ARTICLE IN PRESS

Robotics and Computer-Integrated Manufacturing 26 (2010) 2438

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing

journal homepage: www.elsevier.com/locate/rcim

Cost-efcient drilling using industrial robots with high-bandwidth

force feedback
Tomas Olsson a,d,f, Mathias Haage b, Henrik Kihlman c,e, Rolf Johansson a,, Klas Nilsson b,
Anders Robertsson a, Mats Bjorkman c, Robert Isaksson c, Gilbert Ossbahr c, Torgny Brogardh d,f
a
Department of Automatic Control, Lund University, Box 118, SE-22100 Lund, Sweden
b
Department of Computer Science, Lund University, SE-22363 Lund, Sweden
c
Department of Management and Engineering, Linkoping University, SE-58183 Linkoping, Sweden
d
ABB Robotics, Department RC, SE-72168 Vasteras, Sweden
e
DELFOi, Soerogatan 3D, SE-41251 Goteborg, Sweden
f
ABB Corporate Research, SE-72178 Vasteras, Sweden

a r t i c l e in f o a b s t r a c t

Article history: Here we present a method for high-precision drilling using an industrial robot with high-bandwidth
Received 16 November 2007 force feedback, which is used for building up pressure to clamp-up an end-effector to the work-piece
Received in revised form surface prior to drilling. The focus is to eliminate the sliding movement (skating) of the end-effector
30 November 2008
during the clamp-up of the end-effector to the work-piece surface, an undesired effect that is due to the
Accepted 30 January 2009
comparatively low mechanical stiffness of typical serial industrial robots. This compliance also makes
the robot deect due to the cutting forces, resulting in poor hole position accuracy and to some extent in
Keywords: poor hole quality. Recently, functionality for high-bandwidth force control has found its way into
High-precision drilling industrial robot control systems. This could potentially open up the possibility for robotic drilling
Force control
systems with improved performance, using only standard systems without excessive extra hardware
Feedback
and calibration techniques. Instead of automation with expensive xtures and precise machinery, our
Motion control
Industrial robotics approach was to make use of standard low-cost robot equipment in combination with sensor feedback.
The resulting sliding suppression control results in greatly improved hole positioning and quality. The
conceptual idea behind the force control is useful also in many other robotic applications requiring
external sensor feedback control.
& 2009 Elsevier Ltd. All rights reserved.

1. Introduction condition that no unmodeled external forces act on the robot,

The traditional application areas for industrial robots involve
highly repetitive operations such as spot welding. Hence, robotic
development has been focused on high precision (repeatability) in
repetitive operations. For example, a standard ABB IRB4400 robot
for 60 kg payload has a repetitive accuracy of 0:05 mm. If,
however, the same robot is given a new coordinate that it has
never visited before, the accuracy is in general around 3 mm,
which is 60 times the size of the repetitive accuracy. Today it is
possible to buy the same robot with an option pack that includes
calibration for high accuracy. This will improve accuracy to
become within 0:5 mm, which is still 10 times the repetitive
accuracy. In addition, this accuracy is only guaranteed under the
 Corresponding author. Tel.: +46 462228791; fax: +46 46138118.
E-mail addresses: Tomas.x.Olsson@se.abb.com (T. Olsson),
Mathias.Haage@cs.lth.se (M. Haage), Henrik.Kihlman@delfoi.com (H. Kihlman),
Rolf.Johansson@control.lth.se (R. Johansson), Klas.Nilsson@cs.lth.se (K. Nilsson),
Anders.Robertsson@control.lth.se (A. Robertsson), Mats.Bjorkman@liu.se
(M. Bjorkman), Gilbert.Ossbahr@liu.se (G. Ossbahr), Torgny.Brogardh@se.abb.com
(T. Brogardh).
i.e., only during motion in free space. For contact tasks such
as polishing, drilling, and riveting, the effects of the limited
mechanical stiffness of the robot must be taken into account in
order to maintain the positioning accuracy, which is not feasible
unless a detailed model of the particular robot specimen is
available.
Systems for automatic drilling have a long history both in
industry and the research community. In particular, the use of
industrial robots for drilling is interesting due to their exible
programming and the comparatively low cost of industrial robot
systems. However, robot drilling is a very challenging task due to
the comparatively low mechanical stiffness of the typical serial
industrial robots in use today. In general, clamp-up is necessary in
robotic drilling to avoid vibration during the drilling process as
the drill tool generates vertical, horizontal and axial forces during
the cutting process. The compliance makes the robot deect,
sometimes up to several millimeters, due to the externally applied
forces during clamp-up and drilling. Due to the bending
of the robot links and the elasticity in the gears, the local
deection at the contact point does not necessarily occur in the
(axial) direction of the applied force, but may have tangential

0736-5845/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.rcim.2009.01.002
 ARTICLE IN PRESS

T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438 25

components which are on the same order of magnitude as the 1.1. Problem formulation
axial deection. This tangential deformation results in poor hole
quality and inaccurate positioning. In contrast, aerospace toler- The purpose of this work is a fully developed industrial
ances require drilled holes to be accurate within 0:2 mm [1]. prototype of robotic drilling, based on the use of high-perfor-
As for a state-of-the-art comparison on robotic drilling and mance force/torque control and light-weight industrial robots.
fastening, we refer to the recently reported robot capability test The idea presented in this paper is based on applying a
targeting applications in the aerospace industry with the test dynamically controlled pressure against the work piece with a
robots KUKA KR240, KUKA KR60, ABB IRB7600 and Staubli RX170 tripod attached to the drilling tool, while a self-feeding mechan-
[1]. In this study from Airbus UK, limitations related to static and ism is used to feed the drill. This setup is as shown in Fig. 1. When
dynamic deection, repeatability, absolute accuracy, temperature used together with a metrology system for absolute accuracy in
error, and hysteresis were surveyed, the conclusion being that an the initial positioning, the system should be able to satisfy the
absolute accuracy 0:2 mm was not achievable. Using state- accuracy requirements of 0:2 mm, even in the presence of
of-the-art anti-skating approaches, Atkinson and co-authors from external load during clamp-up or drilling. The method of dynamic
Boeing-Hawker de Havilland concluded that absolute accuracy sensor-controlled drilling represents a different approach com-
remained on the edge of acceptability for aircraft assembly [2]. pared to current static (and expensive) systems. The purpose of
A drilling process involves moving a drilling end-effector to the the force control is threefold: (i) to control the normality to the
correct position of the hole. Prior to drilling, a pressure foot is surface; (ii) to avoid the drilling end-effector sliding on the surface
used to press the parts together in order to avoid burrs entering in (skating) during the drilling and clamp-up phases; and (iii) to
between the plates. In addition, the pressure foot assures that press the parts together so that burrs do not enter in between the
the drilling machine is kept stable throughout the drilling cycle. plates. The control is accomplished by an open robot controller
A self-feeding mechanism is normally used to feed the drill interface with a sampling rate of 250 Hzsee [7,17]. Further, the
through the stack of materials. Automated drilling in the aero- system allows for reuse of much of the existing safety and
space industry today uses large robots for two major purposes: to programming functionality of the robot system, with extensions
handle the large assemblies, and to accurately counter balance the made to allow exible programming of sensor-based tasks. The
drilling forces involved in the drilling process. There are many off-line programming (OLP) environment DELMIA V5 Robotics
different ways to overcome forces in drilling and fastening using was selected for simulation and planning of the drilling process,
industrial robots. One approach is to divide the process in two thereby simplifying robot programming. The robot hole-to-hole
steps, where in the rst step the robot stiffness is mapped programming includes meta-data for the force control in pre-
by applying forces to the robot TCP and measuring its deection, aligning the pressure foot. The force data are included in the robot
while in the second step the robot is adjusted back to the nominal program with our extension to the ABB Rapid program language
position under load. These compensation values are then applied called ExtRapid. The goal is to avoid additional force feedback
as a lter to program the robot during process execution [3]. programming after the OLP in DELMIA.
Mapping the robot stiffness in this way can, however, be extremely
time consuming. Other methods to solve the skating problem have
been tested by using metrology systems to supervise the robot, see
for example [4,5,1,6]. In such methods, a metrology system is 2. System topology
connected to the robot controller via an external feedback loop to
update the nominal position of the robot with measurements in real
The robot system includes a robot with its controller, two
time. However, using metrology is not a straightforward solution, as
computers and a force/torque sensor. The solution presented in
the robot will deect as the pressure foot is engaged.
Common to the traditional approaches is the lack of high-
performance sensor feedback to the robot, whereby the robot
cannot be updated fast enough to cope with the dynamic process
that drilling involve. Robot control systems are traditionally
closed, a circumstance which has hampered system integration
of manipulators, sensors and other equipment, and such system
integration has often been made at an unsuitably high hierarch-
ical level. As a more cost-effective solution, high-bandwidth
feedback techniques can be used to control the properties of the
drilling process. Research and development on force-controlled
drilling has not received as much attention as many other
applications of industrial force control, such as assembly, deburring,
milling or polishing [7,8]. The reason is probably the difculties
involved in robotic drilling, and the lack of available industrial robot
systems with capacity for high-bandwidth force control [911].
Some results on force control for special drilling machines have
been reported in [12]. Experimental systems for force-controlled
robot drilling have been presented in [13], where a force controller
with inner position control was used for the drilling thrust force
control, and in [14], where an application to bone drilling in
orthopedic surgery was presented.
While there exist commercially available products for force
controle.g., from ABB Robotics [15]none of the available
packages include the particular features and/or level of exibility
required for the drilling application, i.e., the ability to redesign the
inner-loop servo control for improved disturbance rejection. Fig. 1. Overview of the end-effector prototype for the drilling experiments.
 ARTICLE IN PRESS
26 T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438
this paper has the sensor installed in the end-effector. Fig. 2 shows
an overview of the system. In detail, the system includes:
External computer: Master PC for ExtRapid program execution.
Robot: ABB IRB 4400 industrial robot.
Robot controller: Modied ABB S4CPlus control system.
Power PC card: G4 processor with memory, PMC interface.
PMC-PCI card: Interface between the Power PC card and the PCI
backplane.
Sensor: JR3/160M50 force sensor and corresponding computer
interface.
Network environment: Congured from the master PC.
Master PC software environment: Making it able to execute and
supervise ExtRapid programs.
The architecture of the ABB S4CPlus control system and its
extensions are shown in Fig. 3. The force-control robot system
consists of the robot and its controller, extended with an extra
processor card (G4). The setup uses two additional PCs, MasterPC
for executing the robot language extension (ExtRapid) and
SensorPC for handling the force/torque sensor, responsibilities
that are to be handled by the controller and the G4, but for
G4 Robot and Tool
Master PC
Complete System
Fig. 2. Overview of the robot control system. The sensor PC is an optional
component not used in the drilling application, as support for JR3 force/torque
sensors has been integrated into the robot controller system, and signals can be
read directly into the Power PC card over the PCI bus. However, the ability to
include a Sensor PC gives extra exibility in interfacing other types of sensors
typically used in laboratory environments, such as vision sensors for CPU-intensive
real-time applications.
Fig. 3. Extension of industrial robot controller with sensor interface and support for external computations and synchronization, using a Motorola PPC-G4 PrPMC-800
processor board mounted on a Alpha-Data PMC-to-PCI carrier board with a local PCI bus. Schematics overview from [7].
development and testing purposes have been kept separate. The
MasterPC has a double role, acting as cell controller.
Task descriptions, as expressed in the Rapid language, are
passed through the trajectory generation, and turned into
references for the low-level servo controllers. The extensions
were implemented by modifying references on the joint level with
a 4 ms sampling interval, which gives a suitable trade-off between
engineering effort, preserved safety and performance. The com-
munication between the master PC and the S4CPlus controller is
over the Ethernet using TCP/IP and UDP/IP. The master PC
evaluates the ExtRapid instructions and synchronizes the robot
program execution with the low-level force control along the path
interpolation. When executing force feedback with a sampling
rate of 250 Hz, the master PC uses the received path coordinate as
basis for the 4-ms advancement along the path, maintaining the
overall path coordinate and computing force control set points
parameters accordingly. The actual low-level external control
functionality is present in the force control block in Fig. 3.
The controllers can be dened in Matlab/Simulink blocks and
translated for embedding by using MathWorks Real-Time Work-
shop, to be compiled and linked with external libraries. In
the resulting system, the control engineer can graphically edit
the force control block diagram, and then build and deploy it into
the robot controller. In order to accomplish interrupt driven hard
real-time execution with shared memory communication, the
force controller is run as a Linux kernel module on an added
Motorola PPC-G4 processor board on a local PCI bus (Fig. 4). In
order to maintain safety, the internal safety functions are
activated also during sensor-based motions, and the modied
references for the arm-level control are subject to the same safety
logic as ordinary trajectories. A detailed system description is
provided in [7].
The extended system offers an open experimental platform
for robotics research explored on many hierarchical levels (from
control algorithms with high bandwidth, to robot programm-
ing and task modeling with on-line sensor information). The
preserved high-level support and the integration with the super-
vision and safety system of the standard industrial robot system
constitute a major difference to most Open Robot systems
which have been reported for academic research. To our knowl-
edge, there are no other systems that provide the same high
sampling rate and low inputoutput latency (within a factor of
ten) together with all the user-programming features and super-
visory functions [16,17]. Currently, the system is used in Lund,
 ARTICLE IN PRESS
T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438
Fig. 4. Robot control software architecture based on the S4CPlus robot control
with sample rate 250 Hz (h 4 ms).
Linkoping, and Valencia. Moreover, the platform was used for pre-
product force control application development at ABB.
2.1. Robot system programming
Robot force control programming is performed on three levels:
(i) specication of a Simulink controller acting within the ABB
industrial robot controller; (ii) steering of the controller through
language extensions in the ABB robot programming language; (iii)
task-level force annotations in the Delmia simulation software.
The robot controller extension allows arbitrary Simulink models
to affect the robot trajectory at the 4 ms joint level. Developed
models are compiled using Real-time Workshop and transferred
to the G4 le system (NFS, FTP, etc.) prior to execution. Execution
time is not dependent upon ABB hardware but relies on the G4
processor, allowing for quite CPU-intensive models. Compiled
models can easily be switched between task executions. A
Simulink library offers blocks for reading and writing ABB
trajectory data. External sensor data can be integrated as extra
blocks, with data provided, for example, through the G4 Ethernet
port. A low-level Java GUI (OpCom) allows direct access to the G4
for switching models, logging model parameter data and changing
model parameters for tuning and validation in a running system.
At a higher level, model parameters are exported dynamically
to the robot programming language extension (ExtRapid) where
they can be accessed and modied as part of a regular robot
program. Usage of a model is encapsulated in a controller block.
Outside the controller block, regular ABB RAPID code is executed
as usual. Inside the controller block, RAPID code is executed under
the inuence of the model. Since the model is not allowed to
deviate too much from the planned trajectory for supervisory and
safety reasons, RAPID code is needed for dening a default
trajectory within the block. The interface between regular code
and model-inuenced code is currently solved by requiring the
robot system to be in a known state (non-moving, not-in-process)
when turning on and off the model.
The ExtRapid extension is implemented to be in conformity
with standard RAPID on ABB S4 and IRC5 systems. How to express
the extension in the original language is still an open question.
During the course of the experiment two alternatives were tested.
In the rst alternative, controller blocks were specied using
a XML-like annotation hidden in Rapid comments. This made it
easy to use existing tools for simulation and OLP for specication
of default trajectories. The blocks were implemented using
shallow parsing techniques (regular expressions)cf. Fig. 3. In
the second alternative, a full parser using aspect-oriented
27
compiler-compiler tools were used to express both the original
language and the extension as imperative and declarative aspects,
thus enabling force annotation to be added as language constructs
to the original language [18,19]. Potentially, this approach allows
sensor-specic interfaces to be modularized as language aspects
and enter the language with full syntactic and semantic checks.
Both approaches allow force constructs to be added to the original
language, either manually by the operator, and/or included as part
of trajectory meta-data in a tool with tags generated during code
generatione.g., post-processing in DELMIA. As for programming
environments, we refer to examples to be found in Appendices A
and B.
This project used the OLP method, with DELMIA V5 Robotics as
the OLP system. The force-feedback programming method in this
project was aimed at making the programming environment as
powerful and user-friendly as possible. In addition to the default
installation of DELMIA a small interface was implemented to
incorporate ExtRapid programming as part of the robotic drilling
simulation process. In DELMIA there is a functionality to include
comments to a robot program, as the comment types PreCom-
ment and PostComment. The PreComment includes a comment
prior to a program line in the post-processed robot program.
PostComment does the same, but includes the comment text after
the program line. For each position in the simulation model to be
drilled the user clicks Insert ExtRapid and the ExtRapid menu is
shown, see Fig. 5. Using PreComments and PostComments enables
the user to view ExtRapid data directly in the PPR-tree. Using
PreComment and PostComment was the key to avoid customizing
the post-processor in DELMIA, see Fig. 6 and Appendix. When the
program is executed in the robot, the master controller interprets
the ExtRapid data. One advantage having the master controller
reading directly from the program in the robot controller is the
ability to include ExtRapid data in the teach-in programming
method. The method to include the ExtRapid data in DELMIA
made it possible to avoid the additional programming of force
data on the workshop oor.
3. Drilling system design
The end-effector prototype, an overview of which is shown in
Fig. 1, was designed with the purpose of generating force and
torque sensor data to the robot controller. The drilling tool was
equipped with a tripod with three antennas positioned symme-
trically around the drilling tip. When each antenna is pressed to
the surface a force will build up and propagate to the force sensor.
Three antennas were used to recognize asymmetrical forces
around the drill tip in order to compensate for lack of normality.
The Coromant Captot interface was used for attachment to the
robot chuck. The force sensor was a JR3 160/50M, and was
installed between the drilling unit and the pressure foot. The end-
effector was congured as a mix between a pointing conguration
and a hanging conguration [4], see Fig. 7. The sensor data from
the force/torque sensor were used in the following way:
Torque: different amount of force on each antenna will cause
errors in normality.
Z-force: the total amount of pre-load force exerted on the
surface.
XY-force: the force indicating the skating effect.
All the six degrees of freedom (6-DOF) of the force/torque
sensor are used to compensate for the different phenomena in the
process. The torque will increase if the end-effector is rotated
around the X- or Y-axis (dened in the work-piece plane), which
corresponds to the end-effector force not being normal to the
surface. Thus, the controller will make use of the torque
measurements to even out the force on the three antennas. The
 ARTICLE IN PRESS

28 T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438

Fig. 5. The Delmia interface with ExtRapid programming functionality.

Fig. 6. ExtRapid task-level programming in Delmia software.

 ARTICLE IN PRESS
T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438
Fig. 7. Conguration of the end-effector prototype for the drilling experiments.
controller will control the Z-force to ramp up the clamp-up force
to the reference value and use the X- and Y-force values to control
the forces in the X- and Y-directions to zero, thus eliminating
skating. Without compensation, due to the compliance in the
robot transmission and links, deformations of up to several
millimeters could occur, which would seriously degrade hole
quality and positioning. Although the high-friction contact
between the tripod and the surface improves the sliding
suppression and effective stiffness of the robot, sliding may still
occur if the tangential forces become larger than the break-away
force of the friction contact. This is frequently the case in practice,
as the axial cutting forces will make the tripod ascend from the
surface, thus breaking the friction contact.
3.1. Dynamic modeling and control
The external force control interface exploits the existing inner
motion control structure of the robot system, extended to include
the possibility for external actuation through changing joint-level
servo motion references at a 4 ms sampling time [7]. For the
purposes of simulation and model-based control design, a
dynamic model of the system responses to external forces and
motion references is required. In order to illustrate the expected
properties of such a model, a local model is assumed in the form
Ma x a D1 x_ a Kxa Kxm D2 x_ m f e
Mm x m D3 x_ m Kxm Kxa D4 x_ a f c
where xm and xa are the motor- and arm-side positions in local
Cartesian coordinates, and f c and f e are the (transformed) control
torque and external force on the tool, respectively. Ma and Mm
represent the arm and motor inertias, Di are damping matrices
and K represents elasticity in the transmission and links. Together
with a feedback/feedforward type motion controller
Z  the resonance frequencies.
f c f fb xm ; x_ m f ff xr ; x_ r f I xr  xm dt f ffw
it leads to a full model of the motion-controlled robot. The
position/velocity signals xr and x_ r , corresponding to the references
for the inner motion controllers, are the control inputs to be used
by the external control, f ffw indicating control from measurable
exogeneous disturbances. In the presence of a constant non-zero
external disturbance force f e f 0 and zero reference xr 0, in
29
stationarity we obtain the equilibrium
xm 0 (4)
xa K1 f 0 (5)
As the Cartesian stiffness matrix K will not be diagonal in general,
this means that the deection in the tool position will not be in
the direction of the external force. In particular, the axial forces on
the drill and pressure foot will make the robot bend and deect
also tangentially to the surface, causing a sliding motion which
must be compensated for.
The model in Eqs. (1)(3) is not sufciently detailed for high-
bandwidth model-based control design. The presence of dry and
viscous friction in motors and transmissions, mechanical back-
lash, and partially unknown parameters make it necessary to nd
a more detailed model. By exploiting the special structure of the
Eqs. (1)(3) using a typical robot motion controller, it is possible
to tune models that capture the true behavior of the controlled
robot more accurately [20,21]. In this way, models capturing the
behavior of the controlled robot were experimentally obtained. In
order to obtain better reliability, the tuning procedure was divided
into a static and a dynamic step. In the rst stage, an algorithm
for static calibration was used to nd the stiffness matrix K. In
the second stage of the tuning, a dynamic model including the
motion-controlled rigid robot dynamics, resonances, and cou-
plings between the motion directions was tuned, given motion
references xr and external tool forces f e as inputs. A pseudo-
random binary excitation signal was sent as a motion reference to
the unconstrained robot, and both motor- and arm-side positions
were measured. A discrete-time inputoutput model of the arm-
side motion was found by minimizing the quadratic criterion
2
X
N Xna  
JA; B; F; G; H 4 Aj xa k  j  Bxm k  1
kna j0
32
nm
X zX
m 1
Fj xm k  j  Gi xa k  i  Hxr k  15
j0 i1
(6)
subject to the constraints
X
na X
nm zX
m 1
B Ak ; Fk H; Gk 0 (7)
k0 k0 k1
stating the physical property that the system was statically
decoupled by the integral action, giving a system with unit static
gain. An extended approach including friction parameters was
outlined but was not used in the nal experiments, as friction was
shown to be handled more accurately on a lower level by the
(1) position/velocity control servos of the built-in controller. For this
(2) reason, the identied model used in the experiments was a linear
MIMO system.
For validation of the obtained model, the true reference step
responses were compared to the step responses predicted by the
obtained model. The resulting reference step responses for a
3-DOF model of translation only can be seen in Fig. 8, and the true
and simulated arm-side responses to an external force f e can be
seen in Fig. 9. A good t was obtained both statically and around
(3)
3.1.1. Environment properties
During stiction contact between the tripod and the drilled
component, the contact behavior was similar to a very stiff and
poorly damped spring, as predicted by many friction models such
as the LuGre model [22,21]. When the tangential forces became
larger than the break-away forces of the stiction, the tripod started
to slide across the surface, with poor hole quality and positioning
 ARTICLE IN PRESS

30 T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438

1.5
1
0.5

xza [mm]
0
0.5
1
1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t [s]

0.5
xzm [mm]

0.5

1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t [s]

Fig. 8. True arm-side (top) and motor-side (bottom) reference step responses (solid lines) in the z-direction, and the corresponding step responses of the model obtained by
the tuning procedure (dashed lines). A modied and aggressively tuned velocity feedforward, in combination with the increased exibility represented by the drilling tool,
gave a faster and more resonant response to motion references than the built-in controllers.

as a result. Therefore, it was important both to control the was given by

tangential forces so that sliding was avoided, and to control the " #!
X
3 Ki Xi def:
moments to keep the tripod in contact with the surface at each of fe xa Ke xa (10)
the three contact points. Once such a contact had been achieved, skewXTCP
i Ki Xi
i1
the dependence of the contact force f e on the tool position xa was with the skew-symmetric matrix skewx skewxT represent-
expressed through the environmental dynamics. The resulting ing the cross-product, skewxy x  y. For the setup described
high-gain feedback loop affected the stability and performance of in this work, where the contact points were placed symmetrically
the manipulator. For a point contact, the environment dynamics
on a circle in the xy-plane with radius r around the origin, and
can often be approximated by a local stiffness, or as a (non-linear) where each point stiffness matrix Ki diagkx ; ky ; kz was as-
spring-damper [23]. For the tripod contact of the drilling tool used
sumed to be completely decoupled, the contact model was also
in this work, it was necessary to take also the geometry of the decoupled with stiffness matrix
contact into account. The contact was considered as a combina-
tion of three-point contacts, where the force acting at each point Ke diag3kx ; 3ky ; 3kz ; 1:5r 2 kz ; 1:5r 2 kz (11)
contributed to the effective force and moment acting at the robot
The contact model in Eq. (10) provided a useful local approxima-
TCP point. In the drilling application, using a position/velocity-
tion during stiction, and the objective of the control was to keep
controlled robot, the contact properties of the small tripod could
the system in this stiction regime.
be expressed as a (non-diagonal) stiffness matrix. For the model of
the 5-DOF system used in the experiments, the tool position
3.1.2. Control design
In the control design it is necessary to take both robot
xa tTa ua;x ua;y T (8)
dynamics and environment properties into account. Therefore,
in addition to the model-based control, the option to tune
was represented by the three translations ta and the two rotation
controllers manually in order to account for poorly modeled or
angles ua;x and ua;y around the x- and y-axes of Fig. 1, expressed in
varying environment parameters is desired. Automatic design
a xed coordinate system which was taken to coincide with the
procedures for force controllers have been presented previously
initial position of the tool, xa 0. The coordinates Xw i of each
[24], but they are not suitable for the drilling application due to
contact point in the world coordinate system were related to xa by
the signicantly higher contact stiffness and special control
objectives. Instead, we suggest the use of a control strategy based
Xw
i Rx ua;x Ry ua;y Xi
TCP
ta
2 3 on an easily tunable force controller, for adapting the performance
1 0 0 0 zTCP
i to the environment properties, in combination with an inner
6 7
6 0 7 def: motion controller with model-based disturbance rejection and
 6 0 1 0 zTCP i 7xa Xi xa (9)
4 5 decoupling. Instead, we propose a control strategy based on an
0 0 1 yTCP i xTCP
i easily tunable force controller using an inner motion controller
with model-based disturbance rejection and decoupling. In inner-
where XTCP i xTCP
i yTCP
i zTCP
i  describes the TCP coordinates motion force control, the measured contact force and force
of the contact points of the tripod. Assuming a linear stiffness reference are used as inputs in the integration of a motion- or
f i Ki Xw
i at each contact point, the full contact stiffness model impedance equation. This relation is often chosen as a passive
 ARTICLE IN PRESS

T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438 31

0.5

xxa [mm]
0

0.5
6 8 10 12 14 16 18
t [s]

0.5
xa [mm]
y

0.5
6 8 10 12 14 16 18
t [s]

1.5

1
xza [mm]

0.5

0.5

1
6 8 10 12 14 16 18
t [s]

Fig. 9. True arm-side external force responses (solid lines) in the x-, y- and z-directions, and the corresponding responses of the model obtained by the tuning procedure
(dashed lines).

second-order system in order to emulate the behavior of a passive ef

mass-spring-damper. The robot motion controller is set to track
the output position from the impedance equation. Because of fe
the limited bandwidth of the motion control system and the + + JR3 F/T Environment
fr
deformations of the robot caused by external forces, the tracking
of the desired motion may be poor when the robot is in contact
with a stiff environment. In order to improve the tracking xa
performance, compensation of the effects of the measured Force ctrl.
external forces was included in the inner-loop motion control. fe
This can be seen as trying to increase the stiffness of the robot
as seen from the tool, which improves the ability to control
contact forces and moments. Although the compensation can only
improve the rejection of disturbances up to a frequency limited by
+ r
Dist. comp. xr
the mechanical and controller bandwidth, it is still effective as xre f Robot
long as the variation of the sliding forces is slower than the
xm
bandwidth, which is usually the case in practice. We used a +
controller structure which includes this inner-loop compensation
as in Fig. 10. In order to track the desired position obtained by
em
integrating the impedance relation, the inner motion controller Fig. 10. Simplied structure of the control system for drilling, with an outer force
should have both a fast arm-side response to motion commands, control loop and inner-loop disturbance compensation.
and good suppression of external forces up to the desired
bandwidth of the system. In addition to force sensors, which measurements were used and an inner controller was designed
can be used to obtain improved disturbance suppression through to give a faster and more decoupled response in contact [20]. The
feedforward, feedback from arm-side position measurements controller was chosen to give a proper suppression of distur-
could be used in the inner controller to improve the absolute bances f e at the arm-side position xa , for frequencies up to
accuracy of the positioning. Such measurements could be approximately 25% of the mechanical bandwidth, in order to
obtained from, e.g., cameras or laser trackers. Here, force suppress the dominant sliding effects.
 ARTICLE IN PRESS

32 T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438

The force controller was designed by tuning a desired transfer matrix

decoupled impedance in the form
2
d d
MI Dx DI Dx f e;f  f r
dt
2 dt
r xref Kdc Dx
with MI and DI diagonal matrices. Since the inner-loop design was
based on a 3-DOF model with translation only, a static decoupling
matrix Kdc was included for improved decoupling between the
control of xy-torques and xy-forces. The use of a purely static
method for the decoupling of the rotation can be motivated by the
much lower demands on the bandwidth of the moment control,
as compared to the linear forces. The reason for the difference is
that the moment control only needs to keep the tripod aligned
and in stable contact with the surface, and is not as strongly
affected by the fast variation of the forces that cause the deection
and sliding. The addition of the matrix Kdc in the feedback loop
will result in a modied impedance with respect to external
forces on the TCP. Used together with the dynamic compensator
in Fig. 10, this means that the dynamic force response will
exhibit a better decoupling, so that sliding does not occur when
external forces or torques are applied in other directions. The
matrix Kdc should therefore be designed to remove the couplings,
while the diagonal elements should remain the same in order
not to remove the desired compliance in the force-controlled
directions.
A proper choice for Kdc could be found from the static
calibration data, as described in [20,2527]. Here, Kdc was chosen
such that for all unit basis vectors ei
Dxm;i Kdc eTi Dxm;i ei
where Dxm;i was the motor-side motion required for a force
change Df e ei in stationarity. The arm-side response of the
full system to external forces in Fig. 10 was described by the
Gtot z Gc;af z Gc;ar zGI z (15)
where Gc;af z and Gc;ar z were the responses in the tool position
(12)
xa of the closed inner loop to forces f e and motion references r,
(13) and GI z represented the discretized dynamics in Eqs. (12) and
(13) from applied force to r. The stability of the resulting system
could be analyzed by considering a system with Gtot z connected
to the environment dynamics in a simple feedback loop.
4. Experiments
The rst part of the experiments were carried out at Lund
University, using a smaller ABB Irb 2400 industrial robot equipped
with a pneumatic Atlas Copco LBL25 drilling machine with 4 mm
drill diameter. The contact forces were measured using a JR3
force/torque sensor, and the work piece was a 3.5 mm plate
of high-strength aluminum. In this part of the experimental
program, a number of experiments were performed in several
different robot congurations, with the purpose of evaluating the
effect of the sliding suppression control. The evaluation was done
by comparing the results from two different sets of experiments,
corresponding to controllers with and without sliding suppres-
sion, respectively.
In the rst set of experiments, the built-in motion control of
the robot was used without force compensation in the inner loop.
The approximate stiffness of the robot and tool, with respect
to forces applied at the drill tip, was 160 N/mm in the axial
z-direction, and 100 and 50 N/mm in the tangential x- and
y-directions. There was also a signicant static coupling, resulting
(14) in a tangential deection when axial forces were applied to the
drill. In Figs. 11 and 12 the results from one of the drilling
experiment with 3-DOF force/torque control are shown. The axial
z-force and the moments around the x- and y-axes were
controlled such that a stable contact was achieved with a total

150
100
50
0
fe [N]

50
100
150
200
250
0 5 10 15 20 25 30
t [s]

2
c [Nmm]

2
0 5 10 15 20 25 30
t [s]

Fig. 11. The forces during a drilling experiment using the built-in motion controllers for the inner-loop control, with no control of the sliding forces. Top: sliding forces in
the x-direction (solid) and in the y-direction (dashed), and normal (axial) forces (dash-dotted). Bottom: contact moments acting on the TCP point around the x-axis (solid)
and y-axis (dashed). In this case, large tangential forces are built up in the uncontrolled x-direction.
 ARTICLE IN PRESS

T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438 33

axial force of 200 N. However, although the axial force in Fig. 11
was accurately controlled to the desired value, the friction forces
between the tripod and the surface were not sufcient to be able
to suppress the sliding motion of the tool. This can be seen in
Fig. 12, as sliding occurred primarily in the x-direction, both
during the application of the pressure foot onto the surface, and
when the cutting forces were applied during the drilling. When
forces were applied, the tripod contact switched from stiction to
slip and back again several times. This behavior is also indicated in
Fig. 11, which shows the presence of large x-forces with fast recoils
0.5
0
x a, xy [mm]
at transitions between stiction and slip. The total deection in the
experiment was approximately 1.6 mm, of which 0.8 mm occurred
during the drilling phase. Both the positioning and quality of the
resulting holes were unsatisfactory.
In the next set of experiments a 5-DOF force/torque control
with an inner-loop force compensation was used, in which the
forces in the x- and y-directions were controlled in order to
suppress sliding. Figs. 1315 show the results of a drilling
experiment where this controller was used. Except for this change
of controller, all other parameters were identical to the previous
set of experiments. In Fig. 15 it can be seen that the tool deection
when forces were applied during the force build-up was reduced
to approximately 0.1 mm, and sliding during the drilling phase
was reduced to 0.1 mm. Having performed a number of experi-
ments in different congurations, the model-based force con-
troller was always able to control the sliding forces so that the
tripod contact remained in the stiction regime during the entire
drilling operation, and the tangential deformation was always
below 0.3 mm. This resulted in greatly improved mechanical

0.5 stiffness and vibration suppression, leading to signicant im-

provements in hole quality and positioning.
A second, more rigorous, test program was performed in the
1 robot lab at Linkoping University. Test coupons were installed in a
test rig at four different locations in the robot work envelop, see
Fig. 16. In each experiment the skating deviation was measured
using two LVDT sensors (Linear Variable Differential Transducer).
1.5
The test program included four stations and on each station three
tests were made:
0 5 10 15 20 25 30
t [s] (1) to press with different clamp-up forces on the surface using a
Fig. 12. The linear deection of the tool in the x- and y-directions during a drilling
hard foundation;
experiment using the built-in motion controllers for the inner-loop control, with (2) to press with different clamp-up forces on the surface using
no control of the sliding forces. Undesired sliding of around 1.6 mm occurs in the an elastic foundation;
tangential x-direction, caused by the axial forces. (3) to keep the clamp-up force constant and drill;

150
100
50
0
fe [N]

50
100
150
200
250
0 5 10 15 20 25 30
Time [s]

2
c [Nmm]

2
0 5 10 15 20 25 30
Time [s]

Fig. 13. The forces during a drilling experiment using an inner-loop controller with compensation for the robot compliance, and with active control of the sliding forces.
Top: sliding forces in the x-direction (solid) and in the y-direction (dashed), and normal (axial) forces (dash-dotted). Bottom: contact moments acting on the TCP point
around the x-axis (solid) and y-axis (dashed). The tangential forces are controlled to keep the friction contact in the stiction regime.
 ARTICLE IN PRESS

34 T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438

Bode Diagram
101

102

xx
103

104
Magnitude (abs)
xy

105
xz

105
102 101 100 101 102 103
Frequency (rad/sec)

Fig. 14. Bode diagram for identied transfer functions of robot in non-constrained motion from x-component of external force f e to TCP velocities x_ x , x_ y , x_ z for reference
model of decoupled impedance (or admittance) (red dash-dotted graph); impedance control without inner-loop compensation of compliancecf. Fig. 10 (black dashed
graph); decoupled impedance control (blue solid graph).

0.5
xa,xy [mm]

0.5

1
0 5 10 15 20 25 30
t [s]

Fig. 15. The linear deection of the tool in the x- and y-directions during a drilling
experiment using an inner-loop controller with compensation for the robot
compliance, and with active control of the sliding forces. The drill sliding is Fig. 16. Test coupons installed in a test rig at four different locations in the robot
reduced in the critical drilling phase by a factor of ve as compared to the previous work envelop.
case.

with measured outputs
(1) forces and torques in all 6-DOF;
(2) skating using LVDT sensors.
The elastic foundation constituted four rubber pillows and two
square shaped plates. When the robot pushed the test coupon
intentionally moved. The idea of a moving test coupon was to
simulate exible skin panels. In this case the LVDT moved along
with the mobile test coupon. The outputs of the tests were the
measured forces and torques in all 6-DOF, and the skating as
measured using the LVDT sensors.
In order to measure the skating during clamp-up, the drill bit
was replaced with a short rod having square cross-section, as seen
in Fig. 17. This solution enabled the LVDT to measure skating as
close to the drill tip as possible. The robot was moved close to the
surface before the measurements were initiated. This small
movement before build-up of the clamp-up pressure caused a
small movement, which was neglected when measuring the
actual skating effect. The force feedback control was initiated as
soon as the force was detected in the force sensor. The test results
 ARTICLE IN PRESS
T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438
showed an LVDT motion of 0.1 mm. It was concluded that most
of the 0.1 mm movement was due to the orientation control of the
end-effector. The reason for this was that the LVDT could not
measure exactly at the TCP point, but rather at a point 34 mm up
along the square shaped stick. This phenomenon was increased
when drilling on exible surfaces, where larger orientation control
actions were necessary.
The drilling was performed with a Guhring cutter with a 5 mm
bore diameter, and the holes where drilled 15 mm apart. The hole-
to-hole cycle time was around 12 s. The actual time to control the
clamp-up without skating and with the end-effector orthogonal
to the surface took around 2.5 s. Note in the force diagrams in
Figs. 1320 that the thrust force from the drill reduces the
pressure between the pressure foot antennas and the work-piece.
Keeping the clamp-up pressure higher than the thrust force is
critical to avoid skating effects. In addition to the coupon drilling
tests drilling was also made on a very exible plate. This test was
made to investigate how well suited this method is to be used in
Fig. 17. Set-up of the LVDT sensors in test conguration.
Fig. 18. Task-level program set-up for testing of drilling in exible airframe skin panels.
35
exible airframe skin panels, as seen in Fig. 19. In the cases where
the structure has low stability, the normal direction of the skin
will change during clamp-up. This is one of the strengths of using
force/torque feedback. In this case the end-effector maintains
orthogonality with respect to the surface during clamp-up,
through rapid rotation of the tool in order to adapt to the
changing surface normal. This was conrmed in laboratory
experiments, drilling on the very exible skin panel.
5. Discussion
As an alternative to controlling the position using arm-side
position feedback, the controller attempts to achieve sliding
Fig. 19. Set-up for testing of drilling in exible airframe skin panels according to
task-level program of Fig. 18.

36
50
0
50
Fx, Fy, Fz, [N]
100
150
200
250
300
350
400
8
6
4
Mx, My, Mz [Nm]
2
0
2
4
6
8
10
Fig. 20. Drilling operation forces F x , F y , F z , M x , M y , M z vs. time for the end-effector with three antennas also reecting the drilling operation cycle time, the plate thickness
being 4 mm with holes at a distance of 15 mm.
suppression by making sure that the tangential interaction forces
are always small enough to keep the contact in the stiction
regime. In practice, the achievable bandwidth of the force control
is limited by the mechanical bandwidth of the robot, as well as by
the bandwidth of the inner motion control. Therefore, the force
control and sliding suppression must be combined with proper
mechanical construction in order to obtain a sufciently stiff
system. In the proposed solution, the stiffness is effectively
increased by the tripod high-friction contact. The contact will
damp and suppress small disturbances, such as vibrations from
the feeding and rotation of the drilling tool, while the force control
and active sliding suppression handles large disturbances at lower
frequencies, such as the slower variations of the cutting forces.
Thereby, a system which is able to reject disturbances over a wide
frequency range is obtained, at a potentially very low cost as the
solution requires only a standard force/torque sensor integrated
into the drilling tool. Besides, the accuracy of force and position
made possible by force control eliminates the need for integration
of expensive, extensive support systems. Whereas there are
difculties with coupling between different directions in the
Cartesian space and the linear-model approximation could be
challenged, it should be stressed that in spite of these approxima-
tions and error sources it was possible to obtain the performance
needed for the process.
The drilling force control system differs from most other
applications of force control, such as polishing, grinding and
assembly, where the force control is used to increase the
compliance rather than to improve the stiffness to force
disturbances. In the drilling system, the model-based inner-loop
compensation improves the stiffness, using one or several local
models of the robot stiffness and dynamics. In order to
experimentally acquire and tune such models in the approach
presented, extra arm-side position measurements must be
available during the off-line tuning phase, as methods based on
motor positions only have proved not accurate enough for
determining the dynamic response to force inputs. As an
alternative approach, joint-space stiffness models could be
converted to Cartesian space and used instead, although this
ARTICLE IN PRESS
T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438
0 5 10 15 20 25 30 35 40
0 5 10 15 20 25 30 35 40
Time [s]
approach has not been investigated due to the lack of available
models for the experiment robot. In cases when dynamic models
or arm-side measurements are not available, acceptable results
can in some congurations be obtained without inner-loop
compensation using proper tuning of the outer force controller.
However, the couplings between different degrees of freedom
may lead to poor performance, and attempts to increase the
bandwidth result in limit cycles and oscillations in the force
control. The model-based approach includes two types of
decoupling, the rst being the dynamic model-based inner-loop
disturbance compensation scheme for the three translational
degrees of freedom, the second the static decoupling of rotation
and translation needed because of the tripod geometry. This
division of the decoupling action into a static and dynamic part is
justied, since the bandwidth requirements on the fast sliding
suppression are much greater than on the comparatively slow
control of the contact torques, used to keep the tripod aligned
with the surface. This structure is a great simplication compared
to a full 5-DOF dynamic model and decoupling, and its usefulness
is validated by the experiments.
In addition to the coupon drilling tests, drilling was also made
on a very exible plate. This test was made to investigate how
well-suited this method is to use with the exible airframe skin
panels, as depicted in Figs. 18 and 19. In the cases where the
structure has low stability, the skin will not maintain the same
normal direction as it had prior to clamp-up. This is one of the
strengths in using force feedback control measuring all 6-DOF
simultaneously. In this case, the end-effector maintains orthogon-
ality with respect to the surface. When the surface normal
changes its direction due to clamp-up, there will be a detection of
normality error in the force/torque control and the end-effector
will rotate to adapt to the changing surface normal. This was
conrmed in the laboratory experiments by drilling on the very
exible skin panel (Figs. 18 and 19). Drilling operation 6-DOF
forces and cycle time are shown in Fig. 20. To our knowledge, use
of force control for drilling in exible sheets has not been reported
elsewhere and, thus, the results cannot be compared with any
such previous results.
 ARTICLE IN PRESS

T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438 37

In addition to the robotic force control interface [7,17], robotic
force control interfaces and their usage have been reported
elsewheree.g., the DLR effort [24], the Comau force control
interface [28], and the ABB force-controlled machining [8].
add this last in the template (line 1394):
oxsl:call-templatename "processComments"4
oxsl:with-param name "prefix" select "Ext"/4
oxsl:with-param name "attributeListNodeSet"
select "AttributeList"/4
o/xsl:call-template4

6. Conclusions In:oxsl:template name "processComments"4

The use of industrial robots in automatic drilling applications modify the line:
has been limited, mainly due to the presence of rapidly varying oxsl:if test "($precomchk PreComment and $prefix Pre)
interaction forces in combination with compliance in gear boxes or ($postcomchk PostComment and $prefix Post)"4
and links. Functionality for high-bandwidth force control in to
oxsl:if test "($ precomchk PreComment and $prefix Pre)
modern industrial robot control systems could potentially lead or ($postcomchk PostComment and $prefix Post)
to robotic drilling systems with signicantly improved perfor- or ($extrapidchk ExtRapid and $prefix Ext)"4
mance, without the use of costly hardware modications and and add in the choose statement:
calibration procedures. In this paper, we have presented methods oxsl:when test "$extrapidchk ExtRapid"4
oxsl:text4 o/xsl:text4
and systems for force-controlled robot drilling. Using a 6-DOF
oxsl:value-of select "$attrvalue"/4
force/torque sensor, an outer force control loop and a model-based oxsl:value-of select "$cr"/4
inner-loop disturbance compensation scheme have been de- o/xsl:when4
signed, and used to control the axial contact force and suppress
the sliding of a tripod contact while the drilling is performed. The
advantage of the proposed controller is demonstrated in repro-
ducible drilling experiments using an industrial robot system.
Moreover, the application potential for high-precision robotic Appendix B. ExtRapid application
drilling operations in airframe assembly has been demonstrated.
MoveJ pos10, v50, fine, drill_TCP
MoveL pos20, v20, fine, drill_TCP
FORCE
Acknowledgments FORCESET ramp_speed : 150;
FORCESET glob_gain : 1;
FORCESET f_switch : 1;
This research was made in the framework of the project FlexAA WaitTime 0.1;
supported by the Swedish Foundation for Strategic Research FORCESET forceRef : -400;
under the program ProViking. This nancial support is gratefully FORCEWAITUNTIL forceOuto 390;
acknowledged. WaitTime 1;
SetDO bus2do8, 1;
The authors are grateful to Magnus Engstrom, Saab Aero-
WaitTime 0.5;
structures, for problem denitions and cooperation within the SetDO bus2do8, 0;
FlexAA project, and to Anders Blomdell for implementation of FORCESET glob_gain : 0;
sensor interfaces and real-time software. WaitTime 8;
FORCESET glob_gain : 1;
FORCESET forceRef : 0;
WaitTime 5;
Appendix A. XML modications and ExtRapid programming FORCESET f_switch : 0;
WaitTime 1;
ENDFORCE
Software example of Delmia XML code MoveL pos30, v50, fine, drill_TCP

In: oxsl:template match "ActivityList"4

The example force tag above denes a force scope in ExtRapid
for pos20. The force is ramped up with the speed of 150 mm/s;
add after first if statement, after line 1048:
f_switch command starts the impedance controller; force
oxsl:variable name "extrapid"
select "AttributeList/Attribute[AttributeName Ref 400 is the clamp-up force; forceOut is a trigger value to
ExtRapid1]/AttributeValue"/4 continue with the next ExtRapid program line; SetDo bus2do8,1
activates the drilling cycle; glob_gain 0 stops the force
Add this if statement at line 1088:
oxsl:if test " $ postcomment ! "4 control loop, which keep the robot steady during drilling.
oxsl:call-template name "processComments"4 As glob_gain 1 and forceRef 0 is executed, the force
oxsl:with-param name "prefix" select "Ext"/4 control starts again and the clamp-up force is reduced to zero
oxsl:with-param name "attributeListNodeSet" before moving the end-effector to pos30 in the Rapid program;
select "AttributeList"/4
pos30 is a position some distance away from the airframe.
o/xsl:call-template4
o/xsl:if4
References
In:oxsl:template match "Action"4
[1] Summers M. Robot capability test and development of industrial robot
add this last in the tamplate: positioning system for the aerospace industry. In: SAE 2005 AeroTech
oxsl:call-template name "processComments"4 Congress & Exhibition, Grapevine, TX, SAE Technical Papers 2005-01-3336,
October 2005.
oxsl:with-param name "prefix" select "Ext"/4
[2] Atkinson J, Hartmann JL, Jones S, Gleeson P. Robotic drilling system for 737
oxsl:with-param name "attributeListNodeSet"
aileron. In: SAE 2007 AeroTech Congress & Exhibition, Los Angeles, CA, USA,
select "AttributeList"/4 SAE Technical Papers 2007-01-3821, September 2007.
o/xsl:call-template4 [3] Degoulange E, Dauchez P, Pierrot F, Prat P. Determination of a reference model
for controlling the deformation of an industrial robot. Application to riveting
In:oxsl:template match "Activity"4 in aeronautics. In: Proceedings of the IEEE/RSJ international conference on
intelligent robots and systems, vol. 1, 1994. p. 34351.
 ARTICLE IN PRESS
38 T. Olsson et al. / Robotics and Computer-Integrated Manufacturing 26 (2010) 2438
[4] Kihlman H. Affordable automation for airframe assemblydevelopment of
key enabling technologies. PhD thesis, Linkoping University, Linkoping,
Sweden, PhD thesis 953; 2005.
[5] Van Duin S. A comparison of indoor GPS versus laser tracking metrology for
robotic drilling. In: Aerospace manufacturing and automated fastening
conference and exhibition, Toulouse, France, 2006.
[6] Van Duin S, Kihlman H. Robotic normalizing force feedback. In: SAE 2005
AeroTech Congress & Exhibition, Grapevine, TX, 2005.
[7] Blomdell A, Bolmsjo G, Brogardh T, Cederberg P, Isaksson M, Johansson R, et
al. Extending an industrial robot controllerimplementation and applica-
tions of a fast open sensor interface. IEEE Robotics and Automation Magazine
2005;12(3):8594.
[8] Zhang H, Gan Z, Wang J, Zhang G. Machining with exible manipulator:
towards improving robotic machining performance. In: Proceedings of the
2005 IEEE/ASME international conference on advanced intelligent mecha-
tronics, July 2005. p. 112732.
[9] Siciliano B, Villani L. Robot force control. Dordrecht, NL: Kluwer Academic
Publishers; 1999.
[10] Caccavale F, Natale C, Siciliano B, Villani L. Integration for the next generation:
embedding force control into industrial robots. IEEE Robotics and Automation
Magazine 2005;12(3):5364.
[11] Johansson R, Spong M. Quadratic optimization of impedance control. In:
Proceedings of the 1994 IEEE international conference on robotics and
automation, San Diego, CA, May 813, 1994. p. 61621.
[12] Kawaji S, Arao M, Chen Y. Thrust force control of drilling system using neural
network. In: Proceedings of the IEEE/ASME international conference on
advanced intelligent mechatronics, vol. 1, Como, Italy, 2001. p. 47681.
[13] Alici G. A systematic approach to develop a force control system for robotic
drilling. Industrial Robot: An International Journal 1999;26(5):38997.
[14] Lee W, Shih C. Control and breakthrough detection of a three-axis robotic
bone drilling system. Mechatronics 2006;16(2):7384.
[15] ABB Robotics. Application manualforce control for assembly. Technical
manual, Ref. 3HAC025057-001, 2005.
[16] Nilsson K, Johansson R. Integrated architecture for industrial robot program-
ming and control. Journal of Robotics and Autonomous Systems 1999;29:
20526.
[17] Johansson R, Robertsson A, Nilsson K, Brogardh T, Cederberg P, Olsson M, et al.
Sensor integration in task-level programming and industrial robotic task
execution control. Industrial Robot: An International Journal 2004;31(3):28496.
[18] Ekman T, Hedin G. The JastAdd systemmodular extensible compiler
construction. Science of Computer Programming 2007;69(13):1426.
[19] Haage M. Flexible interactions with productive robots in partly unstructured
environments. Licentiate thesis LU-CS-LIC:2004-2, Department of Computer
Science, Lund University, Lund, Sweden, ISSN 1652-4691; June 2004.
[20] Olsson T, Robertsson A, Johansson R. Flexible force control for accurate low-
cost robot drilling. In: IEEE international conference on robotics and
automation (ICRA 2007), Rome, Italy, 2007. p. 47705.
[21] Shiriaev A, Robertsson A, Johansson R. Friction compensation for passive
systems based on the LuGre model. In: Proceedings of the 2nd IFAC workshop
on Lagrangian and Hamiltonian methods for nonlinear control, Seville, Spain,
2003. p. 1838.
[22] Canudas de Wit C, Olsson H, Astrom K, Lischinsky P. A new model for control
of systems with friction. IEEE Transactions on Automatic Control
1995;40(3):41925.
[23] Diolaiti N, Melchiorri C, Stramigioli S. Contact impedance estimation for
robotic systems. IEEE Transactions on Robotics 2005;21(5):92536.
[24] Natale C, Koeppe R, Hirzinger G. A systematic design procedure of force
controllers for industrial robots. IEEE/ASME Transactions on Mechatronics
2000;5(2):12231.
[25] Olsson T. High-speed vision and force feedback for motion-controlled
industrial manipulators. PhD thesis, Department of Automatic Control, Lund
University, TFRT-1078; May 2007.
[26] Olsson T, Johansson R, Robertsson A. Force/vision based active damping
control of contact transition in dynamic environments. In: Dynamical vision.
Lecture notes in computer science, vol. 4358. Berlin, Heidelberg, New York:
Springer; 2007. p. 299313, ISBN 978-3-540-70931-2.
[27] Kihlman H, Brogardh T, Haage M, Nilsson K, Olsson T. On the use of force
feedback for cost efcient robotic drilling. In: SAE 2007 AeroTech Congress &
Exhibition, Los Angeles, CA; 2007 SAE Paper 2007-01-3909.
[28] Lippiello V, Siciliano B, Villani L. An open architecture for sensory feedback
control of a dual-arm industrial robotic cell. Industrial Robot: An Interna-
tional Journal 2007;34(1):4653.