 3, pgs.

939948. Monterey, CA, 1993.
[144] J. L. Wiederrich and B. Roth. Design of low vibration cam profiles. Conference on Cams and Cam Mechanisms. Liverpool, England, 1974.
[145] J. L. Wiederrich and B. Roth. Dynamic synthesis of cams using
finite trigonometric series. ASME Journal of Engineering for Industry, pgs. 287293, Feb. 1975.
[146] Stephen Yurkovich, Anthony P. Tzes, and Kenneth L. Hillsley.
Controlling coupled flexible links rotating in the horizontal plane.
American Control Conference, pgs. 362367. San Diego, CA, 1990.
[147] A. Zverev. Handbook of Filter Synthesis. John Wiley & Sons,
1967.

@Created by PDF to ePub