Proceedings of the 2005 IEEE/ASME TD5-04
International Conference on Advanced Intelligent Mechatronics
Monterey, California, USA, 24-28 July, 2005
Machining with Flexible Manipulator: Toward
Improving Robotic Machining Performance
Hui Zhang, Jianjun Wang, George Zhang, Zhongxue Gan Zengxi Pan, Hongliang Cui, Zhenqi Zhu
ABB Inc. Corporate Research Stevens Institute of Technology
Robotics and Automation Group Castle Point on Hudson,
2 Waterside Crossing, Windsor, CT 06095 Hoboken, NJ 07030
noisy, dusty and unhealthy environment (Figure 1).
Abstract—This paper presents the critical issues and Therefore, automation for these operations is highly
methodologies to improve robotic machining performance desirable. However, due to the variations and highly
with flexile industrial robots. Compared with CNC machines,
irregular shape of the automotive casting parts, solutions
the stiffness of industrial robots is significantly lower,
resulting in unacceptable quality and lower productivity. The based on CNC machining center usually presented a high
problem is treated with a novel methodology that consists of cost, difficult-to-change capital investment.
stiffness modeling, real-time deformation compensation for To this end, robotics based flexible automation is
quality and controlled material removal rate for efficiency. considered as an ideal solution for its programmability,
Experimental results show that higher productivity as well as adaptivity, flexibility and relatively low cost, especially for
better surface accuracy can be achieved, indicating a
the fact that industrial robot is already applied to tend
promising and practical use of industrial robots for machining
applications that is not possible at present. foundry machines and transport parts in the process.
Nevertheless, the foundry industry has not seen many
I. INTRODUCTION success stories for such applications and installations.
Currently, more than 80% of the application of industrial
The automotive industry represents the fastest-growing
robots is still limited to the fields of material handling and
market segment of the aluminum industry, due to the
welding. (Figure 2)
increasing usage of aluminum in cars. The drive behind this
is not only to reduce the vehicle weight in order to achieve
New Orders -UNITS
lower fuel consumption and improved vehicle
performance, but also the desire for more sustainable Other, Inspection &
transport and the support from new legislation. Cars Assembly < 10 lbs.
4%
Assembly > 10 lbs.
2%
produced in 1998, for example, contained on average about Material Removal
Spot Welding
27%
85 Kg of aluminum. By 2005, the automotive industry will 4%
be using more than 125 Kg of aluminum per vehicle. It is
estimated that aluminum for automotive industry alone will Material Handling
>10 lbs
be a 50B$/year market. 29% Arc Welding
18%
Material Handling
< 10 lbs. Dispensing/Coating
8% 8%
Figure 2: 2003 Robot Applications in North America. A total of 12,367
robots valued at $876.5 million were ordered. When sales to companies
outside North America are added in, the total for North American robotics
suppliers is 12,881 robots valued at $913 million. NOTE: These numbers
include results from North America and Outside North America. Source:
Robotic Industries Association
The major hurdle preventing the adoption of robots for
Figure 1: Manual cleaning operations in a foundry plant material removal processes is the fact that the stiffness of
today’s industrial robot is much lower than that of a
Most of the automotive aluminum parts start from a standard CNC machine. The stiffness for a typical
casting in a foundry plant. The downstream processes articulated robot is usually less than 1 N/Pm, while a
usually include cleaning and pre-machining of the gating standard CNC machine center very often has stiffness
system and riser, etc., machining for high tolerance greater than 50 N/Pm.
surfaces, painting and assembly. Today, most of the Most of the existing literature on machining process,
cleaning operations are done manually in an extremely such as process force modeling [1, 2], accuracy
0-7803-9046-6/05/$20.00 ©2005 IEEE. 1127
Authorized licensed use limited to: UNIVERSITE DE TUNIS. Downloaded on August 28,2025 at 14:06:09 UTC from IEEE Xplore. Restrictions apply.
improvement [3] and vibration suppression [4] are based on reduce cycle time. Every part, including the first, is
the CNC machine. Research in the field of robotic optimized automatically, eliminating the need for manual
machining is still focused on accurate off-line part program optimization.
programming and calibration [5, 6]. Akbari etc al [7] To summarize, in this paper, our focus is to address two
describe a tool angle adjustment method in a grinding issues: 1) to improve the robotic machining quality with the
application with a small robot. In that case the process low stiffness, low accuracy robot; 2) to improve the robotic
force is very small. Matsuoka etc al [8] study the characters machining efficiency by providing real time optimization
of an articulated robot in a milling process avoiding large to maximize material removal rate. If industrial robots
process force by using an end mill with small diameter and could be made to provide end-effector position accuracies
high spindle speed. Without the capability of realtime under contact situations equal to the end-effector position
force control, the method to eliminate the force effect on repeatabilities they already provide, then robotic machining
the robotic machining process has not been fully addressed would result in significant cost savings for a lot of potential
in the research community or in industry. applications. Moreover, if such applications can be proven
Machining processes, such as grinding, deburring, to be economically viable and practically reliable, one
polishing, and milling are essential force tasks whose would expect to witness many success stories in the years
nature requires the end effector to establish physical to come.
contact with the environment and exert a process-specific This paper is organized in six sections. Following this
force. The inherent lower stiffness of the robot has introduction section, section II provides the in-depth
presented many challenges to execute material removal analysis for the modeling and measurement techniques of
applications successfully. The first one is the structure the robot stiffness matrix, which would determine the
deformation and loss of accuracy due to the required structure deformation when a machining process is taking
machining force. The predominant cutting action in place. Building on that result, Section III addresses the
machining involves shear deformation of the work material technique for real-time deformation compensation so that
to form a chip. The contact between the cutting tool and the the quality and accuracy of the robotic machining operation
workpiece generates significant forces. As a result, a can be improved. Section IV presents the real-time
perfect robot program without considering contact and optimization methodology for a controlled material
deformation will immediately become flawed as the robot removal rate so that the efficiency of such operations can
starts to execute the machining task. Unlike multi-axis be improved. Experimental results are presented in
CNC machine centers, such deformation is coupled and Sections V followed by a summary in Section VI.
varies even subjected to the same force in different
workspace locations. Such coupling results in deformation II. ROBOT STIFFNESS MODEL
not only in the direction of reaction force and can generate As was stated before, one of the focuses for this paper is
some counter-intuitive results. to improve the robotic machining accuracy by reducing
Secondly, the lower stiffness also presents a unique machining force induced deformation. While thermal
disadvantage for machining of casting parts with complex induced error is the largest error component for CNC
geometry, which means non-uniform cutting depth and machining, motion error due to machining force contributes
width. As a result, the machining force will vary most of the total machining errors in robots. For example, a
dramatically, which induces uneven robot deformation. 500N cutting force during a milling process will cause a 1
What this means in one example is that the flatness of the mm position error for a robot instead of a less than 0.01mm
machined plane is so inferior that it renders the robotic error for a CNC machine. In order to achieve higher
process unable to meet the minimum requirement. dimensional accuracy, the deformation due to the
In general practice, machine tools maximize the material interactive force must be compensated.
removal rate (MRR) during roughing cycles by applying all Since force measurement and subsequent compensation
of the available spindle power to the machining process. is carried out in 3-D Cartesian space, a stiffness model,
When machines use carbide tools for roughing operations, which relates the force applied at the robot tool tip to the
the available spindle power is usually the limiting factor on deformation of the tool tip in Cartesian space, is crucial to
MRR. In conventional robot programming and process realize deformation compensation. The model should be
planning practice, the cutting feed rate is constant even accurate enough for the prediction of robot structure
with significant variation of cutting force from part to part, deformation under arbitrary load conditions. At the same
which dictates a conservative cutting feed rate without time, it needs to be simple enough for real time
violating the operational limits. Therefore, it is desirable to implementation. Detailed modeling of all the mechanical
maximize MRR and minimize cycle time by optimizing the components and connections will render a model too
cutting feed rate based on a programmed spindle load. By complicated for real-time control, and difficult for accurate
optimizing the feed rate in real time, one could compensate parameter identification.
for conservative assumptions and process variations to help Industrial robotic systems are designed to achieve high
1128
Authorized licensed use limited to: UNIVERSITE DE TUNIS. Downloaded on August 28,2025 at 14:06:09 UTC from IEEE Xplore. Restrictions apply.
positioning accuracy and high strength. Elastic properties gives us:
of the arms are insignificant. Therefore, the dominant F T 'X W T 'Q (4)
contribution factor for a large deflection of the manipulator From (1), (3), (4), we have:
tip position is the joint compliance, e.g., due to gear
transmission elasticity. Modeling of robot stiffness could Kx J (Q) T K q J (Q) 1 (5)
be reduced to six rotational stiffness coefficients in the For an articulated robot, K x is not a diagonal matrix and
joint space. From the control point of view, this model is it is configuration dependent. This means that: first, the
also easy to implement, since all industrial robot controllers force and deformation in Cartesian space is coupled, in
are decoupled to SISO joint control at the servo level. As a other words, the force applied in one direction will cause
result, the joint deformation could be directly compensated the deformation in all possible directions; second, the
on the joint angle references passed to the servo controller. stiffness is also a function of robot kinematics J (Q ) , it
Note here that the axis of the force sensor coincides with
changes significantly in the entire workspace. However,
the axis of joint 6, the stiffness of the force sensor and its
even though at different locations, the stiffness matrix will
connection flange could be modeled into joint 6. Figure 3
take different values (see Table 1 for one example), these
shows the structure of a 6-DOF ABB IRB 6400 robot with
changes can be sufficiently modeled by Eq. (5), with the
black arrows representing the location of compliant joints.
assumption that K q (i, i ) , representing the stiffness of joint
i, is a constant value. Thus, if K q can be measured
accurately, the deformation of robot TCP under external
force at any location in the workspace could be estimated
as,
1
'X J (Q) K q J (Q) T F (6)
2.86E+02 -8.78E+02 8.12E+02 1.49E+06 3.01E+05 3.97E+05
-8.78E+02 2.39E+02 5.48E+03 -1.85E+05 -5.46E+05 -4.09E+05
8.12E+02 5.48E+03 4.91E+02 -4.56E+06 4.31E+05 8.62E+05
1.49E+06 -1.85E+05 -4.56E+06 7.53E+07 5.85E+08 -7.62E+08
3.01E+05 -5.46E+05 4.31E+05 5.85E+08 1.68E+08 3.20E+08
3.97E+05 -4.09E+05 8.62E+05 -7.62E+08 3.20E+08 1.30E+08
Table 1 One example of the Cartesian Stiffness Matrix (Units are N/mm,
N/rad, N·mm/mm, and N·mm/rad respectively.)
Figure 3: Structure of 6-DOF ABB IRB 6400 manipulator Experimental determination of joint stiffness parameters
is critical in fulfilling real-time position compensation. In
Next, we will derive the stiffness model in Cartesian this model, the joint stiffness is an overall effect
space based on joint compliance parameters. contributed by motor, joint link, and gear reduction units. It
In joint space, the model could be represented as: is not realistic to identify the stiffness parameter of each
W K q 'Q (1) joint directly by dissembling the robot; the practical
method is to measure it in Cartesian space.
Where: W is the torque load on the each joint; K q is a
6×6 diagonal matrix; 'Q is the 6×1 deformation vector.
While in Cartesian space:
F K x 'X (2)
Where F is the 6 D.O.F. force vector, 'X is the 6
D.O.F. deformation of the robot in Cartesian space, and
K x is a 6×6 stiffness matrix.
From the definition of the Jacobian matrix, we have:
'X J (Q) 'Q (3)
Where J (Q) is the Jacobian matrix of the robot.
At the steady state, after compensating the tool gravity
force, the robot joint torques will exactly balance external
forces applied on the tool tip. The principle of virtual work
Figure 4: Experiment Setup of robot stiffness measurement
1129
Authorized licensed use limited to: UNIVERSITE DE TUNIS. Downloaded on August 28,2025 at 14:06:09 UTC from IEEE Xplore. Restrictions apply.
controller of industrial robot.
To be able to measure small deformations in 3-D space,
the end-effector is equipped with a sphere-tip tool shown in Stiffness Frame Filter
Fms
Figure 4. The tool tip is set to a fixed point in the 'q r Model Transform
workspace, and the manipulator joint values are Gravity
recorded. A given load in the range of 100N~400N is qrold qrnew Model
Robot
applied to the tool, causing the sphere-tip to move away Controller
from the original point. The original and deformed
positions are measured with ROMER, a portable CMM 3-
Figure 6: Principle of real-time deformation compensation
D digitizer, and the 3-DOF translational deformations are
calculated. From, Eq. (6), K q could be solved by least The block diagram of real time deformation
square method. The same experiment was repeated at compensation is shown in Figure 6. After filtering the force
several different locations in the robot workspace. sensor noise and compensating the gravity of the spindle
and the cutter, the force signal was translated into the robot
The same procedure is taken at different locations in the tool frame. Based on the stiffness model identified before,
robot workspace, the deviation of the results is small, the deformation due to machining force is calculated in real
which means a set of constant model parameters could time and the joint reference for the robot controller is
model the robot deformation with small error. (Figure 5) updated.
IV. CONTROLLED MATERIAL REMOVAL RATE
In pre-machining processes, maximum material removal
rates are even more important than precision and surface
finish for process efficiency. MRR is a measurement of
how fast material is removed from a workpiece; it can be
calculated by multiplying the cross-sectional area (width of
cut times depth of cut) by the linear feed speed of the tool:
MRR w d f (7)
Where w is width of cut (mm), d is depth of cut (mm),
f is feed speed (mm/s).
Conventionally, feed speed is kept constant in spite of
the variation of depth of cut and width of cut during
foundry part pre-machining process. Since most foundry
Figure 5: Error of stiffness modeling
parts have irregular shapes and uneven depth of cut, this
will introduce a dramatic change of MRR, which would
III. ROBOT DEFORMATION COMPENSATION
result in a very conservative selection of machining
The major position error sources in robotic machining parameters to avoid tool breakage and spindle stall. The
process can be classified into two categories, (1) concept of MRR control is to dynamically adjust the feed
Machining force induced error, and (2) motion error speed to keep MRR constant during the whole machining
(kinematic and dynamic errors, etc.). The motion error, process. As a result, a much faster feed speed, instead of a
typically in the range of 0.1 mm, is inherent from the robot conservative feed speed based on maximal depth of cut and
position controller and would appear even in non-contact width of cut position, could be adopted. (Figure 7)
cases. While the machining force in the milling process is Since the value of MRR is difficult to measure, the MRR
typically over several hundreds of Newton, the force- is controlled by regulating the cutting force, which is
induced error, which could easily go up to 1 mm, is the readily available in real-time from a 6-DOF strain gage
dominant factor of surface error. Our objective is to force sensor fixed on the robot wrist. Placing the analysis
estimate and compensate the deformation in real time to of the material removal process on a quantitative basis, the
improve the overall machining accuracy. characterization of cutting force is important for research
The existing research of robot deformation compensation and development into the modeling, optimization
is focused on gravity compensation, deflection monitoring and control of metal cutting.
compensation of large flexible manipulators, etc. Not The challenges for designing a robust controller for
much attention has been paid to the compensation of MRR is the fact that cutting process model varies to a large
process force induced robot deformation due to the lack of degree depending on the cutting conditions. Efforts for
understanding and model of robot structure stiffness, the designing an adaptive controller will be presented in a
lack of real time force information and limited access to the separate paper.
1130
Authorized licensed use limited to: UNIVERSITE DE TUNIS. Downloaded on August 28,2025 at 14:06:09 UTC from IEEE Xplore. Restrictions apply.
the cutting force Fc , while force process gain T changes.
Aggressive Failure and V. EXPERIMENTAL RESULTS
Dangerous
Condition In the previous sections, the robot deformation subject to
an arbitrary process force loading is modeled and the
Optimal Power Limit
model parameter is experimentally measured. With this
model, the online deformation scheme is implemented on
Conservative Ft =400N the robot controller. Secondly, the concept for controlled
Ft =350N material removal rate is presented and implemented. In this
Ft =300N
Ft =250N section, the experimental results are presented to validate
Ft =200N
Ft =150N
the aforementioned schemes.
Figure 9 shows the setup of a milling test. A spindle is
fixed on the robot arm and the workpiece is fixed on a steel
Variation in depth of cut
table. For illustration, a 6063 aluminum block is used for
Figure 7: Controlled material removal rate
testing purpose.
As the feed speed f is adjusted to regulate the
machining force, MRR could be controlled under a specific
spindle power limit avoiding tool damage and spindle stall.
Also, controlled MRR means predictable tool life, which is
very important in manufacturing automation. Figure 8
shows the block scheme of machining force control with
controlled material removal rate (CMRR).
Figure 9: Experimental setup for robotic milling
Tests on an aluminum block with the depth of cut
changed from 2 mm to 3 mm shows, when force control is
activated, the cutting force is regulated in spite of the
variance of depth of cut. (Figure 10) The milling test of
aluminum with variation of width of cut shows similar
results.
Figure 8: Force control for robotic machining As a result, the feed speed could always be setup as fast
as the limit of spindle power. In a foundry parts milling or
The structure of cutting force in a milling operation is deburring process, the robot won’t have to move at a very
represented as linear first-order model: conservative speed to avoid tool breakage or spindle stall.
1 (8) The cycle time decreased by CMMR is typically around
Fc K w d f
W ms 1 30% to 50% for different workpieces.
In the deformation compensation test of milling an
where W m is the machining process time constant. Since
aluminum block, a laser displacement sensor is used to
one spindle revolution is required to develop a full chip measure the finished surface. The surface error without
load, Wm is 63% of the time required for a spindle deformation compensation demonstrates counter-intuitive
revolution [9]. Since the force control is implicitly results; an extra 0.5mm was removed in the middle of the
implemented, the control loop bandwidth is limited by a milling path. Conventional wisdom says that a flexible
position servo control, which is around 10 Hz for an machine would also cut less material due to deformation,
industrial robot. The force process gain may be seen as since the normal force during cutting will always push the
cutter away from the surface and cause negative surface
T K w d , which is sensitive to the process inputs.
error. However, in the articulated robot structure, the
With the proper selection of reference feed speed f r and
deformation is also determined by the structure Jacobian, in
reference force Fr , a PI controller is adopted to regulate a lot of cases, a less stiff robot could end up cutting more
1131
Authorized licensed use limited to: UNIVERSITE DE TUNIS. Downloaded on August 28,2025 at 14:06:09 UTC from IEEE Xplore. Restrictions apply.
material than programmed. The coupling of the robot surface accuracy. The results outline a promising and
stiffness model explains this phenomenon, the force in feed practical use of industrial robots for machining applications
direction and cutting direction will result in positive surface that is not possible at present.
error in that robot configuration. Since the feed force and
the cutting force are the major components in this setup,
the overall effect will cut the surface 0.5 mm more than the
commanded depth. In our definition, negative surface error
means less material was removed than the commanded
position. The result after deformation compensation shows
a less than 0.1 mm surface error, which is in the range of
robot path accuracy. (Figure 11) Further test conducted on
the foundry cylinder head workpiece shows that the surface
accuracy improved from 0.9mm to 0.3mm, which is below
the 0.5mm target accuracy for pre-machining application.
Figure 10: Force control result of variant depth of cut
VI. SUMMARY
In this paper, critical issues and methodologies to
improve robotic machining performance with flexile
industrial robots are presented, where previous force
control research is expanded into machining application.
Practical machining experiments were conducted in the lab
to validate the concept and design methodology; both
deformation compensation and controlled MRR algorithm Figure 11: Deformation compensation results
are demonstrated in experiments. The experiment results
showed great reduction of cycle time, as well as better
[6] M. Sallinen, T. Heikkilä, 2000, "Flexible Workobject Localisation
for CAD -Based Robotics", Proceedings of SPIE Intelligent Robots
REFERENCES and Computer Vision XIX: Algorithms, Techniques, and Active
[1] Sung I. Kim, Robert G. Landers, A. Galip Ulsoy, 2003, “Robust Vision. Boston, USA, 7 – 8 Nov. 2000. USA. Vol. 4197 (2000), pp.
Machining Force Control with Process Compensation,” Journal of 130 - 139
Manufacturing science and engineering, Vol 125, pp. 423-430 [7] Ali A. Akbari, Shizuichi Higuchi, 2000,”Autonomous Tool
[2] Jeffrey L. Stein, Kunsoo Huh, 2002, “Monitoring Cutting Forces In Adjustment in Robotic Grinding,” The int. conf. on Precision
Turning: A Model-Based Approach,” Journal of Manufacturing Engineering(ICoPE) ,121-126
science and engineering, Vol 124, pp. 26-31 [8] Shin-ichi Matsuoka, Kazunori Shimizu, Nobuyuki Yamazaki, Yosh
[3] Seung-Han Yang, 1996, “Real-time compensation for geometric, inari Oki, 1999, “High-speed end milling of an articulated robot and
thermal, and cutting force induced errors in machine tools,” Ph.D. its characteristics,” Journal of Materials Processing Technology
dissertation, The University of Michigan Volume: 95, Issue: 1-3 pp. 83-89
[4] E. Budak, Y. Altintas, 1998, ''Analytical Prediction of Chatter [9] L.K. Daneshmend, H.A. Pak, 1986, “Model Reference Adaptive
Stability Conditions for Multi-Degree of Systems in Milling. Part I: Control of Feed Force in Turning,” ASME Journal of Dynamic
Modeling, Part II: Applications,” Transactions of ASME, Journal of Systems, Measurement, and Control, Vol. 108, No. 3, pp. 215-222.
Dynamic Systems, Measurement and Control, vol.120, pp.22-36 [10] D. E. Whitney, “Force feedback control of manipulator fine
[5] Y.H. Chen, Y.N. Hu, 1999. "Implementation of a Robot System for motions,” ASME J. Dynamic Syst. Measure. Control, vol. 99, no. 2,
Sculptured Surface Cutting. Part 1. Rough Machining". Int. Journal pp. 91–97, 1977
of Advanced Manufacturing Technology, Vol 15. Pp. 624-629
1132
Authorized licensed use limited to: UNIVERSITE DE TUNIS. Downloaded on August 28,2025 at 14:06:09 UTC from IEEE Xplore. Restrictions apply.
