Int J Adv Manuf Technol (2013) 64:447–458

DOI 10.1007/s00170-012-4030-6

ORIGINAL ARTICLE

A simulation platform for optimal selection of robotic belt

grinding system parameters
Shuihua Wu & Kazem Kazerounian & Zhongxue Gan &
Yunquan Sun

Received: 23 November 2011 / Accepted: 24 February 2012 / Published online: 16 March 2012
# Springer-Verlag London Limited 2012

Abstract Robotic belt grinding is an effective process for material removal to meet dimensional accuracy requirements
removing material from geometrically complex workpieces. on workpieces.
However, due to the relatively low stiffness of the system, the
grinding quality is prone to inaccuracies caused by system Keywords Robotic belt grinding . Conformance grinding
dynamics. In order to control the quality of the grinding
process, a profound understanding of the system is required.
This paper presents a platform for comprehensive modeling 1 Introduction
and simulation of the robotic belt grinding system. The system
kinematics model is based on the CAD model of the work- Grinding broadly refers to the final machining or finishing
piece in composition with robot kinematics. The dynamics process that removes material by utilizing hard abrasive par-
model is a comprehensive combination of the dynamics of the ticles as the cutting medium. As an important machining
robot, the grinder, and the interaction between the grinder and method, grinding has been investigated for decades. Studies
the workpiece. A material removal model of the grinding have focused on the subjects of grinding force, energy, thermal
process, which can adapt to workpieces with complicated model, friction, chatter, and material removing process, among
shapes, is also developed and presented. The system simula- others [1–3].
tion shows that optimal selection of key control parameters of The grinding of a workpiece with free-form geometry
the grinder and proper selection of robot control strategies can typically requires manual intervention. The manual grinding
efficiently suppress chatter in the grinding process. Further- process is time and labor consuming and exposes operators to
more, having the ability to predict material removal rate, the high noise levels and metal dust environments. In recent years,
comprehensive simulation platform is also demonstrated to be several groups have worked on the development of robotic
a strong tool in selecting the grinding process key parameters, grinding systems [4–6]. In these systems, belt grinder was
namely, robotic velocity and contact force, for the control of selected as the machining tool because of its superior material
removal ability and its lower overall system stiffness.
S. Wu (*) : K. Kazerounian However, due to the relatively low stiffness of the robotic
Department of Mechanical Engineering, belt grinding system, the grinding quality is prone to inac-
University of Connecticut, curacies caused by system dynamics. Chatter, which has
Storrs, CT 06269, USA
e-mail: [email protected]
critically adverse effects on grinding accuracy and surface
finish, has been extensively studied for conventional grind-
K. Kazerounian
ing processes, such as surface grinding and plunge grinding
e-mail: [email protected]
[2, 7]. On the other hand, for belt grinding, research on
Z. Gan : Y. Sun chatter has been rarely found in literature. Sun [6] has
InterSmart Robotic Systems Co. Ltd, revealed that the ratio between the robotic grinding system
Langfang, Hebei, China 065001
stiffness and the contact stiffness greater than 1 is sufficient
Y. Sun for a belt grinding process to be stable. However, the study
e-mail: [email protected] has not fully considered the dynamic model of the grinder.
448 Int J Adv Manuf Technol (2013) 64:447–458

In addition, the analysis of the effect of the system param- This paper presents a simulation platform for robotic belt
eters and control strategies on chatter remains absent. grinding systems. In this platform, comprehensive dynamics
In manufacturing of mechanical components with complex models of the robot, the grinder, and the interaction between
geometries, such as turbine blade manufacturing, there are the grinder and the workpiece are fully included. A material
frequently strict dimensional accuracy requirements on the removal model of the grinding process, which can adapt to
workpiece. Before a workpiece is finalized, measurement workpieces with complicated shapes, is also developed and
machines are usually adopted to measure its dimensional error. presented. Simulation studies are conducted for the cases that
As an example, an error map of a turbine blade is shown in the workpiece–grinder contact force has to change rapidly to
Fig. 1. The color variation on the turbine blade surface stands remove different amount of material from different regions to
for the different amount of material that should be removed to meet dimensional accuracy requirements on workpieces. The
achieve the dimensional accuracy of the workpiece. Confor- simulation results show that high-magnitude and high-
mance grinding, which refers to a grinding process that can frequency chatters can be generated in these cases. However,
remove specific amount of material from specific regions of a if the key control parameters of the grinder and the control
workpiece, is typically employed as the final machining pro- strategies of the robot are properly selected, the chatters can be
cess to locally adjust the dimension of workpieces to fit into the efficiently suppressed. Furthermore, having the ability to pre-
accuracy tolerance. Currently, no automatic system is available dict material removal, this presented platform is also demon-
that is capable of performing such a process. strated to be a strong tool in selecting the grinding process key
The material removal of the belt grinding process is affected parameters, namely, robotic velocity and contact force, for the
by grinding process parameters, such as belt type, belt speed, control of material removal to achieve the goal of conformance
workpiece feed velocity, and contact force between the contact grinding.
wheel and the workpiece. For workpieces with free-form sur- In this paper, the kinematics model of the system is
face, Zhang et al. [8] proposed a two-step procedure to model introduced in Section 2. The dynamics models of the system
the material removal: (1) solve the pressure distribution in the are analyzed in Section 3. The material removal model for
contact area between the workpiece and the contact wheel and the belt grinding process is presented in Section 4. Case
(2) combine the solved pressure distribution with other grind- studies and discussions are presented in Section 5.
ing parameters to estimate the material removal. In the local
model, Zhang and Cabaravdic et al. [8, 9] suggested that the
force (pressure) distribution in the workpiece–wheel contact 2 Kinematics model
area is calculated locally. However, the performance of the
model is not always sufficient, since obtaining global “best-fit” A CAD model-based methodology is adopted in our study
of the material removal could be infeasible, especially for to generate the robotic grinding path and describe the sys-
workpieces with complicated free-form geometries. tem kinematics. Targets (desired contact points on the work-
Simulation of the belt grinding process can be used for piece surface) are created based on the CAD model of the
material removal prediction. It is of great help in selecting workpiece, as shown in Fig. 2a. The local (body) coordinate
grinding process parameters to perform conformance grinding systems of the targets are selected in such a way that one
to achieve the dimensional accuracy on workpieces. Such axis is perpendicular to the workpiece surface and the sec-
simulation research has been reported by Zhang et al. [8] ond axis is parallel to the direction along which the surface
and Ren et al. [10, 11]. However, in these presented simula- has the minimum principle curvature. This axis selection
tions, the dynamics of the system, which has significant ensures optimal contact between the workpiece and the
effects on the grinding process, has not been considered. contact wheel of the grinder. The robot is then controlled
by closing the kinematics loop:

T ¼ T0  Wobj  Tgt ð1Þ

where, as shown in Fig. 2b, T is a 4×4 homogeneous

position/orientation matrix demonstrating the pose of the
tool with respect to (w.r.t.) the base frame of the robot and
can be calibrated for a specific system setup; T0 is the pose
of the robot end effector w.r.t. the base frame of the robot
and can be controlled by the robot controller; Wobj is the
pose of the workpiece w.r.t. the end effector frame T0 and
can be calibrated for a specific system setup; and Tgt is the
Fig.1 Geometric error map of a turbine blade surface pose of the created targets w.r.t. the workpiece frame Wobj.
Int J Adv Manuf Technol (2013) 64:447–458 449

3.2 Belt grinder dynamics

A belt grinder for robotic grinding systems is shown in

Fig. 3, where the contact wheel is the tool that makes contact
with the workpiece; the tension wheel is used to set and
maintain the tension in the belt. Two air cylinders are
included (as shown in Fig. 3b), one for the belt tension
and the other one for passive compensation of the contact
force (a large contact force will push the “movable” grinder
head back along the slider to avoid grinding burn from
excessive contact force).
The belt grinder dynamics in the horizontal direction,
which plays a significant role in the grinding process, is
modeled in Fig. 4. The characteristics of this model are
described below.
& The grinder base mounted on the floor is considered
rigid
& m1 is the mass of the grinder head
& C1 is the damping coefficient of the air cylinder for the
contact force compensation
Fig.2 Path generation and system kinematics. a CAD model of a
typical workpiece and the desired kinematic tool path. b The kinemat-
ics relation through position matrices of the robot end effector, the
workpiece, and the grinder

During the grinding process, based on the feedback of the

robot joint positions (q), the relative pose between the
workpiece and the tool (contact wheel) can be calculated
with Eq. (2):
Tool
¼ T1  T0  Wobj  Tgt
Target T ð2Þ
T0 ¼ forward kinðq1 ; q2 ; :::; q6 Þ
Tool
where Target T is a 4×4 homogeneous position/orientation
matrix demonstrating the pose of the target on the workpiece
w.r.t. the tool (contact wheel) frame T; foward_kin is the
robotic forward kinematics process; and q (i.e., q1 ; q2 ; :::; q6)
are the six joint positions of the robot.

3 Dynamics model

3.1 The robot system dynamics

The dynamics of robotic system has been well-studied [12,

13]. Neglecting the friction, the equation of motion of the
robot can be given by:


MðqÞ q þ h q; qÞ þ gðqÞ ¼ t þ JT f ð3Þ

where q are the joint positions; M(q) is the symmetric


positive definite inertia matrix; hðq; qÞ represents the cen-
trifugal and coriolis effects; g(q) represents the gravity load;
t is the joint-driving torques; J is the Jacobian matrix; and f Fig. 3 Belt grinder and its CAD model. a Belt grinder. b CAD model
is the external force/torque exerted onto the robot. of the belt grinder
450 Int J Adv Manuf Technol (2013) 64:447–458

Fig. 4 Physical model of the

belt grinder

& Fc is the interaction force between the grinder base & Ks is the stiffness between the grinder head and the stop
and the grinder head exerted by the air cylinder for on the grinder base
the contact force compensation; Since the length of & X1 and X2 are the displacements of the grinder head and
the cylinder is much larger than the dynamic shift of the tension wheel components, respectively
the grinder head, the force is considered constant at & Dw is the displacement of the contact point (target) on
all times the workpiece w.r.t. the contact wheel in horizontal
& m2 is the mass of the tension wheel component direction, which can be solved with Eq. (2)
& C2 is the damping coefficient of the air cylinder for belt & Fh is the contact force between the workpiece and the
tension contact wheel; Fg is the grinding force; They both will
& Ft is the interaction force between the grinder head and be discussed in Section 3.3
the tension wheel components exerted by the air cylin- & Belt section AE is horizontal and ED is vertical
der for belt tension and is considered constant at all
times The equation of motion of the grinder can be given by
& Kw is the stiffness of the contact wheel Eq. (4):

  
  
m1 X 1 þ C1 X 1 þ C2 X 1  X 2 ¼ Fc þ Ft  Fb ðtÞ  sin a  Fs ðtÞ  Fh ðtÞ

   ð4Þ
m2 X 2 þ C2 X 2  X 1 ¼ Fb ðtÞ  ðsin a þ sin bÞ  Ft

where Fb(t) is the tension force of the belt; a and β are the
acute angles between the vertical direction and the belt
sections BC and CD, respectively (as shown in Fig. 4);
and Fs(t) is the force between the stop (see Fig. 4) and the
grinder head, which can be obtained from Eq. (5):

   
sin a
Fs ðtÞ ¼ max 0; Fc þ Ft  Ft þ Ks  X1 ðtÞ
sin a þ sin b
ð5Þ
The tension force of the belt Fb can be calculated with
Eq. (6):

Ft
Fb ðtÞ ¼ þ Kb  ΔSðtÞ ð6Þ
sin a þ sin b Fig. 5 Contact between two cylinders
Int J Adv Manuf Technol (2013) 64:447–458 451

Fig. 6 Procedure of material

removal modeling

where Kb is the belt stiffness and ΔS(t) is the change of the 2. The deformation which occurs during the workpiece–
belt length, which can be approximated by Eq. (7): wheel contact is much less than the radius of the contact
wheel and the local radius of the workpiece.
ΔSðtÞ ¼ ΔSAB ðtÞ þ ΔSBC ðtÞ þ ΔSCD ðtÞ þ ΔSDE ðtÞ þ ΔSEA ðtÞ For the contact between cylinders, Johnson [14] pre-
ΔSAB ðtÞ   max ð0; X1 ðtÞ þ Dw ðtÞÞ  sin g sented a set of equations based on Hertz method (Eq. (8))
ΔSBC ðtÞ  ðX1 ðtÞ  X2 ðtÞÞ  sin a
ΔSCD ðtÞ  X2 ðtÞ  sin b
and the analysis of non-Hertzian normal contact (Eq. (9)).
ΔSDE ðtÞ  X1 ðtÞ  sinð0Þ ¼ 0 These equations can precisely estimate the contact area, the
ΔSEA ðtÞ   maxð0; X1 ðtÞ þ Dw ðtÞÞ  sinð p2 Þ ¼  maxð0; X1 ðtÞ þ Dw ðtÞÞ pressure at the central contact point (point O in Fig. 5), part
ð7Þ compression, and interaction force.
2 2
1
¼ þ R12 ; E1 ¼ 1u
1
E1 þ E2 ; a ¼
1 1u2 2 4PR
where ΔSAB , ΔSBC , ΔS CD , ΔS DE , and ΔS EA are the R R1 ( pE
2 1=2 ð8Þ
pHertz ðxÞ ¼ p a2 ða  x Þ ; j xj < a
2P 2
changes of the belt length in belt sections AB, BC, CD,
DE, and EA, respectively; γ is the acute angle between belt 0 ; j xj  a
section AB and the vertical direction as shown in Fig. 4.

3.3 Workpiece–grinder interaction dynamics 1u1 2

d¼P f2 lnð4R1 =aÞ  1g ð9Þ
pE1
In the horizontal (normal) direction, the workpiece–grinder
interaction involves modeling the contact mechanics be-
where, as shown in Fig. 5, in robotic belt grinding applica-
tween the workpiece and the grinder (contact wheel). In
tions, R1 and R2 are the radiuses of the contact wheel and the
robotic belt grinding applications (such as turbine blade
local radius of the workpiece, respectively; R* is the equiv-
grinding), the workpiece–wheel contact can be approximat-
alent radius; v1 and v2 are the Poisson’s ratios of the contact
ed by the contact between two cylinders based on the
wheel and the workpiece, respectively; E1 and E2 are the
following facts in the grinding process:
Young’s modulus of the contact wheel and the workpiece,
1. The grinding path is well planned so that when the respectively; P is the contact pressure which can be further
workpiece–wheel contact occurs during the grinding formulated as FWh (Fh is the contact force; W is the thickness
process, the workpiece surface has minimum princi- of the contact wheel); a is half of the width of the contact
pal curvature along the axial direction of the contact area; pHertz(x) is the pressure at the point with an X coordi-
wheel. nate of x (the central contact point O is assumed to be the

Fig. 7 Discrete-time material

removal process
452 Int J Adv Manuf Technol (2013) 64:447–458

Fig. 8 Calculation of pressure distribution. a Contact around central contact point O. b Pressure calculated with FEA software Abaqus. c Pressure
calculated with the Hertz method

origin); and δ is the part compression occurring in the model can greatly reduce the calculation time and also better
contact situation and equals to max (0,X1(t)+Dw(t)) (X1 adapt to workpieces with complicated shapes.
and Dw are shown in Fig. 4); In obtaining Eq. (9), the
workpiece is considered rigid compared to the contact wheel 4.1 Material removal process
that has much lower stiffness.
In the tangential direction, the grinding force Fg (shown The calculation of material removal is mathematically formu-
in Fig. 4) is considered proportional to the normal contact lated as an integration of the material removal rate. In discrete-
force (Fh) with a constant coefficient, according to the time cases, the robotic grinding process can be considered as
grinding process model [15]: the robot running the targets through the contact wheel one by
one with certain velocities and contact forces. When one target
Fg ðtÞ ¼ Kg & Fh ðtÞ ð10Þ
is reached, contact occurs at the neighbor targets nearby. As
shown in Fig. 7, contact between the contact wheel and target i
occurs from the time when the robot reaches target i−m to the
4 Material removal model time when the robot reaches target i+n. Therefore, the total
material removal at target i can be formulated as:
 
For the modeling of material removal, we proposed a material X
n Tgtiþjþ1  Tgtiþj 
MRðTgti Þ ¼ rðVw;iþj ; Pd ðTgt iþj ; Tgt i Þ; Cg Þ   
removal model, which is introduced in more detail in [16]. This Vw;iþj 
j¼m
model consists of a superposition sub-model and a shape-
ð11Þ
dependent sub-model (as shown in Fig. 6), following the two-
step procedure suggested by Zhang et al. [8]. In particular, this where MR(Tgti) is the total material removal at target i; r is the
presented model concentrates on the analysis of the two key instant material removal rate, which is a function of the grind-
grinding parameters: the robotic (workpiece) velocity and the ing parameters; Vw, i+j is the robotic velocity from target i+j to
contact force between the contact wheel and the workpiece. target i+j+1; Pd(Tgti+j, Tgti) is the pressure at target i when
Other parameters, such as the belt type and the belt speed, target i+j is reached; Cg is a constant that stands for the effect
are considered constant in a predefined grinding process. Com- of other grinding parameters; and ∥Tgti+j+1 −Tgti+ j∥ is the
pared to the existing models, the proposed material removal distance between target i+j+1and target i+j.

Fig. 9 Performance of superposition model (R1_wheel 050 mm, R2_workpiece_local 067.9 mm, F050 N, W025.4 mm). a Pressure calculated with FEA
software Abaqus. b Pressure calculated with superposition model. c Pressure errors (psuperposition −pFEA)
Int J Adv Manuf Technol (2013) 64:447–458 453

4.2 Superposition model for three-dimensional contact cases, the Hertz method cannot
predict the variation of the pressure along the contact lines (e.g.,
Finite element analysis (FEA) is the traditional method used to axis Y in Fig. 8) which are parallel to the axial direction of the
solve contact problems for pressure distribution in the contact contacted cylinder.
area. However, as an iterative algorithm, it is very time- The proposed superposition model formulated below intro-
consuming and cannot be employed in real-time applications. duces a term to compensate for the errors in the Hertz method:
On the other hand, the Hertz method is not computationally
psuperposition ðx; yÞ ¼ pHertz ðxÞ & ð1 þ cðyÞÞ
expensive and can also precisely estimate the contact area and
 2y  ð12Þ
cðyÞ ¼ R  polynomial W 
k 
the pressure at the central contact point with Eq. (8). However,
where, psuperposition(x, y) is the pressure at contact point (x, y);
pHertz(x) is the pressure at the contact point calculated with the
Hertz method indicated by Eq. (8); c(y) is the error compen-
sation function; R* is the equivalent radius; W is the thickness
 
of the contact wheel; and Polynomialð 2y 
W Þ is a polynomial of
2y
degree 5 in the absolute value of W .
It is shown [16] that compared to the pressure distribution
calculated by FEA software (shown in Fig. 9a), the superpo-
sition model result (shown in Fig. 9b) only has an average
relative error (formulated as absolute value of the ratio
Pressure error
Pressure at the center contact point ) of 4.8% with standard deviation
of 4.2%.

Fig. 10 Material removal modeling (F0100 N, Vw 010 mm/s). a Fig. 11 Material removal prediction using shape-dependent model. a
Using shape-dependent model. b Using local model. c Modeling errors F050N, Vw 010 mm/s. b F0100N, Vw 020 mm/s
454 Int J Adv Manuf Technol (2013) 64:447–458

4.3 Shape-dependent model where R is a matrix indicating the instant material removal
rate at contact points in the contact area; KL is a local coeffi-
To estimate material removal for the grinding of geometrically cient describing the material removal ability of the system at a
complex workpieces, we propose a shape-dependent model. certain location, which is determined by the contact situation
This model introduces local coefficients to denote the material (i.e., the local shape of the workpiece); PA is a matrix describ-
removal ability of the system in certain contact situations. ing the pressure distribution throughout the contact area; Cg
These local coefficients essentially give the model the ability stands for the effect of other grinding parameters and is
to adapt to the complex free-form geometry of the workpiece. considered constant; and u and v are the mesh size of the
The shape-dependent model is formulated as follows: discredited contact area in two directions.
0 1 0 1 Integrating Eq. (13) into Eq. (11) yields:
r11 r12 ::: r1v p11 p12 ::: p1v
B r21 r22 ::: r2v C B p21 p22 ::: p2v C  
R ¼ KL  PA  Cg
B
R ¼ B .. .. ..
C
. C
B
PA ¼ B . .. ..
C
.. C
X
n
 Tgtiþjþ1  Tgtiþj 
@ . . . .. A @ .. . A MRðTgti Þ ¼ KLiþj  Pd Tgtiþj ; Tgti  Cg   
. . Vw;iþj 
ru1 ru2 ::: ruv pu1 pu2 ::: puv j¼m

ð13Þ ð14Þ
which can be further formulated in matrix form as follows:

2 3
Pd ðTgt1 ; Tgtmþ1 Þ Pd ðTgt2 ; Tgtmþ1 Þ ::: Pd ðTgtmþ1þn ; Tgtmþ1 Þ 0 ::: ::: 0
6 0 Pd ðTgt2 ; Tgtmþ2 Þ :::: Pd ðTgtmþ1þn ; Tgtmþ2 Þ Pd ðTgtmþ2þn ; Tgtmþ2 Þ 0 ::: 0 7
6 7  ð15Þ
4 ::: ::: 5
0 ::: ::: 0 Pd ðTgthnm ; Tgthn Þ ::: Pd ðTgth ; Tgthn Þ
2 kTgt2 Tgt1 k
3
 KL1 2 3
6 kVw;1 k 7 MRðTgtmþ1 Þ
6 : 7
6 7 6 : 7
6 : 7  Cg ¼ 6
4
7
5
6 : 7 :
4 5
kTgthþ1 Tgth k
 KLh MRðTgt hn Þ
kVw;h k

Fig. 12 Robotic controller. a

Position controller. b Position–
force controller
Int J Adv Manuf Technol (2013) 64:447–458 455

In Eq. (15), the KLi(i01…h) are not totally independent virtual attractive forces for the robotic position and velocity
to each other, since the local coefficient depends on the local errors and can effectively control the robot with non-
shape, and on a smooth surface, local shapes do not change expensive computational cost [18, 19]. In Fig. 12, xd is the
dramatically. The K-mean clustering method is employed to reference robot position in Cartesian space; xa is the feed-
reduce the number of independent KLi based on the local back of the robot position in Cartesian space; q is the robot
contact situations [16]. In the modeling phase, the independent position in joint space; Kp, Kv, and Ki are diagonal matrices
KLs can be solved using experimental data. Once KLi are presenting the P, I, and D parameters for position control,
known, Eq. (15) can be utilized to predict material removal
for future grinding process.
It is experimentally verified [16] on a robotic belt grind-
ing system (consisting of an ABB IRB4400/60 robot, a 6D
ATI force sensor, a belt grinder, and a turbine blade as the
workpiece) that the shape-dependent model has a much better
adaptability to workpieces with complex geometries than the
existing local model [9] which fails to find a good global best-
fit. As shown in Fig. 10, the maximum modeling error, mea-
MR  Actual MR
sured by absolute value of the ratio ModeledActual MR ,
dropped from 17% (using local model) to 2% (using shape-
dependent model). For material removal prediction, the pre-
diction errors of the presented shape-dependent model, mea-
MR  Actual MR
sured by absolute value of the ratio ModeledActual MR , are
below 8% for all the targets along the grinding path, as shown
in Fig. 11.

5 Simulations

A simulation platform was developed to study the dynamic

behavior of the robotic belt grinding system. The robotic
kinematics and dynamics modules in Corke’s Robotics
Toolbox [17] were adopted in the platform, along with the
PUMA 560 parameters, which can be readily modified for
any 6-DOF industrial robot system.
For the control of the robot, two controllers are adopted
and compared to each other. One is the target tracking
position controller [18] shown in Fig. 12a. The second one
steams from the target tracking position–force controller
[19] and is shown in Fig. 12b. These controllers introduce

Fig. 14 System response at spark-in phase. a Robotic position errors

(unit of X, Y, and Z: millimeter; unit of roll, pitch, and yaw: degree). b
Workpiece–wheel contact force (case 1, Fc 020 N, C1 0200 Ns/m; case
2, Fc 0100 N, C1 0500 Ns/m). c Displacement of the grinder head (i.e.,
X1 in Fig. 4; unit, millimeter; case 1, Fc 020 N, C1 0200 Ns/m; case 2,
Fig. 13 Procedure of the robotic belt grinding simulation Fc 0100 N, C1 0500 Ns/m)
456 Int J Adv Manuf Technol (2013) 64:447–458

respectively; J is the Jacobian matrix; hðq; q Þ represents the For the first case, the position controller shown in
centrifugal and coriolis effects; g(q) represents the gravity Fig. 12a was adopted. The workpiece was required to
load; fd is the reference contact force; fa is the contact force feed-in 2 mm (in the horizontal direction) in ten control
feedback; Kpf is the parameter for force control; I is a 6×6 cycles. As shown in Fig. 14, although the robotic position
identity matrix; and S is the 6×6 selection matrix. errors were pretty small for both simulations, high-
The simulation follows the procedure shown in Fig. 13. magnitude chatter occurred when the control parameters of
The robot controller generates the joint-driving torques the grinder were not properly selected (see the simulation
based on the planned (reference) robotic trajectory and the results of case 1 in Fig. 14b, c). By evaluating various
current system states (position, velocity of the robot joints, settings of key control parameters of the grinder on the
the grinder head, and the tension wheel component). The simulation platform, it was found that optimal selection of
contact model solves the workpiece–grinder interaction the key grinder parameters (increasing the support force Fc
forces. The driving torques and the interaction forces are and the damping C1 shown in Fig. 4) could efficiently
put to the system dynamics model. Then, the accelerations suppress the chatter (as shown by the simulation results of
of the system are calculated and used to update the system case 2 in Fig. 14b, c).
states. The total material removal of the simulated grinding For the second case, cyclic inputs were adopted to sim-
process is calculated by the integration of all the instant ulate the requirements of swift contact force change in the
material removal. grinding process. The position controller shown in Fig. 12a
The control cycle of the system is set to 10 ms. Two cases and the position–force controller shown in Fig. 12b were
which are prone to generate vibrations have been studied. adopted respectively to compare their performances in han-
One is at the spark-in phase, when the workpiece is required dling chatters. For the position controller, a cyclic move-
to feed-in to make quick contact with the contact wheel. The ment of workpiece in horizontal direction (shown in
other one is during the grinding process, when rapid move- Fig. 15a) was super-positioned to the grinding path shown
ment of the workpiece in the horizontal direction (direction in Fig. 2a. For the position–force controller, cyclic expected
of Fh in Fig. 4) is required to swiftly change the contact (reference) contact forces (shown in Fig. 15c) were defined
force to remove different amount of material from different along the same grinding path. The magnitude of the
regions along the grinding path. expected contact forces in both cases was comparable. The

Fig. 15 System responses with

cyclic input. a Horizontal
movement of the workpiece. b
Horizontal movement of the
grinder head when position
controller was adopted. c
Predefined expected (reference)
contact force and actual contact
force obtained from simulation
results. d Horizontal movement
of the grinding head when
position–force controller was
adopted
Int J Adv Manuf Technol (2013) 64:447–458 457

parameters, namely, robotic velocity and contact force, can

be planned by using certain control strategies [16]. The
simulation platform could be applied to predict the material
removal for each of the candidate sequences. Based on the
simulation results, the best candidate sequence can be se-
lected to proceed grinding. An example of the material
removal prediction is shown in Fig. 16, where varying
robotic velocity and contact force sequences were planned
for the grinding path shown in Fig. 2a.

6 Conclusions

A comprehensive simulation platform for robotic belt grind-

ing systems was presented. The platform included the kine-
matics model, the dynamics models of the robot, the grinder,
and the workpiece–wheel interaction, along with the mate-
rial removal model of the grinding process. Simulations of
the grinding process were conducted with full consideration
of the system dynamics. It was revealed that optimal selec-
tion of key parameters of the grinder and proper selection of
robotic control strategies could be of great help in suppress-
ing chatters occurring in the belt grinding process. It was
also demonstrated that the presented platform is a strong
tool for selecting grinding process key parameters for the
conformance grinding process. The authors believe this
study is valuable to industrial development and implemen-
tation of robotic conformance grinding systems, where con-
trolling material removal according to dimensional accuracy
requirements on workpieces is the ultimate goal.

References

1. Brinksmeier E, Heinzel C, Wittmann M (1999) Friction, cooling

and lubrication in grinding. CIRP Ann Manuf Technol 48:581–598
2. Inasaki I, Karpuschewski B, Lee HS (2001) Grinding chatter-
Fig. 16 Material removal prediction. a Planned robot velocity se-
origin and suppression. CIRP Ann Manuf Technol 50:515–534
quence. b Planned contact force sequence. c Predicted material remov-
3. Brinksmeier E, Aurich JC, Govekar E, Heinzel C, Hoffmeister
al along the grinding path (unit: millimeter)
HW, Klocke F, Peters J, Rentsch R, Stephenson DJ, Uhlmann E
(2006) Advances in modeling and simulation of grinding process-
es. CIRP Ann Manuf Technol 55:667–696
4. Chen X, Gong Z, Huang H, Zhou L, Ge SS, Zhu Q (1999) SMART
simulation results in Fig. 15b showed that using robotic
robotic system for 3D profile turbine vane airfoil repair. Int Conf
position controller was prone to generate higher frequency Int Associ Ind Autom 21(4):275–283
chatters of the belt grinder, while no chatter occurred when 5. Huang H, Zhou L, Chen X, Gong Z (2003) SMART robotic system
the robotic position–force controller was adopted as shown for 3D profile turbine vane airfoil repair. Int J Adv Manuf Technol
21:275–283
in Fig. 15d.
6. Sun Y (2004) Development of a unified flexible grinding process.
Furthermore, having the ability to predict material re- Ph.D. Thesis. University of Connecticut
moval rate, the presented simulation platform can also be 7. Li H, Shin YC (2007) A study on chatter boundaries of cylindrical
used as a strong tool for selecting grinding process param- plunge grinding with process condition-dependent dynamics. Int J
Mach Tools Manuf 47:1563–1572
eters for the conformance grinding process. In order to meet
8. Zhang X, Cabaravdic M, Kneupner K, Kuhlenkoetter B (2004)
specific material removal requirements along a grinding Real-time simulation of robot controlled belt grinding processes of
path, candidate sequences of the grinding process key sculptured surfaces. Int J Adv Robot Syst 1:109–114
458 Int J Adv Manuf Technol (2013) 64:447–458

9. Cabaravdic M, Kuhlenkötter B (2005) Bandschleifprozesse opti- 15. Chen X, Brian Rowe W (1996) Analysis and simulation of the
mieren. Metalloberfläche 59:44–47 grinding process. Part II: mechanics of grinding. Int J Mach Tools
10. Ren X, Cabaravdic M, Zhang X, Kuhlenkotter B (2007) A local Manuf 36:883–896
process model for simulation of robotic belt grinding. Int J Mach 16. Wu S (2012) Robotic conformance grinding: modeling, control
Tools Manuf 47:962–970 and optimization. Ph.D. Thesis. University of Connecticut (to be
11. Ren X, Kuhlenkötter B (2008) Real-time simulation and visuali- presented)
zation of robotic belt grinding processes. Int J Adv Manuf Technol 17. Corke PI (1996) A robotics toolbox for MATLAB. Robot Autom
35:1090–1099 Mag IEEE 3:24–32
12. Paul RP (1981) Robot manipulators: mathematics, programming, 18. Kazerounian K, Gupta K (1985) A target tracking manipulation
and control. MIT Press, Cambridge theory for robots. Proceedings of the IASTED International Sym-
13. Craig JJ (1989) Introduction to robotics: mechanics and control. posium Robotics and Automation 263–267
Addison-Wesley, New York 19. Giblin DJ, Zongliang M, Kazerounian K (2007) Target tracking
14. Johnson KL (1987) Contact mechanics. Cambridge University manipulation theories for combined force and position control in
Press, Cambridge open and closed loop manipulators. J Mech Des 129:326