Professional Documents
Culture Documents
Smart-Structures Technology For Parallel Robots
Smart-Structures Technology For Parallel Robots
DOI 10.1007/s10846-010-9522-8
Received: 12 January 2010 / Accepted: 15 December 2010 / Published online: 21 January 2011
Springer Science+Business Media B.V. 2011
548
1 Introduction
Industrial robots are a key technology in the automation of handling and assembly
processes. Serial robots with several degrees of freedom (DOF) are usually used for
this kind of application. The improvement of their dynamic behavior is limited due
to the large amount of moved mass. In consequence it is worth to develop parallel
robots for a further increase of the robots acceleration and velocity. A typical parallel
robot has a small amount of moved masses and is therefore suitable for high-dynamic
tasks. On the other hand higher accelerations increase the risk of severe vibration of
the structural parts. Therefore, smart-structures technology is introduced in parallel
robotics for a further enhancement of their features [1, 2].
The introduction of smart-structure technology in robotics involves three main
disciplines.
These principle challenges are addressed within the German Collaborative Research
Center SFB562 Robotic Systems for Handling and Assembly which aims to achieve
a major step forward in the performance of parallel robots. This paper summarizes
the contributions of the Collaborative Research Center to these main disciplines.
The latest test platform is called Triglide [6, 8]. It is a four DOF robot with three
translational and one rotational DOF.
Figure 1 offers a view in the workspace of Triglide. The end effector is in the
middle of this figure. The end effector includes a triaxial accelerometer to measure
549
vibrations of the robot and a serial rotatory drive to realize the fourth DOF. The
end effector is connected to three linear drives by links, each consisting of two active
rods, see Fig. 2. The distance between the two lower bearings is 828 mm, the vertical
distance to the third bearing is 450 mm and the rod length is 600 mm. The x-axis of
the machine coordinate system is collinear and the y-z-plane is perpendicular to the
linear drives.
The active rods have an inner diameter of 25 mm and a wall thickness of 1.5 mm.
They are made of Carbon Fiber-Reinforced Plastics (CFRP). Piezoceramic patch
actuators are applied on the surface to actuate the rod. A single rod reaches a free
stroke (unloaded case) in axial direction of 35 m at 400 V piezo voltage. The piezo
patches of each pair of rods are connected in parallel. Therefore, three control signals
have to be generated to drive the smart components of the system.
Active rods in itself are not broadly investigated in literature. In [13] rods for
active vibration suppression are discussed, albeit for use in an truss structure, where
the actuators in the active rods have to work against outer structural elasticity of the
truss as well as inner elasticity of the rod. The topology of these rods is determined by
discrete piezo stacks as discussed in Section 2.1. Other setups of the rods are required
for parallel mechanisms, because the design goals are different as well as the loads to
be considered. Nonetheless the principle of dimensioning active rods as presented in
this article could be transferred to other load cases.
There are many different ways for modeling the kinematics and dynamics of
structures. Some packages are available in software systems like Matlab or Mathematica [16]. In the field of parallel robot structures with closed loop kinematics,
general purpose multi-body systems like Simpack [12] or Adams are very useful
tools for common tasks and for verification purposes (which we have used quite
frequently to check our tools), but in practice it is often advantageous to have direct
access to the underlying system equations as well as on the implemented data flow
in the equations. This also allows direct control of the exported computer code for
evaluation of the underlying equations of motion and forces. To have control on
550
the data flow, it is important to reuse coded equations for different applications:
For the trajectory based robot control, rigid body models are needed to obtain the
kinematic and dynamic equations. The kinematic equations are essential for the
correct implementation of the robot task. To reduce the deviation of the actual
trajectory path with respect to the planned one, inverse dynamic models are used
in the so-called computed torque control algorithm [14]. And for studying elastic
vibrations, some bodies might be replaced by elastic bodies including several e.g.
modal vibration patterns [9].
With respect to vibration isolation, the source of disturbances is given by the base
where the system is mounted. Regarding parallel robots for handling and assembly,
the inertial forces of the moving robot represent the major source of disturbances.
Beside the Collaborative Research Center Wang und Mills from University of
Toronto investigated vibration control of planar parallel robots with three degrees
of freedom (DOF) for several years. The main part of their work was the exact
modeling of the dynamic and elastic behavior of the robot structure [22]. Beside
Finite-Element calculations analytical approaches were used as well. They chose a
modal control concept [23, 24]. In his PhD thesis Wang [21] described this concept.
In spite of a position dependent model of the vibration behavior, no approach of a
position dependent control can be found. Control performance measurements are
done in certain operating points for which a controller is synthesized previously. A
strategy for vibration control on a trajectory is missing. Wangs successor Zhang
investigates the improvement of the modeling of coupled dynamic and elastic
behavior of a planar parallel robot [25].
In various applications gain-scheduling approaches are used successfully. The
most common approach is based on a linearization of the non-linear plant about a
family of set points or operating points [18]. This linearization leads to a parameterdependent family of linear systems, described as polytopic or affine parameter
dependent state-space models. In some cases, this parameter dependence can be
modeled in terms of linear fractional transformations (LFT). Linear matrix inequalities (LMI) are able to synthesize stable parameter dependent controllers which
gain performance over the entire parameter range [4, 11]. For Triglide system
identification and controller synthesis are automated and implemented in C on the
target system. Due to the complexity of LMI computations they are not feasible for
this kind of application. Therefore, controllers are synthesized with common and fast
Riccati-based H -algorithms. Afterwards global stability of the control loop in LFT
representation is tested by use of small gain theorem.
2 Components
Different components for adaptive systems in parallel robots are investigated. The
application of smart components aims at the vibration suppression which allows
for shorter cycle-times in handling and assembly. The rationale is that vibration
is induced by the high accelerations necessary to realize short cycle-times. These
vibrations must decay before new handling and assembly tasks can be performed
in the new robot position with high precision. Reduction of the vibration allows to
tap the full potential of parallel mechanism for handling and assembly.
551
The following components are investigated for the implementation of smartstructures technologies:
This section focusses on active rods because they are implemented in the
Triglide-structure.
2.1 Choosing Actuators
The stroke of active rods has to be as high as possible for the best performance
in vibration suppression, while the added weight due to active material has to
be minimized. On the other hand all components have to withstand the applied
structural loads. Piezoceramic material is chosen as smart material due to its good
industrial availability and its well known behavior.
Active rods exert a maximal elongation l given by
cA
1
l = l0
= l0
,
(1)
c A + cS
1+V
where l0 is the free stroke of the piezo-actuator, c A the stiffness of the actuator and
c S the stiffness of the load carrying structure. In order to reach high actuation authority a low ratio of V = c S /c A and high l0 is needed. Furthermore, c S determines
the structural deformation under external load. Thus, structural conformity of the
integrated smart materials has to be realized.
Different topologies of rods can be chosen. The upper part of Fig. 3 depicts a
setup with a discrete piezostack. A prestressing is needed in order to overcome the
brittleness of the actuator. Prestressing causes a tensional loaded structure and a
Rod
(compressive load)
Actuators
Structure
(tensional load)
Structure
552
compressed rod, assuming that the stack does not expand to the full length of the
rod, which is a sensible assumption weight-wise. Such a configuration [3, 7] gives a
low ratio V but also a low l0 . The topology of the related configuration (upper
part of Fig. 3) is based on a more complicated design than the one shown at the
lower part of the figure, where thin actuators are attached to the surface of the rods
structure. The ratio V from Eq. 1 is higher than for the discrete stack actuator, but
the l0 is higher, as the actuators expand over the full length of the rod. In order
to get the required prestressing, piezocomposites as shown in Fig. 4 are used. They
provide the electrical insulation and connection voiding the need for the complicating
prestressing topology needed with discrete actuators. For the Triglide demonstrator
rods with distributed actuators are chosen finally.
2.2 Dimensioning of Active Rods
The combination of surface mounted piezocomposites with fiber composite material
leads to an easy dimensioning process of the rods. This process is demonstrated
assuming a fiber layout as shown in Fig. 5 with a length of the rod being 42.4 cm,
a radius of 1 cm and wall-thickness of 2 mm [14].
The thickness of the actuators and the fiber-angle are the design parameters, while
the length (and thus the free stroke l0 of the actuators) is given by the design of
the robot structure. The ratio V is given in Fig. 6 depending on the thickness of the
used actuators and the angle of the fiber orientation [5]. An optimum is reached
for high angles of fiber orientation combined with high thickness of piezoactuators
respectively.
Dimensioning loads have been considered in order to prevent a rod failure.
Models of the structural dynamics are used to obtain these loads early in the
design process when loads are not available. The definition process of these loads
is described below. A 2500 N load is obtained as dimensioning load for the rods with
distributed actuators. The parameters of the design process have to be chosen in a
way that all bearable forces are above the dimensioning load. Of course, matters of
553
+
+
manufacturability can further reduce the number of possible parameters, e.g. the
ability to apply thick ceramics to a circular rod of small radius is limited due to
reaching the critical strain of the embedded actuators during bonding.
Two failure modes are of interest for the active rods. The critical strain under
tensional load of the embedded actuators has to be avoided. The critical strain is
reached when the prestress of the piezocomposite is exhausted and the piezoceramic
fails due to brittleness. Figure 7 shows the parameters where the strain is below the
critical value above the shaded plane.
554
555
556
r1
r2
e3
Platform (enlarged)
e3
r3
C12
C31
yE
C2e2
F, F
e2
C21 -f
P
B3
q4
C1 y E
E,F,F,C3 C2
yF
p1
3
-p3
xE
xF
q3
C22
zE
e1 C1
e1 C11
C32
C3
l3
A3
zE
l1
B1
q1
C1
z
C3
E
yE
l2
xE
B2
C2
x
1
q2
A1
a1
O
A2
b1
b2
a3
2
a2
b3
557
b and c respectively. According to Fig. 10, all the necessary reference coordinate
systems can be obtained by the relations
O1 = T y b 1 a1 Rx 180 O,
O2 = T y b 2 a2 O,
O3 = T y b 3 a3 Rx 90 O,
OAS(GS)1 = R y 180 T y a1 OAS1 ,
OAS(GS)2 = T y a2 OAS2 ,
OAS(GS)3 = T y a3 OAS3 ,
OASi = Tx qi Oi ,
OGSi = Rz i OAS(GS)i ,
OGS(SP)i = Rx 90 Tz ei OGSi ,
OGS(SN)i = Rx 90 Tz ei OGSi ,
OSPi = Rx i OGS(SP)i ,
OSNi = Rx i OGS(SN)i ,
OSP(GP)i = R y 180 Tz i OSPi ,
OGP(SP)i = Rx 90 Tz ei OGPi ,
OGP(SP)i = Rx i OSP(GP)i ,
OSN(GP)i = R y 180 Tz i OSNi ,
OGP(SN)i = Rx 90 Tz ei OGPi ,
OGP(SN)i = Rx i OSN(GP)i ,
OGPi = Rz i OAP(GP)i ,
OAP(GP)1 = T y r1 OAP ,
OAP(GP)2 = T y r2 Rz 180 OAP ,
OAP(GP)3 = R y 180 T y r3 Rx 90 OAP ,
OF = Tz f OAP ,
OEF = Rz OAP ,
OP = Tz p3 T y p2 Tx p1 OEF .
All parallel structures contain kinematically closed loops. Therefore, to solve the
underlying non-linear equations of motion, which can be obtained in a systematic
way using the transformations just given, numerical solution procedures must be
applied in general. In case of the Triglide-structure, the so called direct kinematical
problem (DKP) with prescribed drive position parameters q1 , q2 , q3 and q4 =
as well as the inverse kinematical problem (IKP) with known TCP parameters xP ,
yP , zP and P = from the trajectory planning module can be solved explicitly by
elementary operations.
558
Pk =
3
q i fqi + ,
(2)
i=1
which relates all internal and external virtual power in the system. Here fqi denotes
the translational drive forces, the drive moment of the fourth rotational drive
mounted on the TCP-platform, and the sum includes all rigid and elastic components
or bodies of the system. Especially for rigid bodies with mass m, center of gravity c
and inertia tensor J, the virtual power is given by
T
mI mcT a g
v
m c
P =
+
.
(3)
J
mc J
Note that all vectors and tensors (including the earth gravitation g) are used with
respect to a body fixed reference system, which moves with absolute velocity v,
acceleration a and angular velocity . The tilde operator is defined as usual by the
= c. The inertia tensor J0 with respect to c can be used to
cross product c
calculate the inertia tensor J = J0 + mcc T of the body reference system.
From these equations and the kinematical restrictions on the position parameters
and its variations, a very efficient real time C-code has been derived, which was
used as dynamic model of the rigid body Triglide-structure in the computed torque
control method similar as has been done before for the HEXA-structure (see [14]).
The non-linear dynamic model was also built in the multi body system Simpack to
verify the calculated drive forces of the real-time model. In order to excite all parts of
the systems, complicated three dimensional trajectories have been used as shown in
Fig. 11. After some code debugging, finally the forces of both models matched within
floating point precision. The constraint forces in the joints were calculated as well
and used in the design process of the active rods.
559
Pbeam,k = A
w(z)
dz
0
+ EI
(4)
with the transversal deformation w(z) and the longitudinal elongation u(z). For
simplicity, the mass density , the cross sectional area A, the Young modulus E and
the geometrical moment of inertia I are assumed to be constant in beam direction.
The beam is segmented and using the classical FE-approach, suitable stiffness and
mass matrices are generated. In the present configuration, the active piezoceramic
patches are actuated identically for each beam, which results in an equivalent
external force pair F = U proportional to the supplied voltage U and acting in
beam direction at both ends of each beam. The appropriate actuation factor and
other parameters of this elastic formulation have been obtained by suitable FEmodels of the concrete assembly of the beams in Fig. 9. The external virtual power to
extend the Jourdain principle in Eq. 2 is given by (uk () uk (0))U k . If all these
equations are combined, a position dependent linearized matrix equation of the well
known form
Mq u + Dq u + Kq u = f + Bq x,
y = Cq u,
Gq u = 0
(5)
can be derived. The mass matrix Mq , damping matrix Dq , stiffness matrix Kq are
position dependent, which is indicated by the subscript q of the drive position vector.
x denotes the electric voltages applied to the individual active rods, f is a general force
excitation, which includes inertia forces, y describes the accelerations of the TCP
platform, which should be decreased by the vibration controller, u are the chosen
finite element nodes and Gq represents the linearized equations of the kinematical
constraints. By choosing some vectors q from the kernel of Gq , e.g. modal shapes
from a modal analysis of Eq. 5, these equations can be simplified to
q z + D
q x,
q z + K
q z = qT f + B
M
q z ,
y=C
u = q z
(6)
q = Cq q .
q = qT Bq , C
q = qT Kq q , B
K
(7)
with
q = qT Dq q ,
q = qT Mq q , D
M
560
These models were set up for the Triglide-structure, but they are not yet fine
tuned and there is ongoing work to compare their frequency response functions with
measured data. Once they fit the data, they can ease simulation studies for control
performance and enhance the control layout over the workspace of the parallel robot
structure.
4 Control
4.1 Objectives & Strategies
The objective of the implementation of smart structures in parallel robots is the
vibration reduction at the end effector of the robot. However, the vibration behavior
of Triglide changes during operation. Several influences like the position of the
end effector and changing loads are the sources of such variations. In contrast to
the changing loads, the position of the end effector is known exactly. This fact
is used to improve the control. Instead of synthesizing a single controller for the
entire workspace, several controllers are set up to take control in their region of the
workspace. The advantage of this concept called Robust Gain-Scheduling is that each
of the several controllers can be less robust and therefore gain more performance
than a single one [3].
The challenges of this control strategy, e.g. the workspace segmentation, the
system identification and the controller switching process are addressed in the
following sections.
4.2 Control Environment
An effective vibration control requires the measurement of the effector vibrations.
Therefore, the integration of an accelerometer in the end effector is a common way to
get information about the oscillations. The oscillations have to be controlled in three
translational degrees of freedom, thus a triaxial sensor is used. The orientation of this
sensor is parallel to the corresponding axes of the robots coordinate system. Due to
the setup of the Triglide structure, the end effector platform has always the same
orientation. Complex transformations of sensor data into the machines coordinate
system are not necessary.
The control architecture of Triglide is a proprietary development of the Collaborative Research Center 562. The hardware components of this architecture are a
central control PC with QNX 6.2 real-time operating system [15] and several digital
signal processor (DSP) boards; all connected via IEEE 1394 (FireWire) bus. A single
DSP board is called a node. A node and its software are task specific and managing
the A/D- and D/A-conversion of signals and the digital I/O. A structural control
node is implemented in the Triglide architecture that converts the signals from the
accelerometer from analog to digital and the control variables, the piezo voltages,
from digital to analog domain.
On the central PC a middleware called MIRPA-X and the industrial application
protocol (IAP) are managing the real-time and the transfer of data over FireWire.
MIRPA-X and IAP are developments of the Collaborative Research Center [20].
561
Different controller processes run on the PC. MIRPA-X is responsible for the correct
scheduling in the system cycle which is 1 kHz. Within each cycle step data is read from
the nodes via bus, processed by the real-time threads and sent back to the nodes.
4.3 Workspace Analysis and Segmentation
The system behavior of the parallel robot Triglide depends on the robots end
effector position. Hence a workspace discretization is advisable to get a comprehensive survey of the system dynamics. For this purpose the workspace is divided into
several domains. A system identification for one representative operating point of
each domain is performed.
This approach is motivated by the intention of designing a single controller for
each operating point to control the accordant workspace region. Therefore, the
system properties of the structure in an operating point should hardly deviate from
the properties in other points of the domain. Otherwise the corresponding controller
would have to be designed more robust than necessary sacrificing system performance. Hence, a correlation between the domain size and the system properties
variation should be established. A finer segmentation is recommended for workspace
regions with higher property gradients.
4.3.1 Initial Segmentation
The robot system comprises three translative degrees of freedom, whereby a
workspace subdivision in three directions of space is required. It is furthermore
assumed that the system properties in y- and z-direction vary significantly. A
segmentation is done only in these two directions. This approach is caused by the
fact that the robots stationary basic structure is significantly stiffer than the movable
lightweight part. The assumption that the interesting low-frequency eigenvalues
are induced by the movable structure parts is admissible. According to this the
considered eigenfrequencies are related to the relative positions of the three links to
each other and not to the position of the movable structure in reference to the basic
structure. For example a simultaneous movement of the three linear motors with
the same velocity and direction does not have any effect on the systems vibration
behavior.
The initial segmentation divides the workspace into 25 domains. They are
arranged in a rectangular and nearly equidistant grid. For each domain a system identification at the central operating point is performed. A summary of the identified
system parameters is given in Figs. 12 and 13, where the frequency characteristic of
the relevant eigenvalues are plotted in the y-z-plane indicating the effectors position.
The figures show that both frequencies change quite smoothly. As expected there are
no discontinuities in these characteristics and thus no discontinuities in the system
behavior.
4.3.2 Optimized Workspace Segmentation
A new workspace segmentation is inferred based on the measured system parameters
which depend on the effectors position, (see Fig. 14). Thereby, the operating point
density is correlated to the gradient of the eigenfrequencies, whereby the density of
562
Fig. 12 Frequency
characteristic of first
eigenvalue in y-z-plane
34
24
36
23
33
34
f [Hz]
21
30
10
5
26
15
24
200
30
18
14
9
3
31
19
20
28
29
17
13
12
28
16
27
11
100
-400
-200
0
-100
y [mm]
Fig. 13 Frequency
characteristic of second
eigenvalue in y-z-plane
32
22
25
32
-200
25
200
400
26
z [mm]
64
65
25
24
23
62
22
21
60
f [Hz]
20
15
55
14
10
19
58
17
12
56
16
50
200
18
13
9
5
60
11
7
3
54
100
-400
-200
0
-100
y [mm]
-200
200
400
z [mm]
200
z [mm]
100
0
-100
-200
-200
-100
0
y [mm]
100
200
52
563
the operating points increases with decreasing z-values. This inhomogeneity allows
a reduction of operation points from 25 down to 10. Assuming that the system
properties only depend on the y- and z-position, the x-position is not incorporated in
the following investigation.
The optimization algorithm checks the distances of each considered point from all
operating points of the given workspace. Afterwards the considered point is associated with the closest operating point. The algorithm yields a workspace segmentation
elucidated in the diagram of Fig. 15. The honeycomb structure of the workspace
segmentation is displayed in the figure. Thus, the corners of the regions are obtuse
angled and the number of bordering regions in any point cannot exceed three. So,
the maximal number of involved controllers in any point is three, too.
4.4 System Identification
An important step in the controller development is the system identification. The
controller synthesis depends on the underlying model of the plant. Only if this plant
meets the demands of the control concept, the controller will achieve the desired
performance. The optimized workspace segmentation includes 10 regions with one
operating point each. For each point, measurements have to be carried out, followed
by a system identification. The Triglide plant has three control inputs, for the three
active links, and three measurement outputs, the three translational accelerations.
Hence, nine structural transfer functions have to be determined.
The system identification is automatized in order to save time and to simplify
the procedure. All nine transfer functions of a single operating point are measured
simultaneously by exciting all inputs with bandlimited, uncorrelated noise. The noise
is generated by a superposition of several sine waves with frequencies in the desired
bandwidth and a random phase. The final signals are characterized by flat power
spectra with steep boundaries. The system identification is performed in the target
QNX system. This is possible since all needed modules like signal generation, signal
processing, and system identification are implemented in the C/C++ programming
language [1]. The algorithm used for identification is based on subspace identification
methods [10] enabling an identification of a system with order 30 in less than 50 s.
564
(8)
(9)
for each operating point i. In the following the plant model is called
Ai Bi
.
Gi =
Ci Di
(10)
(12)
The vectors w and z are called performance inputs and outputs. The specification of
objectives is done with so-called weighting functions W. The inverse of a weighting
function describes the desired singular value function of the weighted transfer
565
function. The transfer matrix RS is considered for the limitation of the control
variable. It describes the relation between disturbances entering the loop at the
plants output and the control variable. The weighting scheme which is appropriate
for the synthesis is shown in Fig. 18.
The transfer matrices SG and RS are weighted by the matrices W SG and W RS .
The transfer matrix Pwu is given to the control synthesis algorithm. The inputs and
outputs of it are as follows:
zy
wd
zu = Pwu wr
e
u
(13)
Like the system identification, the control synthesis is ported onto the target
QNX machine and integrated into the real-time environment, too [2]. This guarantees a fast and reliable synthesis chain on the target system. The structural control
on Triglide works in the bandwidth from zero to 100 Hz at a control cycle of 250 Hz.
An example of the control performance is given in Fig. 19. The frequency response
function from link 2 to the acceleration in z-direction is shown there for a single
operating point. The curve of the controlled plant (SG) is significantly smaller than
the one of the uncontrolled (G).
4.6 Controller Switching
An especial controller for each domain of the segmented Triglide workspace has to
be designed. In consequence, a possibility to change the active controller is necessary
in order to suppress vibrations during repositioning of the end effector effectively.
A sudden switching between two controllers may cause discontinuities that excites
the plant. Therefore, a method was developed that provides a smoother switching
process.
566
0
Magnitude in dB
-10
-20
G
1, z2
SG1, z2
-30
Phase in degree
-40
180
135
90
45
0
-45
-90
-135
-180
10
20
30
40
50
60
70
80
90
100
Frequency in Hz
4.6.1 Method
The method creates a combined controller during the switching process. This controller R consists of the two involved controllers R1 and R2 and a scaling block at
the output of each (see Fig. 20). The scaling blocks have a multiplicative system
behavior. The factors are 1 for scaling R1 and 2 for scaling R2 . Both i -parameters
are restricted to provide a convex combination for the mixed controller R.
Figure 21 shows an exemplary characteristic of both i -parameters and the
resulting control variable u. In this example the switching procedure starts at t = 0.3 s
and ends at t = 0.35 s. Initially, the signals u and u1 are identical. After 50 ms which
is the time slot for the switching process u and u2 are identical. This continuous
switching method enables the smooth transfer from u = u1 to u = u2 .
4.6.2 Verif ication of Stability
A proposal of the Robust Control theory is used to verify the stability of this
switching strategy. This approach analyzes the effects of an uncertainty ( j) on
567
i[ ]
0.5
0
0.275
0.3
0.325
t[s]
0.35
0.375
0.4
u
u
1
u2
U[V]
0.2
0
-0.2
-0.4
0.275
0.3
0.325
t[s]
0.35
0.375
a system G( j), which works evidently stable under undisturbed conditions. The
approach is based on the model depicted in Fig. 22 and proves the robust stability
by means of two criteria [17, 26]:
1. The undisturbed controlled system Gwz ( j) is stable (for ( j) = 0)
2. The following inequation has to be fulfilled for all R+ : max (G ( j))
max (( j)) 1
A description of the uncertainty ( j) is needed in order to use these criteria. This
block should be a function of both switching parameters 1 and 2 for the given
switching process.
Since the sum of both factors must always be 1, the number of switching parameters can be reduced to a single one. Again, the range of this new parameter is
0 1.
1 =
(14)
2 = 1
(15)
568
(16)
R = R1 + (1 )R2
(17)
R = 1 (R1 R2 ) + R2
(18)
(19)
(20)
The variable nmin represents the minimal number of intermediate steps, vTC Pmax
the maximal TCP velocity and fCC the sample frequency of the control cycle.
For the calculation of the controller participation the structural control always has
to know the accurate position of the TCP. Hence, the algorithm is fully aware
of the coordinates of the TCP xTC P , yTC P , zTC P and of the n OP operating points
x OP_i , y OP_i , z OP_i , i = 1, 2 . . . n OP at any time.
569
570
The computation consists of two steps. First, the squared distances between TCPposition and all operating points are calculated:
b i_sq = (xTC P x OP_i )2 + (yTC P y OP_i )2
+(zTC P z OP_i )2
b i = b i_sq
(21)
(22)
The squared distances are used instead of the real distances in order to save computation time. This is valid for the sorting of the distances since the quadratic function
is strictly monotonic increasing. Up to three controllers are involved simultaneously
due to the honeycomb workspace structure (see Fig. 15). Consequently, the three
smallest values of distances and the indexes of the corresponding operating points are
buffered. Relative distances in relation to the closest operating point are calculated
next in order to calculate the weighting factor of each controller. As an example
(Fig. 26b) the computation for the operating points 6 and 3 is as follows:
arel63 =
b 6_sq
b 6_sq + b 3_sq
(23)
This value provides the possibility to assess the relationship between the considered
operating points and the current TCP-position. For example, the TCP is located
exactly between the operating points 6 and 3 for a value arel63 = 0.5. For arel63 < 0.5
it is closer to operating point 6 and for arel63 = 0 it is in point 6. The resulting process
for defining criteria for the switching regions is quite simple. At least, constraints
for the relative distances have to been determined. As mentioned above, a relative
distance of arel = 0.5 means that the TCP is at an edge of a region and an arel = 0
571
1.5
one involved
controller
two involved
controllers
one involved
controller
0.5
6
3
0
0
0.2
0.4
0.6
0.8
arel63
Fig. 29 Illustration of the controller participation 4 of controller 4 in spatial (a) and oversight (b)
view
572
of operating point 4. Furthermore, a defined switching area and a wide area with
an 4 of 0 is noticeable. The crossovers in the switching areas are smooth and free
of discontinuities. Additionally, Fig. 30 depicts all controller participations. In this
three-dimensional figure the crossing controller participations can be observed as
seen before in Fig. 28.
5 Conclusion
The article presents solutions for the major challenges of introducing smart-structure
technology in robotics. The steps to design active rods with distributed surface
mounted piezoceramic actuators are discussed and an implementation is shown.
Boundary conditions for the design are obtained using dynamic models of parallel
robots.
Rigid and elastic body models are required for the development of control algorithms. The rigid model has been successfully implemented in a real-time computed
torque control algorithm to increase the obtainable accelerations. The elastic body
model is still under development and can be used as plant for a model based vibration
control approach.
The challenge of vibration control for a system with varying modal parameters
is addressed in this paper by proposing an adaptive workspace segmentation and
an algorithm for smooth controller switching. Based on an experimental system
identification, an optimal working point distribution is defined, to get a feasible
description of the behavior with a small amount of points. Switching rules are
defined for the controllers in the operating points in order to realize a control that
is continuously active during movements of the parallel robot. The stability of the
switching process is proved with the use of the Robust Stability formulation.
The developed hardware and methods are experimentally validated in the test
platform Triglide of the Collaborative Research Center SFB 562 Robotic Systems
for Handling and Assembly.
573
References
1. Algermissen, S., Keimer, R., Rose, M., Breitbach, E., Monner, H.P.: Applied robust control for
vibration suppression in parallel robots. In: Proceedings of 22nd International Symposium on
Automation and Robotics in Construction (ISARC), Ferrara, Italy (2005)
2. Algermissen, S., Keimer, R., Rose, M., Monner, H.P.: Robust control for vibration suppression
on parallel robot triglide. In: Proc. of Adaptronic Congress 2006, Gttingen, Mai (2006)
3. Algermissen, S., Rose, M., Keimer, R., Breitbach, E.: High-speed parallel robots with integrated
vibration-suppression for handling and assembly. In: 11th Annual International Symposium on
Smart Structures and Materials. SPIE, San Diego, Kalifornien, USA, Mrz (2004)
4. Apkarian, P., Gahinet, P., Becker, G.: Self-scheduled H control of linear parameter-varying
systems: a design example. Automatica 31(9), 12511261 (1995)
5. Bachman, S.: Modellierung und bau eines aktiven stabes mit verteilter aktuatorik zur
schwingungsunterdrckung bei parallelrobotern. IB 131-205-45, DLR, Braunschweig (2005)
6. Breitbach, E., Algermissen, S., Keimer, R., Rose, M., Stachera, C.: Adaptive tools in parallel
robotics. In: Last, P., Budde, C., Wahl, F.M. (eds.) Robotic Systems for Handling and Assembly,
2nd International Colloquium of the Collaborative Research Center 562. Fortschritte in der
Robotik, vol. 9, pp. 203219. Shaker Verlag, Braunschweig, Germany, 1011 May 2005
7. Breitbach, E., Keimer, R., Rose, M., Sachau, D.: An adaptronic solution to increase efficiency
of high speed parallel robots. In: Proceedings of 12th International Conference on Adaptive
Structures and Technologies (ICAST). College Park, Maryland, USA (2001)
8. Budde, C., Last, P., Hesselbach, J.: Workspace enlargement of a triglide robot by changing
working and assembly mode. In: Proc. of the IASTED International Conference ROBOTICS
AND APPLICATION, pp. 244248. Cambridge, USA (2005)
9. Dietz, S.: Vibration and fatique analysis of vehicle systems using component modes. FortschrittBerichte VDI, Reihe 12: Verkehrstechnik/Fahrzeugtechnik, 401 (1999)
10. Favoreel, W., Sima, V., Huffel, S.V., Verhaegen, M., Moor, B.D.: Slicot working note 1998-6:
Subspace model identification of linear systems in slicot. Technical report, European Community
BRITE-EURAM III Thematic Networks Programme NICONET (1998)
11. Gahinet, P., Nemirovski, A., Laub, A.J., Chilali, M.: LMI Control Toolbox Users Guide. The
MathWorks, Inc. (1995)
12. GmbH, I.: Simpack the High-end Simulation Software. Homepage (2008)
13. Gtting, H.C., Schtze, R.: Adaptive lightweight cfrp strut for active vibration damping in truss
structures. J. Intell. Mater. Syst. Struct. 7(4), 433440 (1996). doi:10.1177/1045389X9600700408
14. Hesselbach, J., Budde, C., Maa, J., Breitbach, E., Rose, M.: Dynamic performance enhancement
of a hexa-parallel-robot using the computed torque approach. In: Proceedings of Mechatronics
& Robotics, pp. 10061011 (2004)
15. Krten, R.: Getting Started with QNX Neutrino 2: A Guide for Realtime Programmers. PARSE
Software Devices (2001)
16. Nethery, J., Spong, M.: Robotica: a mathematica package for robot analysis. IEEE Robot.
Autom. Mag. 1(1), 1320 (1994). doi:10.1109/100.296449
17. Raisch, J.: Mehrgrenregelung im Frequenzbereich. R. Oldenbourg Verlag GmbH, Mnchen
(1994)
18. Rugh, W.J., Shamma, J.S.: Research on gain scheduling. Automatica 36(10), 14011425 (2000)
19. Schwertassek, R.: Dynamik Flexibler Mehrkrpersysteme. Vieweg Verlag, Braunschweig (1999)
20. Steiner, J., Kohn, N.: Universal communication architecture for high-dynamic robot systems
using qnx. In: Proc. of ICARCV, Kunming, China (2004)
21. Wang, X.: Dynamic modeling, experimental identification, and active vibration control design of
a smart parallel manipulator. Ph.D. thesis, Department of Mechanical and Industrial Engineering, University of Toronto, Toronto (2005)
22. Wang, X., Mills, J.: A fem model for active vibration control of flexible linkages. In: Proc. of
IEEE International Conference on Robotics and Automation (ICRA), vol. 5, pp. 43084313
(2004)
574
23. Wang, X., Mills, J.K.: Active control of configuration-dependent linkage vibration with application to a planar parallel platform. In: IEEE International Conference on Robotics and
Automation (ICRA), pp. 43274332 (2005)
24. Wang, X., Mills, J.K.: Dual-modal control of configuration-dependent linkage vibration in a
smart parallel manipulator. In: IEEE International Conference on Robotics and Automation
(ICRA), Orlando, pp. 35443549 (2006)
25. Zhang, X., Mills, J., Cleghorn, W.: Dynamic modeling and experimental validation of a 3-prr
parallel manipulator with flexible intermediate links. J. Intell. Robot. Syst. 50(4), 323340 (2007)
26. Zhou, K.: Essentials of Robust Control. Prentice Hall (1998)
27. Zhou, K., Doyle, J., Glover, K.: Robust and Optimal Control. Prentice Hall (1996)