This is a repository copy of Robots in machining.

White Rose Research Online URL for this paper:

https://eprints.whiterose.ac.uk/147491/

Version: Accepted Version

Article:
Verl, A., Valente, A., Melkote, S. et al. (3 more authors) (2019) Robots in machining. CIRP
Annals, 68 (2). pp. 799-822. ISSN 0007-8506

https://doi.org/10.1016/j.cirp.2019.05.009

Article available under the terms of the CC-BY-NC-ND licence

(https://creativecommons.org/licenses/by-nc-nd/4.0/).

Reuse
This article is distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs
(CC BY-NC-ND) licence. This licence only allows you to download this work and share it with others as long
as you credit the authors, but you can’t change the article in any way or use it commercially. More
information and the full terms of the licence here: https://creativecommons.org/licenses/

Takedown
If you consider content in White Rose Research Online to be in breach of UK law, please notify us by
emailing [email protected] including the URL of the record and the reason for the withdrawal request.

[email protected]
https://eprints.whiterose.ac.uk/
CIRP 2028 1–24

CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

Contents lists available at ScienceDirect

CIRP Annals - Manufacturing Technology

jou rnal homep age : ht t p: // ees .e lse vi er . com /ci r p/ def a ult . asp

Robots in machining
Alexander Verl (2)a,*, Anna Valente (2)b, Shreyes Melkote (1)c, Christian Brecher (1)d, Erdem Ozturk (2)e,
Lutfi Taner Tunc f
a
Institute for Control Engineering of Machine Tools and Manufacturing Units (ISW), University of Stuttgart, Germany
b
SUPSI —– University of Applied Sciences and Arts of Southern Switzerland, Switzerland
c
George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, USA
d
Laboratory for Machine Tools and Production Engineering (WZL), RWTH Aachen University, Germany
e
Advanced Manufacturing Research Centre (AMRC), The University of Sheffield, UK
f
Integrated Manufacturing Technologies Research and Application Centre, Faculty of Engineering and Natural Sciences, Sabanci University, Turkey

A R T I C L E I N F O A B S T R A C T

Keyword: Robotic machining centers offer diverse advantages: large operation reach with large reorientation
Robot
capability, and a low cost, to name a few. Many challenges have slowed down the adoption or sometimes
Machining
inhibited the use of robots for machining tasks. This paper deals with the current usage and status of
Process planning
Control robots in machining, as well as the necessary modelling and identification for enabling optimization,
process planning and process control. Recent research addressing deburring, milling, incremental
forming, polishing or thin wall machining is presented. We discuss various processes in which robots need
to deal with significant process forces while fulfilling their machining task.
© 2019 Published by Elsevier Ltd on behalf of CIRP.

1. Introduction blasting), as well as metrology and inspection functions, while

flexibly implementing various process strategies by exploiting
Robots are currently used in industrial machining operations several end effectors or even hybrid process chains on different
which involve relatively low cutting force requirements such as metals and non-metallic materials. A manufacturing robot is
trimming, drilling, and polishing on composite parts, as well as typically a robotic arm manipulator on a fixed base that performs
drilling or deburring, grinding, and milling on metal parts. There repetitive tasks within a local work cell [3]. High-speed robots are
are several aspects that stimulate the use of robots in machining. hazardous and demand the exclusion of humans from robotic work
The first and main one is the cost — robots cost less compared to places, where the robots are typically placed inside a cage. In
machine tools with the same work space. contrast, collaborative robots are considered safe around humans
During the last decade, a new global sales record has been because they operate at low speed and stop moving if they
achieved for industrial robots (IRs) almost every year. In 2016, a encounter an obstruction.
total of 294,000 units were sold, an increase of 16% compared to When it comes to defining machining robots, the attempt to
the previous year. This number soared to 381,000 units in 2017, draw a sharp line to differentiate between them and a machine tool
another 30% more than the previous year [1]. More than 3 million is not straightforward in a general sense and depends on the
installed robots are forecasted by 2020, doubling their number in specific scientific and technical discipline (e.g., machine tools are
industrial operation since 2014 [2]. According to the International classified as Cartesian robots in the robotics scientific discipline).
Federation of Robotics (IFR), in 2016, 1.4% of IRs were intended for The very term “machining” relates to specific technologies
mechanical cutting, grinding, deburring, milling, and polishing implemented with various tools (DIN 8580) but embeds in its
[1]. This number is low compared to non-processing operations etymology the fact that such technologies are implemented on
like handling or assembly, which demonstrates the lack of the use conventional machine tools. Machining robots should be consid-
of robots in machining. However, these developments illustrate the ered a subset of the manufacturing robot family. They can generally
importance of IRs for machining applications. be defined as robots executing manufacturing processes, according
Within the large set of operating resources, manufacturing to DIN 8580.
robots are computer numerically controlled (CNC) machines Like machine tools, machining robots can accomplish multiple
conceived for industry for specific material subtraction or machining tasks by exploiting different configurations, kinematic
deposition functions (e.g., milling, coating, welding or powder structures and performance targets while integrating a tool (end
effector according to the robotics nomenclature) that performs the
machining process when in contact with the material. The end
* Corresponding author. effectors integrated on a machining robot can present similar
E-mail address: [email protected] (A. Verl). capabilities as tools mounted on machine tools. Their selection

https://doi.org/10.1016/j.cirp.2019.05.009
0007-8506/© 2019 Published by Elsevier Ltd on behalf of CIRP.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

2 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

relies upon product and process driven criteria, along with robot The kinetic and potential energy are described for each link in
payload and dynamics. By means of the tool, machining robots can the system. The potential energy is defined as the product of the
convert a blank into a final product by executing operations mass mi of the ith link, the gravitational acceleration g, and the
controlled by the robot’s controller unit. Similarly to machine tools height of the center of mass (opposite to the direction of gravity)
[4], the motion of a machining robot can be controlled by a CNC for the ith link:
unit, as the machining strategy and trajectory are generated by a
V i ðqÞ ¼ mi ghi ðqÞ ð2Þ
CAD/CAM (computer aided design/manufacturing) system. The
productivity and accuracy of the machining operations depend on In comparison to the potential energy, the kinetic energy is,
the preparation of the numeric control (NC) programs, the robot however, far more complicated to describe and follows as:
path planning, motion strategy, and dynamics optimization,
1 T Xn
together with the robot behavior assessment in the working area. T¼ q_ ½mi Jvi ðqÞT Jvi ðqÞ þ J vi ðqÞT Ri ðqÞIi Ri ðqÞT J vi ðqފq_ ð3Þ
In metal cutting, for example, there is a growing number of 2 i¼1

successful installations where robots have been used in a similar with the rotational transformation Ri and the inertia matrix Ii for
manner to a machine tool or to a person carrying a metal cutting each link i. A central part in describing the kinetic energy is the
hand tool [5]. The current paper will outline and investigate the manipulator’s Jacobian J which is a 6  n matrix for a robot with n
opportunities and the range of applications where the adoption of links. The Jacobian contains parts for the linear velocity Jv and the
robots for machining constitutes an instrumental leverage. angular velocity Jv, which are described exemplary for rotational
joints in the following equations [9]:
1.1. Standard definition for robots in machining  
J
J¼ v ð4Þ
Jv
The first patent for what we would now consider a robot was
filed in 1954 by George C. Devol and issued in 1961 [6]. Since then, with:
the conception and exploitation of robots in industry has brought
disruptive innovation while posing complex scientific and techni- Jvi ¼ zi 1  ðO0n O0i 1 Þ ð5Þ
cal challenges to the scientific communities.
Nowadays, the robot is universally defined as a goal-oriented Jvi ¼ z0i 1 ð6Þ
machine that can sense, plan and act. According to ISO/IEC
2382:2015 (en), a robot is defined as a mechanical device, usually The Eqs. (5–6) are illustrated in Fig. 1. In this case, is the z0i 1
programmable, designed to perform tasks of manipulation or rotation of joint i-1 and ðo0n o0i 1 Þ is the distance between the tool
locomotion under automatic control. Similarly, ISO 18646-1:2016 center point (TCP) – marked in Fig. 1 as a point at the tip of the
(en) considers a robot as a program actuated mechanism with a manipulator’s nth body – and the joint i – 1. The zero indicates the
degree of autonomy, moving within its environment to perform world frame as reference frame. In the end, the motion equations
intended tasks. An industrial robot is, in turn, described as an are described with partial derivatives of the Lagrangian:
automatically controlled, reprogrammable, multipurpose manip- d @L
ulator, programmable in three or more axes, which can be either ð7Þ
dt @q_ i @@qL ¼ t i
fixed in place or mobile for use in industrial automation
applications [7]. where t i represents the actuator torques and other nonconserva-
In the following sections, we will introduce the theoretical tive, generalized forces acting on the ith joint. These equations of
foundations of robotics and then cover the programming tool chain motion describe the actuator’s dynamics without influences such
in robotic machining. Then, an overview of several robotic as friction or reaction torques and forces induced by the
machining processes is presented. Finally, we look at emerging environment. For more detailed modelling reference is made to
fields of research in this area and draw a number of conclusions. the relevant literature [8].

2. Theoretical foundations

This chapter summarizes the theoretical background needed

for robot machining tasks. Using a robot model enables robot
programs to predict behavior and process parameters. In order to
model a robot, the first section of this chapter describes important
mathematical equations and parameter identification methods.
Considerations on mechanical and geometrical properties of
robots, followed by aspects regarding robot stiffness are then
introduced. Furthermore, the static and dynamic behavior of
industrial robots is explained, and various test results are
presented. Path planning, robot control and process control
strategies are also indispensable when preparing robots for new
tasks, and are also discussed in detail in this chapter.

2.1. Robot modelling

Fig. 1. Calculation of the manipulator’s Jacobian [8].

There are different approaches to describing a mechanical
system analytically. The method of Lagrangian equations described
in this section is widely used and an easy way to derive equations Many system identification approaches have been developed
of motion for rigid body robots. during recent decades in order to determine system parameters
Let q 2 Rn be coordinates of the joint angles for an industrial like masses, components of the inertia matrix, components of the
robot in the n-dimensional vector space. First, the Lagrangian L is moments for each link, as well as friction parameters and motor
defined as the difference between the total kinetic and the inertias. Most dynamic identification methods for robots use the
potential energy of the system (T and V, respectively) [8]: inverse dynamic identification model, which calculates motor
torques with the aforementioned equations of motion. Further-
Lðq; q_ Þ ¼ T ðq; q_ Þ V ðqÞ ð1Þ more, least squares techniques are used to estimate the

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 3

parameters. In some cases, a few parameters like masses or robot dynamics on the instantaneous tool-workpiece interaction
kinematics are already known and therefore do not need to be in robotic milling with a KUKA KR270 industrial robot. They
identified. As the optimal method depends on the robot and on modelled and analyzed the effect of robot compliance on the
these known parameters, reference is made to literature contri- instantaneous tool tip displacement due to periodically varying
butions on these topics [10–12]. In addition, there are many ways milling forces. However, they did not consider the effect of robot
to optimize path planning for data acquisition for the identification compliance on the milling forces themselves. Recently, Cen and
process. To increase the least squares methods convergence rate Melkote [21] modelled and analyzed the effect of robot compliance
and lower the susceptibility to noise, an optimal excitement and on the instantaneous milling forces. Through simulation and
implicitly an optimal trajectory is crucial for the process. Hence, a experiments, they showed that the peak milling force (Fig. 2), and
sequential identification process is often used for estimating the hence the tool tip deflection, can be significantly under predicted if
different parameters with adapted trajectories [13]. this effect is ignored. This effect is important in accounting for
offline robot compliance compensation strategies.
2.2. Methods to describe mechanical and geometrical properties

Machining of a workpiece requires a geometrical description of

the machining task and a reference coordinate system in which the
task is described. The task described in these coordinate systems is
then converted to the machine coordinate system by probing. To
execute the machining task, a further transformation to the
machine axis coordinate system (or joint coordinate system) is
required. Machine tool structures with three translational and
optional rotational axes have known and defined transformations.
This is not the case for robotic structures including mostly
rotational axes with optional translation axes.
In robotics literature, the so-called Denavit–Hartenberg (DH)
convention [14] is considered to be the defining set of rules on how
to attach these coordinate systems to the robots’ linkages. The
Fig. 2. Effect of robot dynamics on predicted and measured milling forces and their
underlying principle of the DH method is describing the
frequency contents [21]. Rigid model (a) ignores robot compliance.
transformation between two linkages with only four parameters,
representing a combination of two rotations and two translations Recently, Cordes et al. [22] presented a detailed analytical study
[15]. The six parameters generally needed are reduced to four by of chatter in robotic milling of aluminum and titanium structures.
applying the convention’s rules — and limitations. Although the Their work shows that the robot's pose-dependent vibration
underlying principle remains the same, four different variants of modes result in chatter only at low spindle speeds (when milling
notations are used in scientific literature [16]. Classification is titanium). In contrast, only pose-independent vibration modes
based on the index linkages and parameters, which, in conse- cause chatter at high spindle speeds (when machining aluminum).
quence, affects the grouping order of the transformation matrices They subsequently developed a stability chart that can be used to
of the rotations and translations. Instead of attaching only one select chatter-free spindle speeds when machining aluminum. In
coordinate system to each joint, the Sheth–Uicker (SU) convention their robotic milling experiments on a KUKA KR210 IR, Cen et al.
[17] attaches two, an input and an output coordinate system. The [23] also found that chatter instability occurs more readily when
relative motion between the two is exactly the joint motion milling in one Cartesian coordinate direction compared to milling
between the two linkages. The DH parameters can still be used to in a direction orthogonal to it. Mousavi et al. [24] have also shown
describe the relative transformation between the two coordinate that the natural frequencies of a serial-link robot can vary
systems attached to one linkage, however, the joint variable term significantly along a tool path due to a continuously changing
of the DH parameters for the linkage transformation is only a robot configuration. They used three dimensional Euler–Bernoulli
constant offset. Although the DH convention operates with half as beam elements with a matrix structural analysis technique. Similar
many transformations as the SU convention, it is worth noting that findings have been reported for a parallel kinematic robot
the SU convention allows for a more convenient way to define (hexapod) [25]. These observations emphasize the importance
coordinate systems of mechanisms that are not strictly serial. It can of accounting for robot dynamics during robot trajectory (motion)
also handle complex joint types more conveniently, with more plan generation for machining applications.
than one degree of freedom.
2.4. Measurement of static properties
2.3. Robot stiffness
Karim et al. [26] analyzed the Cartesian static compliance
Robot stiffness is the key parameter that influences machining behavior of a KUKA KR500-3 MT robot in numerous poses for
accuracy. Specifically, the limited static and dynamic stiffness of tensile and compressive loads in all directions. The detailed
robot joints and links results in insufficient rigidity of the robot measurement methodology is described in Ref. [26]. The results
TCP, which in turn impacts the machining accuracy. As pointed out were interpolated and graphically displayed over the workspace.
by Pan et al. [18], the static Cartesian stiffness of industrial robots is Large nonlinearities regarding static compliance values, as well as
on the order of 1 N/mm while the corresponding stiffness of a a non-symmetric distribution of these to the robot’s x-axis have
typical CNC machine tool is up to 50 times higher. Unlike a CNC been measured experimentally. Fig. 3 shows the static compliance
machine tool, the Cartesian stiffness of an industrial robot varies values in the y-direction (described in detail later in Fig. 11) for
with its joint configuration (or pose) throughout the workspace loads in that direction. In this plane, the measured static
[19], which complicates the task of machining process control. The compliance rises with the x-coordinate. The values at negative
magnitude and nature of robot deflections are a function of the y-coordinates are lower than at positive y-coordinates.
magnitude and frequency content of the process forces generated Fig. 4, in contrast, shows the distribution of the static
during machining. compliance in the x-direction of the xz-plane for loads in the x-
In recent years, the effects of machining process parameters direction for an exemplary result. The maximum static compliance
and robot configuration on the dynamic characteristics of the values in this case are located in the middle of the workspace.
robotic machining process and its stability have been investigated. Overall, the lowest values have been measured for loads in the x-
Bondarenko et al. [20] presented a simulation study of the effect of direction and the highest for loads in y-direction.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

4 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

exploited in order to automate the measurement of numerous

points on the robot’s structure.

2.4.1. Static stiffness modelling and measurement

A key requirement for offline robot compliance-induced error
compensation methods is knowledge of the robot stiffness. An
example of measuring static stiffness comes from Cognibotics [29–
31]. Their approach is that robot mechanics can be modelled and
Fig. 3. Static compliance values in y-direction in the xy-plane (z = 900 mm) for identified, and not just identified assuming stiff arm components.
tensile (left) and compressive (right) loads in the y-direction; x and y axes in [mm], There are three areas to be modelled and identified: (1) arm
respectively.
kinematics, including individual variations due to differences in
the tolerances and assembly of the arm components; (2) joint
flexibilities including backlash and bending inside the gearbox;
and (3) link flexibilities, including bending arm castings and
bearings. Modern symbolic manipulation languages enable the
creation of robot models which can be used in the next step:
identification.
Lehmann et al. [32] used the internal measurement system of
the robot arm, and torque data in one or more clamped positions in
order to bend the robot arm under controlled and repeatable
Fig. 4. Compliance values in the xz-plane (y = 0 mm) for tensile (left) and conditions. Data from controlled bending motions is used in their
compressive (right) loads in the x-direction; x, y axes in [mm] respectively. study to identify the parameters of an elasto-kinematic model of
the arm joints, gearbox, and links. This model and data are then
According to Karim et al. [27], the preferred feed direction for combined to identify the normal or stiff kinematic model of the
machining tasks should therefore be in y-direction to minimize arm. The remaining parameters are then found by relying upon
deflections of the TCP from the commanded path. Although specific measurement approaches. The final step is to apply this
differences between static compliance values for tensile and model to the machining process. This method performs a two-
compressive have been measured, the variation is negligible. The stage machining process. In the first stage, the elasto-kinematic
measured values for the Cartesian static compliance were used to model is not used, but the nominal robot parameters and
calculate the joint compliance values for all six robot axes robot programmed points are used to perform the first stage machining.
using the approach from Ref. [28]. The results are shown in Fig. 5. In the second stage, the knowledge of the recorded process
torques, motor angles, and the elasto-kinematic model allows an
accurate estimation of the amount of tool-tip deflection that
occurred during the process in the first stage. The tool-tip
deflection is then used to modify the program target positions
during the next and subsequent machining operations.
Another approach (Fig. 6) to measure kinematic and joint
stiffness parameters was developed at the Advanced Manufactur-
ing Research Centre (AMRC, University of Sheffield). Forces in two
orthogonal directions and a moment can be applied on a robot in
this setup. Deflections on the robot are measured with a laser
scanner. After this identification, using the redundant degree of
freedom in the robot, the joint configurations can be optimized for
minimum compliance [33].

Fig. 5. Calculated joint compliance values.

The measurement protocol and the pose description can be

obtained from Ref. [28]. The calculated joint compliance values
exhibit a strong dependency on the pose (position and orientation
of the TCP). For instance, the calculated joint compliance for joint
three is almost zero in general, but in some poses, the value Fig. 6. Setup for identification of joint stiffness of a robot.
represents the overall maximum of all calculated values.
Because of the distinct pose dependency and the significant For example, the robot configuration with minimum compli-
differences for loads in different directions model-based ance in X-direction is presented for a given end effector position
approaches, which are presented in many research projects to and orientation in Fig. 7(a). Influence of the orientation is
determine the robot’s static compliance behavior, can only be demonstrated in the color map in Fig. 7(b). The map compares
considered valid in a small area. It is shown that such models need the maximum compliance in x-direction on an xy-plane at a given
to include effects such as nonlinear friction and backlash, the z height and end effector position and orientation achieved by
influence of possible gravity compensations as well as differences selecting the worst joint configuration and the minimum
for all load directions. Even though the determination of these compliance achieved by the optimum joint configuration for the
effects results in a higher effort, an experimental determination of same end effector position and orientation. It presents the ratio of
the static compliance values is necessary for describing the maximum to minimum compliance for the given xy-plane
compliance behavior in a broader work area. Since each robot simulated between 1.6 and 2.6 in Fig. 7(b).
behaves differently this effort increases even more. However, laser This demonstrates that using the redundant degree of freedom
scanners or laser vibrometers, as applied in Ref. [28], can be of the robot and a stiffness model, compliance of the robot in a

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 5

Fig. 9. (a) Effect of cutting parameter (tool feed) on dynamic stability angle, (b)
Schematic illustration of stability enhancement principle [34].

optimize machining results, the expert system recommends

Fig. 7. (a) Robot configuration with minimum compliance in x-direction. (b) Ratio of cutting parameters to plan the process.
maximum to minimum compliance in x-direction in the given 2D work space.
Using experimental modal analysis, Mejri et al. [37] have
characterized the variation in the FRF of the tool tip as a function of
certain direction can be optimized and the improvement can be as vibration direction and robot position (or robot pose) for an ABB
high as 2.6 times. Work in progress aims to validate the IRB 6660 industrial robot equipped with a high-speed machining
improvements in this field experimentally. spindle. For a given robot pose, they found that the modal
frequencies in the two orthogonal directions of excitation can be
2.5. Dynamic behavior different. Fig. 10 shows a typical result for the real part of the FRF,
which governs machining dynamic stability. They also found that
The dynamic behavior of an industrial robot is important for the robot was more stable when machining in the y-direction than
machining quality and accuracy. Especially for milling tasks, in the x-direction, per their convention.
common use-cases in robot machining where high forces occur in a
periodic manner, the robot structure’s dynamic properties are
critical to avoid severe oscillation and consequent machining
errors. Oscillation amplitudes especially escalate when process
forces match the robot structure’s natural frequency and the
direction of vibration modes.
Tunc and Stoddart [25] showed the dynamic stability of the
robotic milling operation to be significantly influenced by the
direction dependence of the frequency response functions (FRF) of
the tool tip along the orthogonal Cartesian directions (x and y)
defined relative to the robot base coordinates. Fig. 8 shows the
oriented frequency response functions of a Fanuc F200iB 6-DoF
(degrees of freedom) parallel kinematic robot as a function of the
feed direction. They proposed an algorithm to determine the tool
path (feed direction) that maximizes the dynamic stability of the
machining operation. This work highlights the importance of
accounting for the continuously varying tool tip dynamics when
determining the most stable cutting conditions for robotic milling.

Fig. 8. Oriented FRF variation, Fanuc F200iB with feed direction [25].
Fig. 10. Variation in the FRF (real part) of the robot tool tip as a function of excitation
direction(top) for an ABB IRB 6660 robot (bottom) [37].
Cen and Melkote [34] presented a stiffness model for the robot
milling process based on the conservative congruence transfor- To identify the detailed dynamic behavior, like natural
mation (CCT). Their work is based on the recognition that the frequencies and modes of a structure, an experimental modal
Cartesian stiffness of the robot is affected by the external force analysis (EMA) can be performed as was the case in Ref. [38] on a
[35]. This, in turn, alters the robot geometry through the KUKA KR 500-3 MT “machine tooling” robot. The basic concept of
differential Jacobian, because of the elastic deformation of the an EMA is to compare a force of excitation with the reaction of a
joints and links. By adjusting the cutting parameters, it is possible structure and compute a transfer function. Hereby modal
to reduce the angle (g ) between the external force vector and the parameters such as natural frequency and mode shape, which
maximum principal stiffness vector (Kmax) of the robot thereby describe a structure’s natural oscillation, can be determined. While
enhancing the dynamic stability of the machining operation. Fig. 9 the excitation in Ref. [38] took place with an impact hammer at one
schematically illustrates the approach and the dependence of the point close to the TCP, the response was measured with
stability boundary as a function of the cutting parameter (tool feed accelerometers on 63 points spread out across the structure and
in this example). The authors demonstrated the suppression of in all three directions x, y and z. This allows insight into the
mode coupling chatter in robotic milling experiments conducted oscillation of all robot parts. Since the tests were focused on
on a KUKA KR210 industrial robot [34]. the direction- and pose-dependency of the dynamic behavior of
In another work, Lienenlüke et al. [36] introduced an expert the robot, the structure was excited in two directions, y and z (x did
system with a static compliance model that is trained with process not show different modes in preceding tests). Experiments were
data to link process planning parameters with process behavior. To performed in 23 different poses within the robot’s workspace. Joint

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

6 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

configurations included fully stretched out as well as retracted configurations can be avoided to minimize dimensional and
poses. Evaluation of the data showed recurring mode shapes in all surface quality errors.
poses. For robot machining applications, this means that further
Fig. 11 shows the manipulator deformation in the first mode. experiments with different robots have to prove whether the
Segments are color coded for increasing comprehensibility. The qualitative dynamic characteristics of all serial six-axis industrial
robot’s first axis (green) is the joint responsible for the oscillation robots are similar to the results of Ref. [34]. Research results also
that occurs at frequencies between 5.8 Hz and 8.1 Hz depending on suggest that the robot’s joints are the major cause for oscillations
the pose. The second mode showed the manipulator pitching when using off the shelf robots for machining.
between 8.5 and 12.3 Hz, caused by a tilt of the first axis’ bearing.
2.6. Path planning methodologies

Path planning is usually optimized considering two main

objectives: reduced processing time and smoother trajectories. In
the first case, the goal is to minimize the time necessary to execute
a certain trajectory in order to match productivity requirements
[39,40]. In the latter case, a path quality index is favored and time is
treated as a constraint [41]. Some optimization methods consist of
a mixture of the aforementioned ones, proposing a trade-off
between execution speed and quality. In Ref. [42] some of these
approaches are compared taking the execution time and the norm
of jerk into consideration. Since the specific kinematic structures of
many commercially available robots induce uneven distributions
of workload on motion axes, the efficiency and quality of
trajectories is not necessarily a by-product of a geometrically
optimal path planning. This aspect is important not only for
machining segments of the tool path but also for rapid movements,
where optimized trajectories could improve overall lead-time and
Fig. 11. Robot arm deformation in the first oscillation mode. Robot top (above) and
reduce structural stresses.
side views (below), respectively.
Various studies focusing on these topics rely upon the use of
piecewise splines, whose coefficients are optimized using different
Fig. 12 shows a linear interpolation of the natural frequency of approaches. For example, in Ref. [43] a B-spline representation of
the second mode as a function of the TCP position. Black dots mark trajectories is used, and a function that concurrently considers
the measurement points. Higher modes at higher frequencies both the total execution time and the minimization of the norm of
show more complex shapes and movements of axes three and four jerk. Depending on the nature of the chosen objective function, the
[38]. The value of the natural frequencies changed with different optimization approach may range from the classical resolution of a
TCP positions. In general, the more stretched out the manipulator quadratic programming problem to the minimization of a highly
was, the lower the measured frequencies were. non-linear functional. In this case, general-purpose optimization
approaches must be adopted, like in Ref. [44], where the norm of
jerk for the robot joints is minimized numerically by a genetic
algorithm. A similar approach is proposed also in Ref. [45], but in
this case, the chosen objective is a function of manipulability
measures and thus involves inverse kinematics.
As mentioned before, process efficiency and quality may also be
improved by optimizing rapid movements. In these cases, authors
target the planning of smooth trajectories in the Cartesian space,
while downgrading the speed of all joints based on the slowest one
[46] and on achieving a higher degree of regularity by increasing
the polynomial degree of the path or of the motion profile [47–
49]. Multiple-axis movements are often managed in the operations
workspace [50], although the definition of trajectories in joint
space is, in some cases, a viable simplification that allows the
implementation of smoother joint trajectories [51]. A different
approach is implemented in Refs. [52] and [53], where kinematic
quantities are optimized in the joint space, based on a sine-jerk
model for joint motion profiles, and the dynamic characteristics
Fig. 12. Second natural frequency as a function of the TCP position. and limits of each joint are considered to enable an even
distribution of workload along the kinematic chain. Manufacturing
applications in which the robot is subjected to continuous,
The results show a high pose-dependency of the robot’s potentially strong excitations, for example, cold spray [54], could
dynamic behavior and different oscillation directions at different also benefit from path planning strategies that consider the
frequencies. This suggests a high dependency between the process stiffness variation across the whole task space. Mekaouche et al.
execution and the machining quality. This is confirmed for milling [55] present a finite-element-method-tool for mapping robot
experiments [27], additionally showing an impact of up to six structure stiffness over the work space for a generic kinematic
modes on the process, unlike conventional machine tools, where chain.
only the first few modes are relevant. Further investigations have
been conducted to exclude that these vibrations were caused by 2.7. Control
the workpiece itself, the fixture or the tooling. In order to exploit a
machining robot’s full potential, the dynamic behavior of the A robot controller is usually a processor-based electronic or a
system has to be well known. This way, tasks can be carried out in standard personal computer (PC) device that can be programmed
the optimal pose and feed direction for a specific system. Critical to drive the motors attached to each robot axis and coordinate their

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 7

motion while ensuring that the TCP executes specific trajectories The procedure of Klimchik et al. from Refs. [61] and [62] is
with assigned motion profiles [56]. Such trajectories are generated illustrative of the general strategy used for offline compensation
at the CAM level and then fed to the robot CNC in the form of a part (Fig. 14).
program.
The controller can additionally command the digital and analog
input and output (I/O) signals to command external devices, such
as a cutting tool or a welding gun, based on a sequence
synchronized with the robot motion. Such signals are sampled
with cycle times related to the application. The robot controller
communicates with other controllers or PCs and uses sensors to
obtain information about the robot environment, in order to
modify the robot tasks accordingly. For example, images coming
from vision sensors are typically streamed to an external PC that is
connected to the CNC. These images are therefore processed
outside the controller: specific information is elaborated on the PC
and collected by the CNC with a sampling cycle time related to the
process dynamics (from 5 to 150 ms). Fig. 14. Offline process control strategy [62].
When available in the CNC, the timing associated to the actual
implementation of the adapted information is enabled in the part A predictive model is used here to estimate the machining
program by integrating targeted check points during its execution. forces anticipated along the nominal machining trajectory neces-
Fig. 13 outlines the major components of a robot CNC. sary to produce the desired part feature. These are then input to
either a static or a dynamic compliance model of the robot to
compute the optimum joint configurations that will yield the
desired Cartesian positions of the robot tool tip in the presence of
the machining forces. The nominal robot configurations along the
tool path are then modified, based on the optimized joint
configurations, to obtain the modified trajectory, which is then
fed to the robot controller for execution using the forward
kinematics model contained in the robot control software.
Klimchik et al. [62] developed and implemented the offline
compensation strategy on a KUKA KR-270 industrial robot with six
degrees of freedom (DoF). They utilized a time-domain simulation
of the dynamic milling forces using the approach of Altintas [63] to
compute the instantaneous milling forces, which were used to
determine the robot configurations that minimize the compliance
induced errors along the desired trajectory using an iterative
numerical approach. Their experimental results demonstrated
Fig. 13. Robot CNC architecture.
>90% reduction in the maximum deviations of the tool tip due to
low frequency (~7 Hz) robot compliance in the presence of milling
forces. While this result is promising, their work does not address
The trajectory strategy behind robot motion control aims at reducing machined surface errors caused by higher frequency
obtaining high quality products and robust processes [57]. This vibration modes that can be excited by periodic milling forces.
frequently demands the robot CNC to embed the process model As part of the EU FP7 COMET project, Schneider et al. [64]
and a set of process key performance indicators (KPIs) that can be developed an offline error compensation strategy based on a
tracked over time and space. Based on these observed KPIs, specific detailed investigation of multiple error sources in robotic
process optimization strategies can be implemented. The following machining. They proposed a modular approach to overcome
section provides an overview of typical process control strategies accuracy issues either in offline or online. Their study was another
in machining robots. example of how laser trackers can be used in real-time position
compensation for machining robots together with an additional
2.8. Process control strategies piezo actuator based high-dynamic compensation mechanism.
The authors also analyzed several sources of error such as backlash,
A literature review of process control strategies for robotic static and dynamic friction, and nonlinearities in joint stiffness [65]
machining reveals two broad approaches to minimizing the and incorporated these effects in their offline compensation
sources of error and process instability. The approaches consist approach. Dimensional errors as low as 0.05 mm were obtained in
in offline and online methods. Robots are currently used more pocket milling tests on aluminum. In more recent work, Diaz
often in milling [22] and drilling [58–60] especially for large parts, Posada et al. [66] presented their offline error compensation
molds and dies. They offer easier set-up and portability than large strategy (Fig. 15) with experimental verification results obtained
machine tools, but are significantly less stiff than these, hence they on a KUKA Quantec KR270 2700 industrial robot.
cannot be used in all machining applications. The following Fig. 16 shows representative improvement in circularity error
approaches outline some of the opportunities to boost the obtained through offline compensation of cutting force induced
adoption of robotics in machining tasks. compliance of the robot TCP when milling a circular contour in an
aluminum workpiece. Similar tool tip error offline compensation
2.8.1. Offline methods strategies have been reported by other researchers [67–69].
Offline error compensation strategies seek to minimize the In robotic machining, the presence of kinematic redundancy
effect of non-kinematic error sources such as robot joint and link enlarges the workspace of the robot and increases the number of
compliance on the robot positioning accuracy, and in turn, the feasible robot configurations (poses) for machining the desired
machining error, through optimal selection of the robot configu- feature. For example, using a 6-DoF robot to perform a 5-DoF
ration, machining direction, and cutting conditions during tool machining task yields a first-order redundancy. Since robot
path generation using improved static or dynamic models of the Cartesian stiffness varies with configuration, redundancy resolu-
tool-workpiece interaction. tion techniques [71] can be utilized to derive stiffer robot poses,

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

8 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

conditioned poses and away from joint limits. Experimental

assessment of their redundancy resolution method was conducted
on a KUKA manipulator for milling a relatively soft polymeric
material. Xiao and Huan [74] have also proposed different criteria
based on singularity avoidance, joint limit avoidance, and collision
avoidance for use in resolving kinematic redundancy in 6-DoF
industrial robots used for machining tasks.
Sabourin et al. [75] presented a kinematic redundancy-based
path planning optimization method to enhance the machining
performance. They discussed and analyzed a number of different
criteria including kinematic, mechanical advantage, and stiffness
to improve various performance metrics of the robot for a
machining task. Work cells with a 9-DoF serial link manipulator
Fig. 15. Offline compliance error compensation approach [66]. and a 11-DoF parallel kinematic machine, respectively, were
analyzed using various redundancy resolution criteria.
Mousavi et al. [76] studied the effect of one and two degrees of
functional redundancy on the dynamic stability limit (for chatter
vibration) in robotic milling. Fig. 18 shows an example of the effect
of two functional degrees of redundancy arising from the presence
of a rotary table and the rotation of the tool about its axis on the
stability limit defined by the depth of cut.

Fig. 18. Effect of 2-DoF functional redundancy on the chatter stability limit in robotic
milling with an ABB IRB6660 industrial robot [76].
Fig. 16. Improvement in circularity during robotic milling of aluminum through
offline adjustment of tool path by compensating for joint stiffness and joint In addition to model-based methods to compensate for robot
hysteresis induced errors [70].
compliance induced machining errors, work on model-based
methods to suppress the onset of dynamic instability in robotic
thereby enhancing the machining accuracy. Practical examples of machining operations has also been reported. While chatter
redundancy in robotic systems include a robot mounted on a linear vibration in machining with CNC machine tools is typically
track, mobile platform, or overhead gantry robot. Fig. 17 shows an characterized by higher frequencies [64] and is usually due to
example of a milling robot on a mobile platform designed for high chip regeneration, the low and comparable stiffness values in the
accuracy machining of large aircraft structures. Such systems are principal stiffness directions of the robot [77] give rise to mode-
being actively researched for industrial applications, particularly in coupling chatter, which typically occurs at much lower frequencies
the aerospace industry. (~10–20 Hz) [78].
It is evident from the literature review on offline methods for
robotic machining process control that much work has focused on
developing model-based methods designed to minimize robot-
compliance-induced machining errors and/or to suppress dynamic
instabilities through modification of the tool path, optimization of
the robot configuration (with or without kinematic redundancy),
and modification of cutting parameters. A rigorous analysis and
comparison of the different offline compensation and chatter
suppression strategies should be developed based on this.

2.8.2. Tool path accuracy

Tool path compensation in robotic milling has been an active
research area in the last decade. In literature, the major sources of
robot compliances have been identified as the gears and bearings
at the link joints [19,79,80]. Offline tool path correction is the
simplest approach, especially for high volume parts, where
Fig. 17. Mobile milling robot: machining of large aerospace structures [72]. repetitive part measurements by coordinate measurement
machines are involved [81]. However, this method may not have
Andres et al. [73] utilized a redundancy resolution method economical merits, especially for large-scale parts targeted by
based on minimization of a position-dependent scalar perfor- robotic milling. Considering this, methods based on robot stiffness
mance index that seeks to keep the robot pose away from ill- modelling have been the focus in the literature.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 9

Belchior et al. [82] applied robotic offline tool path correction

on progressive sheet metal forming process. Barnfather et al.
[83] proposed to use photogrammetry assisted robotic machin-
ing to compensate inaccuracies in robotic milling of two stage
processes. The closest point to the nominal cutting frame on
aligned inspection surface is used as a base for compensation.
Then, the surface dislocation errors generated in the previous
pass are compensated during the finishing pass. It was
demonstrated that accuracy can be increased by using photo-
grammetry assisted compensation. In milling of circular
pockets, the diameter error was decreased from 500 mm down
to 200 mm, whereas cylindricity error was improved from
1000 down to 250 mm.
Zäh and Rösch [84] proposed fuzzy logic-based controllers to
Fig. 19. Demonstration of asymmetrical and symmetrical tool tip FRF in robotic
improve the tool path accuracy in robotic milling by compensating milling and CNC milling, respectively [86].
the static deflection of the robot during robotic milling process. In
some applications where increased stiffness is required, 6-axis the tool diameter, the cross-transfer functions may significantly
parallel kinematic robots are used [85,86]. A drawback of such affect stability diagrams.
solutions is the need for extensive experimental calibration. In a recent study, Maurotto and Tunc [90] investigated the
In an early study on real time correction of the robot position in effects of low frequency chatter introduced by the robot structure
robotic milling, Zhang et al. [87] investigated major topics and on surface integrity in milling of Inconel 690 by comparing the
approaches. They proposed real time compensation of robot surface residual stress and surface roughness in different chatter
deflections and improved the process performance in terms of conditions. They showed that the low frequency chatter governed
accuracy. The results in their study were revolutionary and showed by robot modes increases the magnitude of the residual stress by
that robots can exhibit high performance milling through real time about 40% compared to cases where dynamics is governed by the
path compensation techniques. Recently, laser tracking systems, tool mode. The standard deviation of surface residual stress also
contrary to their high capital investment, have started to be used in increases.
tool path correction for robotic milling applications.
In a later study, Shi et al. [88] used a 3-axis laser tracker for real 2.8.3. Online methods
time tool path correction. They focused on linear and circular type This section reviews online process control strategies employed
tool path geometries, where major improvement was achieved in to improve the accuracy of robotic machining operations. The basic
tool path accuracy from 2 mm levels to the order of 50 mm. approach in online strategies is to utilize sensor feedback to
The second major obstacle in the integration of industrial robots directly measure the robot end effector position or indirectly
in milling processes is the low frequency modes introduced by estimate the machining error and compensate for it in real-time
their flexible structures. High amplitude vibrations that occur at using an appropriate control algorithm. Early efforts aimed at
low frequencies subsequently result in high fluctuations in the implementing sensor feedback-based control of industrial robots
cutting forces. As a result, part surface, tool body, spindle and robot in high force machining tasks (e.g., force controlled grinding or
axis can be damaged. deburring) revealed significant limitations in the bandwidth of the
Addressing dynamics and stability of robotic milling, Pan et al. robot motion control loop. To address this problem, researchers at
[18] were the first to demonstrate the deep marks left on the part Lund University in Sweden developed hardware and software
surface by the robot vibrations in robotic milling. Zaghbani et al. extensions to an ABB S4CPlus control system to enable the
[89] applied a variable spindle speed strategy for vibration control integration of external sensors and control algorithms based on
and minimization in robotic milling, where they used the cutting real-time sensor data [91]. Fig. 20 shows the extended control
force and vibration data as a measure of stability in their statistical architecture, which enables external sensor-based feedback
approach. It is shown that spindle speed variation can be used as a control at the task, joint, cartesian, and motor control current
chatter mitigation strategy in robotic milling. levels. Using this architecture, modification of the servo arm-
Tunc and Shaw investigated the effect of the flexible robot control commands at sampling intervals of 4 ms were realized.
structure on dynamics [85] and stability [86] of robotic milling
while machining AL7075 and AISI316L type materials. In their first
study, a Stewart type hexapod platform was utilized for robotic
milling. The dynamic response at the tool tip and the hexapod
platform were measured using impact hammer tests. It was
demonstrated that the robotic platform can introduce low
frequency modes which may be as flexible as the cutting tool
modes. The position and direction dependent dynamics intro-
duced by the robotic platform have been clearly shown, as well.
With regards to the effects of robotic platform on milling stability,
they demonstrated three fundamental aspects, i.e. position
dependent stability, effects of cross transfer function on stability
diagrams and the role of feed rate direction on the stability lobes,
and absolute stability limits due to the asymmetrical dynamic
response at the tool tip, which means that the measured frequency Fig. 20. External sensor-based feedback control architecture [91].
response function (FRF) along feed (x) and cross feed (y) directions
are different from each other in terms of both amplitude and Bottero et al. [92] analyzed the compensation for periodic
natural frequencies of the dominant modes (Fig. 19). They also torque disturbances, characteristic of electrical permanent magnet
demonstrated the effect of tooling to minimize position depen- motors. The compensation was obtained as a two-stage adaptive
dency of stability diagrams. The cases were handled in two main controller. The final compensator was then obtained via a pure
groups, where the stability was governed by the robot structure predictive feed forward compensation with an internal model of
and the tool modes. Simulations and experiments were used to the disturbance. The stability and robustness properties of the
show that, for the cases where radial depth of cut is less than 50% of overall adaptive compensation technique could then be experi-

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

10 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

mentally validated on industrial manipulators. Today, industrial measured at 440 Hz using a metrology system. Fig. 23 shows the
robot controllers come with external sensor interfacing features system architecture.
(e.g., KUKA RSI [93]) that enable real-time communication
between robot controller and external sensors.
Olsson et al. [94] developed a control architecture to implement
force-controlled drilling of flexible aircraft skin material on an ABB
IRB 2400 IR. A 6-axis force/torque sensor was used to measure the
x–y–z forces and torques produced in drilling. A dynamic model of
the robot response to external forces and moments was used to
design a feedback/feedforward control algorithm that adjusts the
clamp-up force of drilled sheets in real-time. Evidence of reduced
x–y tool deflections was shown, but no direct improvement in hole
quality was found. Pan and Zhang [95] implemented a hybrid Fig. 23. Overall system architecture of the combined macro and micro manipulator
position and force control scheme in an ABB IRC5 robot controller actuated system [97].
to improve robotic milling accuracy. An ATI 6-axis force/torque
sensor was used to measure the milling forces in real-time, after Subsequently, Fig. 24 shows measured profile errors obtained in
compensating for the spindle and tool. Fig. 21 shows the force machining a circular slot similar to Ref. [96]. The authors indicate
control loop utilized in their work. Corrections were made to the that machining accuracy of 100 mm can be obtained with the
robot’s nominal position and velocity via the trajectory generator. combined macro and micro-manipulator system in milling of steel.
The authors implemented real-time robot deformation compen-
sation and reduced workpiece surface error.

Fig. 21. Force control loop implemented in an ABB IRC5 controller [95].

Schneider et al. [96] used a series of LEDs and three cameras

Fig. 24. Deviation of the machined outer circle, measured using a coordinate
(Nikon K600 metrology system) to track robot end effector and
measurement machine. Setup A: only position-control of the robot. Setup C: micro
compensate in real-time the error in position by modifying the and macro manipulation including optical tracking, adapted from [97].
joint servo commands based on the measured position error. Using
a PID control algorithm, end effector position control of a KUKA
KR125 IR was achieved at 500 Hz, acceptable for IRs. Diaz Posada et al. [98] investigated the improvement in robot
Fig. 22 shows the machining error obtained with the position positioning accuracy in robotic drilling using a 3D laser tracker to
error control strategy when milling a circular slot in a steel compensate for the end effector position and orientation errors,
workpiece. Profile accuracy improvement of over 46% was reported and a robot compliance model-based compensation implemented
by the authors of the studies presented in Ref. [96]. in the external controller interfaced with the robot controller. Their
results show the laser tracker-based compensation is able to
reduce the translation errors to less than 0.1 mm and keep rotation
errors to a maximum of 0.2 in robotic drilling of metal sheets.
Work on real time pose control of an industrial robot using 3-
DoF and 6-DoF laser tracking has also been reported by Moeller
et al. [99]. Fig. 25 shows the positioning path accuracy (ATp) and
the repeatability error (RTp) obtained using external closed loop
control based on the laser tracker data. It can be seen that there is
significant reduction in the ATp parameter with 3- and 6-DoF
control but the improvement in repeatability error is not
significant. However, the external control approach outlined by
Fig. 22. Machining error using a tuned PI control algorithm based on real-time end Moeller et al. is shown to produce considerable reduction in the
effector position sensing [96].

Later work by Schneider et al. [97] extended their real-time

position error compensation approach to include a combination of
macro and micro manipulation. A KUKA KR125 industrial robot
was the macro-manipulator used to hold the workpiece, while a 3-
axis translational piezo-actuated system served as the high
bandwidth micro-manipulator on which the milling spindle was
mounted. A parallel-link flexure mechanism enhanced the
dynamic properties of the piezo-actuated micro-manipulator
and to extend the range of compensation in each axis to 0.5 Fig. 25. Path positioning accuracy and robot repeatability error with and without
mm. The robot position and orientation in the workspace was laser tracker-based control [99].

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 11

robot deflection induced error produced in machining polyure- Due to the open kinematics of serial-link-based robots, unlike
thane material. compact machine tools, expansion errors continue to accumulate
Cen and Melkote [100] recently presented a wireless force from the base upwards. The problem is worsened by the uneven
sensing based approach to compensate for robot positioning errors temperature distribution and by the different materials of the
in robotic milling, summarized in Fig. 26. A low-cost polyvinyli- robot arm, which lead to tensions throughout the components
dene fluoride thin-film force sensor developed in prior work by the [102–106]. The resulting displacement contributes significantly to
same group was used to measure the milling forces in real-time. the position accuracy errors [103]. There are various approaches to
The measured forces are fed to a well-established mechanistic compensate for the effects of temperature-related deviations. The
force model for end milling to compute the instantaneous radial procedures generally follow the scheme in Fig. 28.
and axial depths of cut, which gives the actual position of the tool
relative to the workpiece. The error in tool position is then
compensated in real-time by commanding the robot to the
corrected position.

Fig. 28. General procedure for the drift calibration of robots [acc. to 106].

Fig. 26. Milling force model and wireless force sensor-based feedback First, the drift deviations are determined from measurements
compensation method [100]. on the robot. This can be done in two different ways: directly and
indirectly. Based on this data, certain parameters are adapted. A
The approach was implemented on a KUKA KR210 industrial distinction can be made between changing the target points in the
robot and peripheral end milling tests were conducted at different robot program and internal adjustments to the path planning
feed rates. Fig. 27 shows the surface error obtained without sensor component of the control. The first option has the advantage of
feedback-based compensation and with compensation. It can be being robot-independent and does not require deeper access to its
seen that with sensor feedback-based control, the surface error is controller, but the temperature compensation must be considered
reduced by over 70%. individually in each program. The parameters of the transforma-
tion chain for calculating the inverse kinematics can be varied by
optimization algorithms that can process identified position errors
in such a way that the remaining errors are minimal [103].
The methods used for the determination of position errors can
be divided into two categories: direct and indirect methods. Direct
methods are based on the actual measurement of the position
deviations of the robot arm, whereas indirect methods estimate
the position errors from the measurement of quantities propor-
tional to the displacement. Examples of such quantities are the
change in length of individual components [104] or simply the
temperature [102]. With the help of thermos-elastic deformation
models, it is possible to calculate the displacement of the entire
robot arm without additional hardware. In principle, these indirect
Fig. 27. Surface error reduction in a straight peripheral end milling cut using force
methods are not as exact as the results of direct measurements, but
sensor-based feedback [100].
the considerable cost advantage nevertheless makes them
Figs. 22, 24 and 27 provide quantitative examples of the interesting for some applications [102].
reduction in machining error using online compensation strategies The position measurement for one-time adjustment and
reported in Refs. [96], [97] and [100], respectively. It is evident, calibration (after manufacturing) of industrial robots is usually
though, from the literature review on online process control carried out as part of a 6D calibration. In this calibration, spatially
strategies that there is room for further improvement in robotic fixed measuring devices record the position and orientation of a
machining accuracy through sensor feedback and control. This is measuring object attached to the robot, resulting in high fidelity
an area of ongoing research in the community. data from some robot poses [107,108]. However, this method is
only of limited applicability for the continuous calibration of
2.9. Robot temperature compensation robots. The reasons are: (1) the high cost of the measuring systems
and of their installation location, and (2) a reduction of robot
While the repeatability of standard industrial robots is flexibility by mounting sensors close to flanges.
satisfactory for many applications, their absolute accuracy is not A 3D calibration is more practical in this regard. In this method,
sufficient for machining tasks or robot-based measuring systems. the measuring device is mounted on the robot and the measuring
Manufacturing inaccuracies in robot production can often be body is stationary. This method is simpler to integrate, especially in
permanently compensated by a one-time calibration, while drift the context of robot-based measurement systems, where the
effects, for example, due to wear or temperature, must be robots are already equipped with measuring sensors. However, in
periodically compensated. In particular, the heat input of the contrast to the devices required for 6D calibration, the more
drives combined with temperature fluctuations in the environ- compact sensors attached to the robot can usually only sense a
ment lead to relevant volumetric expansions in the structural position without orientation [109]. For this reason, a 3D calibration
components [101]. requires more robot poses to achieve similar accuracy [106].

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

12 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

In contrast to absolute calibration, drift compensation or the machining task itself, the RC has to be complemented with
relative calibration only uses a limited set of recent measurement functions of a CNC, like in machine tools. Thus, the RC and its
values obtained at regular intervals. The robot approaches various advantages can be used for pre- and post-operations but the path
targets on the measuring body and the sensors determine the planning and trajectory generation for the machining part is done
position errors. On this basis, certain parameters particularly by the CNC. This is necessary because the RC is parametrized for the
susceptible to temperature are optimized to minimize the position particular robot that it is commanding and hence, always machine-
errors [101,103,105,106,109]. specific. CNCs, however, are used to ensure the proper machining
In a full calibration, all available targets are scanned and the of the workpiece [112]. The combination of RC and NC enables G-
compensation values calculated. This method is used after code programming of the machining process. Exploiting the
prolonged plant stops to compensate for substantial temperature flexibility of robots in machining G-code programming can be very
changes. To avoid long pauses during operation, usually only a few complex and challenging. For this reason, the use of CAM-software
targets are measured between the primary robot operations. These is required to ensure simplified and collision free path generation.
are subsequently used for calibration together with previously
recorded measuring points, which is why the method is called split 3.2.1. CAD-CAM toolchain
calibration [106]. In the primary application area of this The machining of components via CAM path planning usually
technology, the robot-based inline measurement of components starts by loading a CAD file into the editor of the CAM program.
for quality assurance, the measurement results of the robot cells Various standard CAD data formats, such as STEP or IGS, can be
are compared with measurements from coordinate measurement read. After path planning (Fig. 30, Step2) the program can be
machines in temperature-controlled environments at regular exported in standardized or machine specific output formats.
intervals. By adding offsets, the discrepancies can be further
reduced.

3. Programming in robotic machining

Until recently, industrial robots were used in machining

exclusively for loading and unloading, that is, for supporting
machine tools in machining processes. But with the possibility of
using them for machining tasks with low process forces, like
chamfering or deburring [110], new requirements emerged for
Fig. 30. Process chain CAD-CAM.
path planning and thus for the programming of industrial robots
for machining operations. Some of the particularities of this field
are presented in this section, along with practice examples. As shown in Fig. 30, the output format can be written into the
respective programming language of the robot manufacturer or
3.1. Sensor based programming methods (deburring, polishing) can be an output in G-code adapted for robot applications. The
customized G-code may allow the use of a larger look-ahead, an
The application of a stereo camera system for external TCP-pose extension of the program parser’s look-ahead. Under these
measurement increases the absolute positioning accuracy of an conditions more precise machining results can be achieved. Other
industrial milling robot by up to 0.1 mm [111]. For higher absolute possibilities of increasing the machining accuracy with offline
accuracy up to 0.03 mm and error correction along the tool path programming using CAM tools are shown in Refs. [113–116].
during feed motion, dynamically measuring laser tracker systems In the project PROGEN [117], demonstrators of pressing tools
are a suitable solution. External guidance (Fig. 29) makes the were machined by robots. Processing was carried out on the same
machining robot unsusceptible to external or constant process geometry with the same material. Path planning for the milling
forces, thermal expansion and calibration errors [99]. strategies was carried out using the same CAM program in G-code
compilation, but varied in 3- or 5-axis programming. The results of
the milling processes are shown in Fig. 31. The surface color is,
analog to the scale, the deviation of the surface geometry in [mm]
compared to the CAD design. The evaluation shown by a
structured-light 3D scanner reveals that 3-axis machining has
fewer geometry deviations over the entire component. This is
mainly due to the frequent reorientation of the process head in 5-
axis machining and the associated axis movement of the robot. The
processing quality can be influenced by targeted robot poses [118]
and skilled path planning [87]. In addition to the selection of
optimal robot positions, the clamping position, the position of the
tool path and the process-related technology parameters also
determine the process result. Industrial robots with serial

Fig. 29. External TCP-guidance and accuracy validation by two dynamic laser
tracker systems [99].

3.2. Programming environment

Pre- and post-operations in machining, such as the determina-

tion of the workpiece’s position and orientation as well as
measurement of the geometry after machining, can be performed
by robots using additional vision- and laser-based technologies. Fig. 31. Results of robot-based milling in 3-axis (left) and 5-axis programming
These operations can be executed using the robot control (RC). For (right) [117].

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 13

kinematics have a different static and dynamic behavior, depend-

ing on the process parameters [27].
Current CAM planning systems are unable to simulate static and
dynamic properties of robot systems during process design. One
approach was pursued in the publicly funded research project
HORuS (at WZL RWTH Aachen) with a focus on integrating process
behavior into process planning. The follow-up project HORuS2 will
further develop these findings. The main focus is on the
development of planning assistance modules to increase the
accuracy of robot systems for machining large components.
One exemplary module focuses on the reduction of inverse
movements of the rotary robot axes during tool engagement.
Studies have shown that inverse movements of individual axes
lead to high path deviations [119]. Foresight or simulation of
rotational axis movements during planning is currently only used
for accessibility and collision checks, but not for analysis of critical Fig. 33. Combination of RC and CNC functionalities in KUKA.CNC.
inversion movements. Fig. 32 shows an approach for calculating
rotational paths based on defined tool paths. It is possible to
visualize critical areas and to adjust process parameters or programming and minimize the need for retraining employees,
machining strategies manually or automatically. leading to less reluctance using robots in the machining industry.
Practical examples prove that IRs can substitute machine tools,
given their programming takes machining conditions into account.
In an example from BAE Systems, two robots are used simulta-
neously for countersinking of predrilled holes on aircraft compo-
nent production [120]. One robot performs machining, while the
other assists the process by supporting (Fig. 34).

Fig. 32. Calculation of critical turning points of the rotational robot axes for reducing
path deviations.

3.2.2. Dedicated programming possibilities

The robot control can be complemented by CNC functions to
Fig. 34. Countersinking cell from BAE systems.
make sure that the robot can follow the desired and programmed
path given as robot-specific languages and generated by CAM-
software. This requires a particular architecture, discussed here by For machining very large workpieces, the working space of the
means of the KUKA.CNC. robots may not be enough. For these cases, robots are moved
KUKA is so far one of only a few manufacturers providing an around a part via automated guided vehicles (AGVs). Fig. 35(a)
integrated NC kernel. There are other examples for that such as the shows an application of milling a carbon fiber reinforced
Siemens NX Robotic Milling Module, the announced COMAU Robot thermoplastic structure from the aerospace sector [121], whereas
CNC control, or a similar solution from Stäubli. Another possibility, Fig. 35(b) presents polishing of a mold for a sailing boat [122].
also from KUKA, is to combine CNC and RC. This is achieved with a
Siemens 840D controller, the KUKA.CNC Sinumerik. This is a KUKA-
specific development, to ensure accurate and improved path-
behavior for machining tasks, because in this case, interventions
on a low level of the RC are required. If such options are not
available, the CAM software has to compile the programs into robot
language, as described in Section 3.2.1. Thereby, the achievable
machining quality is strongly dependent on the post-processors
implemented in the CAM software. This is why a combination of RC
Fig. 35. Mobile robots on AGVs: (a) aerospace [121]; (b) naval sector [122].
and CNC should be favored, since offline post-processors do not
have a feedback of the actual trajectory, whereas the RC does.
Fig. 33 illustrates how RC (orange boxes) and CNC (blue boxes) Fig. 36 shows how subscale machining on large aerospace parts
interact within the KUKA.CNC. is made possible by using an accuracy improved industrial robot
The main tasks of a CNC are the calculation of numerous mounted on an AGV [123]. Replacing machining centers for those
support points and the assurance of suitable path-dynamics, for kinds of applications requires highly accurate measurement
example, constant velocity of the TCP, which in turn equates as the systems and referencing strategies [72].
feed rate of machine tools. In order to provide these necessary Although studies such as Klimchik et al. [62] showed that
functions, the CNC requires high computational power, since the compensation algorithms can be applied to improve process
path dynamics are dependent on the total movement and thus the quality, these were only validated in a laboratory environment and
high number of support points have to be considered by an they have not yet been implemented in the industry. Hence,
enhanced look-ahead functionality for the actual path-generation. technology readiness levels of these solutions need to be more
Therefore, such calculations have to be performed in real-time. advanced for wider adoption of robotic machining in the high value
In conclusion, the RC is used to take advantage of robot-specific manufacturing industry. Until then, industrial processes involving
functions such as singularity avoidance or specific transformation low cutting forces or low dimensional and surface quality
algorithms. The RC-CNC combinations support the use of G-code requirements will remain suitable for industrial robots.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

14 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

 Burr positions and dimensions are difficult to predict which

makes the process highly variable.
 Additional sensors require specific parameterization, making the
set-up of robot cells complex, especially given workpiece and
process variations.
 The low stiffness of robot systems in comparison with CNC
machines can cause difficulties for achieving high quality
deburring, especially for iron and steel parts.
 The deburring of complex and large workpieces is difficult to
achieve with robots due to the lack of absolute position accuracy.

Due to these challenges, manufacturing programming efforts

and process quality requirements should be carefully investigated
before applying robots in material removal. Robots still offer
Fig. 36. Positioning accuracy in large workspaces [123].
advantages: the machined part can be grasped and guided relative
to the spindle, so that a fully automated work cycle is possible.
4. Machining processes with industrial robots Furthermore, large workspaces and high tool dexterity could be
made available at a low cost.
In this section, we are offering a detailed overview of known
machining processes that benefit from the deployment of 4.1.1. Offline compensation of workpiece shape deviations
industrial robots. We are also discussing methods of enabling A fundamental problem affecting the process quality of
robots for a wider range of machining applications based on deburring processes is that geometric shape deviations between
current studies and research results. the nominal CAD model and the manufactured parts may be
beyond acceptable tolerances. To overcome this challenge, Kuss
4.1. Deburring et al. [133] proposed an approach in which dimensional tolerance
specifications of the manufactured part retrieved from the product
Deburring is a common manufacturing process which demands design are then used to derive possible variations of the workpiece
characteristics also common to robotic systems (e.g., flexibility or geometry model. The iterative closest point (ICP) algorithm is
re-configurability, among others). Robots are still not widely used applied for identifying matching inaccuracies between CAD-
for performing this manufacturing task. Several advances have generated and sensor-derived point clouds. These identified and
been made for improving the use of robots in deburring processes. modelled inaccuracies are later compensated in an OLP system to
The offline programming (OLP) allows simulating the process in achieve better accuracies. Fig. 37 shows the experimental setup for
virtual environments and automatically generating programs the detection of workpiece shape deviations within a robotic
based on geometric approaches. For example, the use of OLP in deburring process. Deburring with shape deviations demonstrated
conjunction with a touch probe-based strategy for robotic a 57% decrease in the roundness standard deviation error,
deburring of aerospace components [124]. Also, an offline compared to deburring without shape deviation recognition.
simulation of finishing edges has been reported as beneficial for
the automation of this process [125].
The problem of accurately localizing the workpiece is solved in
industry mainly by using in-process measurement sensors that
localize the workpiece for a subsequent accurate generation of the
deburring path. For instance, the use of a vision assisted robotic
system using 2D-images for automatic programming has been
proposed in Ref. [126]. In industrial cases, workpiece inaccuracies
are compensated under the allowed tolerance mainly due to the
Fig. 37. Detecting workpiece shape deviations with a 3D sensor for robotic
introduction of compliant tools that compensate surface irregu- deburring processes (left). Test workpiece after the deburring process with
larities while maintaining a defined force. Other compensation detection of shape deviations (right) [133]. Key for the left image: 1. Industrial robot
tools such as integrated pneumatic actuation systems have been KUKA KR270; 2. Spindle Pferd PWS; 3. Milling tool; 4. Workpiece; 5. 3D sensor
proposed for tackling this limitation [127]. Ensenso N10.
Numerous control laws and strategies have been proposed in
order to compensate robot, process, and workpiece errors. The 4.1.2. Automatic OLP for sensor-based robotic deburring
control methods can be categorized in two major approaches: In order to approach the complexity of programming robotic
impedance control and hybrid position/force control [128]. Exam- deburring systems, an automatic OLP for sensor-based robotic
ples of control applications are, for instance: the robot path deburring has been proposed, facilitating the industrial imple-
generation method using hybrid force and visual servoing for mentation of these systems. Based on the Product, Process, and
reducing programming time [129]; tool-path modification based Robot (PPR) model, paths are generated which considers the
on direct teaching over the workpiece with a force control in constraints and DoF’s of the robotic deburring process.
normal/tangential direction [130]; the control strategy of mimick- A laser scanner sensor is used for workpiece localization in the
ing human behavior during manual deburring [131]. robot cell (Fig. 38). The ICP algorithm is later also used for
Despite the previously introduced state of the art, remaining compensating errors between the TCP and the edge to be deburred.
challenges have to be overcome in order to reduce the negative
effects of the process and to facilitate the implementation of
robotic deburring systems. These challenges are listed below:

 Robot, process, and product dependent tolerances make it

difficult to reach the industry demanded accuracies.
 Robot cell cost efficiency may be limited due to complex settings
and long programming times, especially given changeovers.
 Programming for robot deburring usually demands experts with Fig. 38. Sensor-based robotic deburring (left), pre-programmed in an OLP
further expertise in the process itself [132]. environment (right) [134].

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 15

Output of this approach is an automatically generated program in During the execution of the forming process by the two
which the expected measurements are also encoded and sent to cooperating industrial robots, the forces at the forming tool reach
the robot controller for later compensation [134]. Furthermore, the around 2000 N, using a typical deep-drawing steel with a sheet
OLP system has been developed to embed sensor models for thickness of 1.0 mm. Those forces lead to a significant displace-
measurement optimization, considering sensor specifications and ment of the forming tool’s TCP and therefore a poor geometric
using sample-based planners and the intrinsic process DoF. Sensor accuracy of the produced part. Laurischkat [146] observed a
and tool calibration algorithms have been proposed to provide displacement of the TCP of over 3 mm during the forming process
satisfactory accuracies [134]. using two KUKA KR360-1 (Fig. 41).

4.2. Incremental forming

In sheet metal forming, conventional processes are unable to

meet the growing demand for shortening product development
cycles due to their need for part dependent tools. Their time and cost
intensive manufacturing prevent a reasonable usage in prototyping
and small batch production [135,136]. The incremental sheet metal
forming (ISF) process is able to overcome these challenges due to its
high geometrical form flexibility and part independent tools, making
it applicable for a wide range of industries [137–140].
In ISF, the final shape is produced by the incremental infeed of a
hemispherical forming tool in the depth direction, and by its
movement along the contour in the lateral direction on each level
or on a helical path. This process called Roboforming (Fig. 39) can
be executed by industrial robots, taking advantage of their high Fig. 41. Forming tool direct displacement, based on forming force [146].
flexibility and distribution while losing the stiffness of the often-
used modified CNC-machines [141]. A second supporting tool can To ensure a precise forming process, the displacement of the
be mounted on a second robot. It moves directly opposed to the forming tool needs to be compensated. In literature there are
forming tool and generates a predefined gap and force between the various approaches to predict the resulting forming forces before
two hemispherical tools. This locally substitutes a die to improve the forming process [147,148]. However, those approaches lack the
the accuracy of the formed part [142]. Additional geometric needed accuracy to enable a simulation-based compensation,
elements need to be engineered in preparation of the Roboforming especially for complex parts [82]. Therefore, an online-compensa-
process chain after the target geometry CAD design of the sheet tion of the robot position, based on the measured forming forces, is
metal part is developed. These elements enable CAM-based path inevitable.
planning and the forming of the target geometry from a planar Abele et al. developed a flexible joint multibody dynamics
sheet metal. system model for an industrial robot to describe its behavior based
on the beforehand measured joint stiffnesses [149]. This model
consists of a basic multibody system including the kinematic and
kinetic parameters. The contained standard stiff joints are
extended to flexible joints by the addition of joints in the direction
of motion. Those are coupled to the driven joints by spring and
damper elements. Analogously joints for tilting are added to
describe the tilting behavior of the bearings. This elastic joint
model is implemented into Matlab/Simulink via SimMechanics.
Laurischkat applied this flexible joint multibody system in
Roboforming. The by a force torque sensor measured forming force
is used to calculate the displacement of the TCP during the forming
process. Afterwards, an inverse model is used to adjust the position
Fig. 39. Robot-based incremental sheet forming setup [143]. via the robot sensor interface (RSI). Laurischkat [146] was able to
improve the path accuracy significantly (Table 1) by using this
The applied CAM system determines the positions where the approach. The online stiffness compensation is therefore a
forming tool is in contact with the metal surface. A post processing significant extension for the applicability of the process entailing
step is required for robot program generation. All process an increase in geometric precision and increasing the formable
parameters (e.g., contact force of the supporting tool) are defined sheet thickness making it indispensable in robot-based incremen-
in this step [144]. The forming of the extended geometry is tal sheet forming.
executed by the incremental forming machinery. The sheet metal
is then unclamped, annealed and cut down to the initial target Table 1
geometry (Fig. 40). Positioning accuracy in robot-based ISF [146].

Sheet thickness Online Mean positioning Errors exceeding

[mm] compensation error [mm] [%]

0.3 mm 0.6 mm

1.0 No 0.98 83.6 70.0

Yes 0.04 2.9 0.0
1.5 No 1.35 88.3 73.3
Yes 0.08 13.3 1.0

4.3. Grinding and polishing

Polishing of free form parts like turbine blades, dies and molds
Fig. 40. Toolchain for robot-based incremental sheet forming [145]. constitutes an important percentage of the total production time

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

16 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

to achieve uniform material removal. They obtained 60% improve-

ment in profile error, and the material removal could be predicted
accurately with an error margin lower than 6%.
For path control, Yang et al. [160] applied shape adaptive
motion control with already integrated part measurement. They
proposed a system containing four main blocks: surface measure-
ment, surface reconstruction, tool trajectory planning, and axis
motion control (Fig. 43). They implemented spatial spectral
analysis in the first block, and proposed an optimal digitizing
frequency. They then performed a spatial spectral B-spline method
Fig. 42. Robotic polishing of: aluminum rim (left), car bumper (right) [151]. for surface reconstruction. In the third block, a motion profile was
initially selected, then the tool locations were determined based on
(17–29%) [150]. Various industrial parts require polishing as a last the reconstructed surface for improved accuracy. A software
manufacturing step, as the examples in Fig. 42 show. package was then developed and implemented in robotic polishing
In polishing, the aim is to improve the surface finish rather than processes. They demonstrated benefits of the system for polishing
to alter part geometry or dimensions. Industrial robots are used in the spherical surface of a doorstop.
these operations as an alternative to machine tools. Considering
the compliant tools used in polishing, the required motion
accuracy from the articulated robot is not as high as for milling
processes. As long as the robot can orient the end effector
according to the part surface within the work space, it is very
suitable to perform polishing processes. Literature shows studies
on these processes where the major goal is to improve the
efficiency and surface quality. The research has focused on end
effector application, process analysis and path generation.
The design of end effector systems is of great importance in
robotic polishing operations, especially for keeping a steady
contact and consistent contact forces. Takeuchi et al. [152] used an
air-actuated polishing end effector, where the contact force was
adjusted by an interface of a linear roller slide and an air piston
attached between the robot wrist and the tool. Mizugaki et al. [153]
used loose abrasives in robotic polishing with a compliant end
effector attached to a robot controller communicating with a CAD/
CAM system. The efficiency of the system was demonstrated
experimentally by preventing over-polishing at the edges.
In their work on automated process planning, Saito et al. [154]
developed and implemented an expert system to plan the
polishing of die and mold surfaces. Their system was able to
suggest process plans for minimum cycle times, once performance
requirements like surface quality and initial conditions were
provided. The system was dependent on the knowledge of skilled Fig. 43. Automated robotic polishing system [160].
mold polishers.
Tool path generation is a vital step towards an automated Robotic grinding has also been used in manufacturing large
polishing process. Mizugaki et al. [155] proposed a novel method scale hydro turbines by Alstom and Hydro-Quebec. They improved
that used the planar Peano curve in the xy-plane and its orthogonal turbine efficiency by reducing surface roughness. In a related study
projection onto the free-form surface. They kept the end effector of Sabourin et al. [161], the manual grinding process after the 5-
normal to the part surface considering the robot’s working space. axis milling of hydraulic turbines was robotized to obtain a better
They concluded that this way, the polishing force can be controlled. surface finish with less hydrodynamic friction, at a reduced total
In another study focusing on tool path generation, Tam et al. [156] cost (Fig. 44). They improved grinding process parameters for a
succeeded in even tool path generation by addressing only time desired surface waviness and roughness, which was implemented
efficiency and surface quality issues. They showed that an accuracy on a full-scale Francis turbine blade. It was shown that the surface
of 0.01 mm can be achieved with robotic polishing of curved roughness can be lowered from Ra = 15 mm to Ra = 0.1 mm
surfaces. Marquez et al. [157] offered a complete solution for reaching a grinding rate of 5 h/m2.
robotic polishing of die-mold surfaces. They used the geometrical
part definition obtained from the CAD model. The advantages of
robotic automation for the manual polishing processes were
clearly shown in terms of time and cost.
Tsai et al. [158] developed a robotic path planning system for
the automated robotic polishing of molds. They utilized the IGES
data format to create the geometric data structure. They then
planned the tool path based on the intrinsic properties of the mold
surface and process requirements. They demonstrated uniform
coverage of the mold surface with satisfactory quality. In a study by
Fig. 44. Automated robotic polishing/grinding system for turbines [161]: Faro laser
Tsai et al. [159], a uniform material removal approach to improve
tracker used for proximity (left), polishing process (right).
automatic mold polishing is proposed. They obtained the mold
surface geometry in the IGES format, which was re-generated in a In their work on automated robotic polishing systems, Ryuh
NURBS surface model. They considered the intrinsic properties of et al. [162] integrated a six DoF articulated industrial robot with a
the surface by considering the effective contact area along the pneumatic grinding tool controlled by a PC and a robot controller.
polishing path for uniform material removal. The end effector They proposed a procedure to provide the path data to the robot.
velocity together with the polishing force were controlled in order The data was automatically generated from the NC-data of a

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 17

previous machining process. This was a step forward towards

automatic path planning for robotic polishing processes. They used
an elastic material between the polishing pad and the holder, to
further enhance the performance of the system in terms of the
grinding contact.
In a recent study, Tian et al. [163] investigated solutions to
improve cost and part quality in curved surface polishing. They
established the physical relation between polishing force, robot,
sensor and polishing tool. Then, they modelled the material
removal distribution and combined this model with a suitable path
spacing algorithm. They validated the system performance on
NAK80 steel by showing that constant force control can be
achieved by active and passive compliance force control.
Fig. 46. Static deformation and dynamic instability when milling with robots [38].

4.4. Enabling industrial robots for specific tasks

trajectory accuracy. The lower part of Fig. 46 shows the result of
The low eigenfrequencies and absolute positioning accuracies dynamic instability: a sufficiently strong force excitation leads to a
represent disadvantages of industrial robots in machining. natural vibration of the end effector and therefore to waviness
Measurements on a KUKA KR500-2 robot show that the first formation on the workpiece surface (chatter marks). Since the
eigenfrequency is at 5.5 Hz. Its mode shape is a rotational cutting force increases momentarily whenever the end mill cuts
deformation between the robot base frame and the rotating into one of these chatter marks, the robot structure is excited
column. This low frequency comes from the poor stiffness in the periodically with one of its eigenfrequencies. The vibrations caused
gear box of the first joint [164]. During process planning, Denkena by this self-excitement can become so strong that the machining
et al. consider the errors resulting from the gear boxes, such as process has to be stopped in order to avoid serious damage to the
compliance and backlash [69,165]. They use the results of the machine, the tool, or the workpiece [168]. Despite these qualitative
milling process simulation for a teach-less process monitoring. In shortcomings of milling robots compared to machine tools, their
the case of robot machining, effects like structural deviation and flexible applicability and their lower investment costs motivate a
backlash are predicted based on the tool path and simulated division of work: a milling robot processes a workpiece as much as
cutting forces, and used for offline compensation. The backlash possible, and then a machine tool performs the finishing. The latter
leads to a positioning error of 175–349 mm, which could be can be used more efficiently and more profitably this way
compensated by robot control. This compensation is still limited, [112,169].
because current machining state information like tool wear, Several research studies aim to improve the accuracy, the
clamping and thermal behavior are not regarded. To take process vibration behavior and the achievable quality of milling results of
changes into account, Denkena et al. [164] present an approach for industrial robots. To increase the kinematic stiffness and thereby
online compensation of deviations resulting from process forces. A the static stability, mechanical reinforcements are widespread.
model was investigated for calculating the pose-dependent Machining robots with an increased degree of parallelism are
stiffness for each pose. The machining robot is equipped with a already available on the market (e.g., ABB IRB 6660). Designing the
force sensing spindle holder (Fig. 45) to measure process forces spindle as a sixth axis and removing one axis and its compliance
during the cutting operation and to increase process accuracy with from the flux of forces is another constructional approach to the
an online compensation of the TCP deviation. problem of static compliance-related errors [170]. These measures
can improve the static and dynamic properties of robots, but
reduce the usable workspace and the kinematic flexibility [84].
A common way to deal with instabilities and inaccuracies is the
implementation of mathematical models for the offline simulation
of different robot structure reactions. Kurze [171] proposed a
model-based method to improve path accuracy and disturbance
behavior. Since the model stiffness and damping parameters have
been determined by analyzing free oscillations in only one pose,
the results have been classified as not satisfactory by the author.
Other works carried out experiments in various poses to determine
the static compliances and use the measurement results to derive
analytical models of the robot [118,172].
The common problem faced by model-based approaches is the
dependency of static and dynamic properties on the robot’s joint
Fig. 45. Force sensing spindle holder for a KUKA KR 500-2 milling robot. configuration [27,38,173]. Due to system complexity given by
numerous joints and components, this interdependency is
Milling tasks involve significantly higher load forces than insufficiently described by a mathematical model [174]. Robot
handling and manipulating. This confronts robotic engineering dynamics models are therefore either valid for just a small section
with problems that have to be solved before industrial robots can of the workspace or insufficiently consider nonlinearities [175]. Re-
compete with, or at least complement, conventional machine tools cent studies attempt to identify the pose (end effector cartesian
[112]. Due to their dynamic properties and relatively high position and orientation) in an empirical dependency on process
compliance, serially linked robot arms are more susceptible to stability, with a high number of measurements conducted in
process forces than comparable machine tools [166]. This can lead different machining poses throughout the workspace. For this
to deflections from the commanded trajectory, and to strong purpose, research groups carried out milling experiments in nine
natural oscillations. Fig. 46 illustrates the consequence of these different poses in the xz-plane (y = 0 mm) of the workspace with
two phenomena for a milling application. feed in positive and negative x and y-direction, as seen in Fig. 47.
The bend of the milled path in the upper part of Fig. 46 indicates Milling wedges with linearly increasing cutting depth gives
that the end mill has been statically pushed away from its information about the correlation between depth of cut and
commanded trajectory. This deviation can be up to several processing stability. Oscillations of the robot structure are recorded
millimeters, which is why Puzik et al. [167] refer to it as low with a force-sensing plate underneath the workpiece, and

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

18 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

Fig. 47. Milling pattern (left) and machining poses [27].

acceleration sensors on the robot arm. The measurements lead to

the following conclusions: (1) The robot has a strong dependency
on the static and dynamic behavior of the machining pose and the
feed direction. Further experimental investigations in this field are
necessary. (2) An exclusive consideration of robot compliance Fig. 49. Improvement in surface quality Ra with the support.
values, as it has been assumed in many previous approaches, will
be unsuccessful: for some machining poses, milling quality and
process stability are higher, even though the robot structure is
especially compliant for these joint configurations. Such
approaches require a high amount of measurement and workload,
but they yield promising results.

5. Trends in robotic machining

Having set the scope for deploying industrial robots in

machining processes, we are now focusing in this chapter on
other fields of applications for robots in machining.

5.1. Thin wall machining

Fig. 50. Robotic hand support on a thin wall part.
Robotic assisted machining is particularly beneficial for thin
wall parts, where problems such as high form errors and chatter These influences cannot be compensated with robot control
are very common. Ozturk et al. [176] demonstrated the use of because the ability to control the joint dynamics is limited by the
mobile rubber roller support that follows the feed motion of the cycle time of the robot control. One solution is the optimization of
machine tool. The rubber roller on the robot applies both support the whole robot kinematic chain and also of the drive concept.
force and provides additional damping to the part. Hence, the Denkena et al. [177] proposed an analytical comparison of three
maximum form error was decreased from 90 to 29 mm in a different machine kinematics (Fig. 51). The main objective was the
representative case (Fig. 48). development of a new cost-efficient machine tool for milling large
aluminum and carbon fiber reinforced thermoplastic (CFRP) parts.

Fig. 51. Concepts and evaluation of machine tool kinematics [177].

Fig. 48. Improvement in form error with the support.

Moreover, chatter problems experienced in the center of the

part were eliminated, resulting in an improved surface quality The first concept is based on a serial kinematic. As the sixth joint
(Fig. 49). is not required for five-sided machining, it is eliminated. For the
Support can be provided by different end effectors, such as second concept, a translational axis is used instead of the first
rubber rollers and metal castors, which follow the feed motion of rotational axis. This drives the whole robot kinematic chain,
the machine tool. A dexterous hand can also be used as an increasing the work space and decreasing the required length of
additional support in order to increase the dynamic stiffness of the the links. In the third concept, the linear axis actuates the
part (Fig. 50). workpiece table instead of the robotic arm kinematic chain. A
As both machine tool and robot are in motion during the rotational axis of the robotic kinematic chain is replaced by a rotary
process, there is a risk of collision between them. For this reason, table. The replacement of the first rotational axis increases the
the integration of the machine tool control and robot control needs stiffness by up to 35%, while the replacement of the guiding shoes
to be established to achieve synchronous motion. on the next larger model has an insignificant influence. The third
concept is the best solution for higher productivity, but is also more
5.2. Dedicated kinematics expensive. The second concept is useful when an inexpensive
machine with optimized stiffness is required. Fig. 52 shows an
Machine tool productivity of industrial robots is still limited by analysis of the influence of machine components on the stiffness at
the dynamic properties, as they influence the process stability. the robot TCP.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 19

Fig. 54. Tolerances for machined Ni-alloys parts with industrial robots. (© PTW TU
Darmstadt).
Fig. 52. Impact of the machine tool parts on the stiffness at TCP.

Due to the long-arm kinematic chain of industrial robots, joint net shape components for the finishing process (Figs. 53 and 54)
stiffness has a big influence on the TCP position in machining [180].
applications. Measurements on a KUKA KR500-2 show that the Other process combinations include drilling, reaming, assembly
stiffness at the TCP varies in a range of 0.5–1 N/mm depending on and quality assurance steps in one robot cell [182]. Supported by
pose and force direction. The low joint stiffness is the reason for laser line scanning and qualified CAM tools, these processes are
low eigenfrequencies in industrial robots. These vibrations reduce applied in one robot cell, thus enabling a hybrid manufacturing
the surface quality in milling operations. process in one clamping.
Denkena et al. developed a piezo-based, actuated robot arm to Skilled path planning in CAM-tools increases the accuracy – in
compensate for high dynamic vibrations of the TCP of an industrial terms of geometry and surface quality – due to lower tool
robot [178]. Highly flexible piezo-based actuator foils and sensors deflections. Furthermore, the process combination and integration
were integrated into the robot structure. As CFRP is predestined for in one cell offers high resource savings. The hybrid production
the integration of sensors, the robot arm was replaced by a CFRP concept combined with robot compliance- and temperature
component. The end effector can be positioned with sub- compensation enables a high productivity and accuracy process
micrometer precision, enabling dynamic compensation of posi- for individual components made of difficult-to-machine alloys.
tioning errors. It can also be used to actively damp the structural The large working space of robots can be used for the machining
vibration and to increase dynamic stiffness. of structural lightweight parts — starting with medium sized parts
A further approach deals with the optimization of the drives. like spring-strut domes for the automotive industry [103] up to
Industrial robot drives use gearboxes to achieve the high drive very large components for aerospace or the shipbuilding industry.
torques required to actuate the large mass of the serial structure. The work space can be increased by adding a mobile self-driven
Denkena et al. [179] developed a new hybrid drive concept, which platform carrying the machining robot [72]. Another field of
combines the advantage of a torque motor with the high torque of a application is the machining of large carbon composite and
servomotor with gearbox. In order to show the advantage of the ceramic parts, due to the advantage of protecting the revolute
new drive concept, it was tested in a two axes robotic kinematic joints of robots by excess pressure against superfine particles
chain. Each joint was equipped with a torque motor with a load- compared to sealed protections and bulky covers for linear
sided high-resolution encoder in addition to the conventional machine ways and ball screws [183,16].
harmonic drive gearbox. The gear motor is used for axis
positioning, as it can provide a high torque. The torque motor 5.4. Reconfigurability
compensates for static and dynamic errors measured on the load-
side. The hybrid drive concept increases the static joint stiffness in Modular robots are conceived as the composition of multiple
control ten times compared to the stiffness of an industrial robot. blocks with uniform interfaces allowing for the transfer of
mechanical forces and torque, electrical power, and communica-
5.3. Hybrid manufacturing tion through the robot. Building blocks consist of a primary
structurally actuated unit and additional specialized units such as
The most economic application of machining robots can be grippers, vision systems and energy storage units [184].
achieved when high flexibility or a large working space are needed. The adoption of modular and reconfigurable robots presents
The combination of two or more process steps can be the basis of a number of advantages [184], such as versatility, robustness
these applications. A hybrid manufacturing concept, combining and lower costs over time. Examples of major results in the field
additive and subtractive processes based on an industrial robot, of reconfigurable robotics can be found in Refs. [185–194]. More
represents an innovative approach to manufacturing technology. recent works deal with: (1) nesting mechanisms for
Exemplary applications are changing the geometry of large robot modules [195–199]; (2) direct and inverse kinematics
molding dies for composite components [84] or body sheets, or the [200–203]; (3) communication and control [204–207]; (4)
production of structural-components, especially for Ni-based motion planning [208,209]; or (5) optimization of robot
alloys [117]. For the robot-based hybrid-manufacturing concept, performance [210,211].
different cladding techniques (e.g., friction surfacing) can be Since accuracy and reliability are more critical for a reconfigur-
combined with a robotic milling process in order to produce near- able robot, the choice of the most appropriate structural
configuration has an importance that goes beyond the mere
satisfaction of workspace reach [212]. Also, their complexity
demands further research efforts in the development of condition
monitoring, fitting this kind of techniques to the peculiar needs
and advantages of modular solutions [213–218].
In conclusion, the research on reconfigurable robotics for
machining will face various open challenges: (i) developing
flexible software frameworks for controlling flawlessly a wide,
potentially unlimited range of kinematic chains; (ii) turning the
concept of reconfigurability in a cost-effective, industrialized
robust solution for manufacturing that can achieve out-of-the-box
Fig. 53. Robot-machining of laser-cladded paths [181]. guaranteed, high-end performances.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

20 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

5.5. Robot programming with function blocks the robots their initial task and then support their continuous
learning over the robot’s deployment. The goal is to enable
As of today, the three main available robot programming autonomous robots to learn a large repertoire of behavioral skills
methods are: lead-through programming with a teach pendant, with minimal human intervention. However, robotic AI applica-
walk-through programming by guiding the robot at its end- tions often compromise the autonomy of the learning process in
effector, and offline programming with a CAM software. favor of achieving training times. This typically involves introduc-
A novel approach is function block programming. It bridges the ing hand-engineered logic representations and human demon-
gap between task planning systems and execution systems, and strations. In robotic learning, research work has explored model-
empowers robot controllers with intelligence. Whilst more based and model-free learning algorithms.
common in human-robot-collaboration applications, it can also Model-based algorithms include a variety of dynamics estima-
provide an additional advantage in robot machining. In function tion schemes [220] such as Gaussian processes, mixture models,
block programming, a non-linear task sequence plan can be and local linear system estimation. Deep neural network policies
mapped automatically to a set of function blocks, which can be have been combined with model-based learning in the context of
readily dispatched to a chosen robot controller for task execution. guided policy search algorithms [221,222]. Such approaches are
The embedded algorithms of the function blocks are triggered exploited for example for scenarios demanding close coordination
inside the robot controller to accomplish the planned tasks. between vision and control (developing policies that map raw
Feedback from human workers is processed and passed to the image observations directly to torques at the robot's motors), or in
function blocks at the robot side for timely robot control. general in situations where it is necessary to learn complex
A function blocks-based automation architecture differs from its feedback control policies mapping high-dimensional sensory
classical counterparts as it enables adaptiveness at multiple levels. inputs to motor torques.
This ensures coping with high rates of component malfunctioning Model-free algorithms include policy search [223] and function
and handling complex system behavior and dynamics based on the approximation methods [224] that have recently been combined
co-ordination of a rich set of signals. Such adaptiveness, both with deep neural networks for learning complex tasks [225]. Direct
architectural and behavioral, is driven by a need to accommodate policy gradient methods offer the benefit of unbiased gradient
external and internal changes, either deterministic or unpredictable. estimates, but tend to require more experience, since on-policy
Each function block contains the control algorithms ruling the estimators preclude reuse of past data.
behavior of the resources at different levels in the automation system In manufacturing, machine learning can serve an additional role
from a single control loop up to the supervision and planning levels. A in design and optimization of processes [226–229], but demands
robotics example can be seen in Fig. 55. future research on real-time processing and control. In current
applications, the robot motion strategy and machining parameters
are adapted in real time based on a sensing system supporting in-
process data collection, their fusion and interpretation. The
optimization logics rely on specific reference quality KPIs. Any
anomalous behavior corresponding to the violation of the quality
indicators results in triggering the optimization model and
enabling the process adaptation logic. The optimization model
can be designed based on various techniques, including AI. Future
works will also consider non-time-critical functions and particu-
larly additional awareness and computational power [230].

6. Conclusions

An increasing role of robots in manufacturing is noticeable,

especially in machining, as exemplified throughout this paper. We
have outlined and investigated opportunities of exploiting the
advantages of robots in this area. One of the trends in robotic
manufacturing is the implementation of accurate and easy-to-use
intuitive programmable systems. The approaches presented in
Fig. 55. Function block distributed control architecture of the robotic cell.
chapter 4 are examples of such systems. Based on this trend and
In robotic applications, examples of function blocks can be the remaining challenges, future directions in this field could be:
elementary functions such as “move to” or “load tool_ID”, that can
be implemented to populate a library of tasks, up to more complex  Simulation of robotic systems could be interfaced with analytical
functions addressing the process recipes and sequence of or numerical models to compute robot motion adaptations for
manufacturing tasks. These function blocks are then sequenced increased tool path accuracy.
in a task schedule to execute specific machining processes such as  Accessible implementation of robot systems with intrinsic
the polishing example reported in Ref. [219]. sensors (i.e. force/torque or vision) capable of automatically
adapting motions in respect to the machining task, thus enabling
5.6. AI in machines and collaborative robots improved process control.
 The path planning process could be optimized using process
The new generation of industrial robots will need to embed redundancy for improving the process quality.
highly advanced control systems, and be able to autonomously  After automatically checking product quality, machine learning
adapt to unexpected situations including performance degrada- algorithms for process optimization could be offered in a cloud-
tion. This can be achieved by adapting the motion parameters. This service platform.
typically limits their joint performance by reducing the most  Development of input devices for facilitating the programming
relevant factors (e.g., acceleration, speed, and deceleration). Path experience. Such devices could record process characteristics
planning adaptation alternatively allows to reduce the movement and update these in robot systems. Furthermore, robot systems
of the degraded joints as much as possible by recomputing could learn and be trained by mimicking professional workers.
trajectories. Artificial intelligence (AI) techniques such as machine
learning and reinforcement learning frameworks can be developed This paper addresses these challenges and shows how they can
to allow task experts without AI or programming expertise to teach be overcome for a wider implementation of robots in machining.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 21

After a theoretical insight, we have presented the possibilities of [14] Hartenberg RS, Denavit J (1965) Kinematic Synthesis of Linkages. McGraw-Hill
Series in Mechanical Engineering, McGraw-Hill, New Yorkp. 435.
identifying and determining system parameters. Different [15] Lipkin H (2005) A Note on Denavit–Hartenberg Notation in Robotics. Volume
approaches for overcoming inherent shortcomings of robots in 7: 29th Mechanisms and Robotics Conference Parts A and B 921–926. http://dx.
machining are distinguishable. The limited static and dynamic doi.org/10.1115/DETC2005-85460. ISBN 0-7918-4744-6.
[16] Siciliano B, Khatib O, (Eds.) (2016), Springer Handbook of Robotics, Springer.
stiffness of robot joints and links still results in insufficient rigidity [17] Bongardt B (2013) Sheth–Uicker Convention Revisited. Mechanism and Ma-
of the TCP, in turn impacting machining accuracy. We have also chine Theory 69:200–229. ISSN 0094-114X.
shown that the static Cartesian stiffness of industrial robots is up to [18] Pan Z, Zhang H, Zhu Z, Wang J (2006) Chatter Analysis of Robotic Machining
Process. Journal of Materials Processing Technology 173:301–309.
50 times lower than the corresponding stiffness of a typical CNC [19] Abele E, Weigold M, Rothenbücher S (2007) Modeling and Identification of
machine tool. However, manufacturing inaccuracies in robot Industrial. Robot for Machining Applications. CIRP Annals Manufacturing
production can be compensated by various methods. Technology 56(1):387–390.
[20] Bondarenko D, Pashkevich A, Briot S, Ritou M, Furet B (2012) Elasto-Dynamic
Due to system complexity, state of the art offline compensation
Model of Robotic Milling Process Considering Interaction between Tool and
is always a simplification and approximation of the real situation. Workpiece. Proceedings of ASME 2012 11th Biennial Conference on Engineering
On the other hand, online compensation can use sensor data from Systems Design and Analysis 1–10.
the real setting, while system complexity hinders the deployment [21] Cen L, Melkote SN (2017) Effect of Robot Dynamics on the Machining Forces in
Robotic Milling. Procedia Manufacturing 10:486–496.
of complex algorithms. Numerous solutions are available, but [22] Cordes M, Hintze W, Altintas Y (2019) Chatter Stability in Robotic Milling.
always for special purposes (i.e. increasing stiffness but not Robotics and Computer-Integrated Manufacturing 55(A):11–18. ISSN
accuracy; or increasing accuracy due to control problems, but not 0736-5845.
[23] Cen L, Melkote SN, Castle J, Appelman H (2018) A Method for Mode Coupling
due to process forces). These solutions still require specialist Chatter Detection and Suppression in Robotic Milling. Journal of Manufactur-
knowledge and are not yet applicable in the industry. ing Science and Engineering 140. 081015-1.
However, not only their lower cost offers robot applications an [24] Mousavi S, Gagnol V, Bouzgarrou BC, Ray P (2017) Dynamic Modeling and
Stability Prediction in Robotic Machining. The International Journal of Ad-
advantage in machining tasks. Several efforts towards improving vanced Manufacturing Technology 88:3053–3065.
machining robot accuracy could be identified with promising [25] Tunc LT, Stoddart D (2017) Tool Path Pattern and Feed Direction Selection in
results. Studies show that only a small fraction of all industrial Robotic Milling for Increased Chatter-Free Material Removal Rate. The Inter-
national Journal of Advanced Manufacturing Technology 89:2907–2918.
robots deployed worldwide in recent years were intended for [26] Karim A, Corcione E, Jäger J, Verl A (2018) Experimental Determination of
machining tasks, a small number compared to their potential in Compliance Values for a Machining Robot. IEEE/ASME International Conference
this area. In order to deploy robots to directly perform on Advanced Mechatronics (AIM), .
[27] Karim A, Schmid S, Verl A (2017) Pose and Feed-Direction Dependency
manufacturing tasks like machining, further investigations are
Analysis for Milling Tasks With Industrial Robots. 24th International Confer-
necessary. Current trends show efforts being made towards ence on Production Research (ICPR), .
developing dedicated kinematics, new drive concepts and new [28] Dumas C, Caro S, Cherif M, Garnier S, et al (2012) Joint Stiffness Identification
fields of application. The collected findings presented here can of Industrial Serial Robots. Robotica 30:649.
[29] http://cognibotics.com/technology/dynamic-modeling/. Retreived 25 Sep-
offer support in this direction. tember 2018.
[30] Nilsson K. Method and System for Determination of at Least One Property of a
Joint, Patent No. BR112015008933 (2), issued Jul. 4 2017.
Acknowledgements [31] Nilsson K, Nilsson A. Method and System For Determination of at Least One
Property of a Manipulator, Patent No. ES2644939 (T3), issued Dec. 1 2017.
[32] Lehmann C, Olofsson B, Nilsson K, Halbauer M, Haage M, Robertsson A,
Contributions to this keynote paper from the following Sörnmo O, Berger U (2013) Robot Joint Modeling and Parameter Identification
colleagues are gratefully acknowledged: Ali Karim, Michael Using the Clamping Method. 7th IFAC Conf on Manuf Modelling, Management,
gan, Eberhard Abele, Berend Denkena, Bernd and Control, 813–818.
Neubauer, Mihai Dra
[33] Celikag H, Sims ND, Ozturk E. Cartesian, Stiffness Optimization for Serial Arm
Kuhlenkötter, Thomas Bauernhansl, Martin Hägele, Eckart Uhl- Robots, Proc CIRP, 8th CIRP Conf. on High Perf. Cutting (HPC 2018).
mann, Wolfgang Hintze, Klas Nilsson, Hendrik Van Brussel, [34] Cen L, Melkote SN (2017) CCT-Based Mode Coupling Chatter Avoidance in
Christian Baier, Felix Hähn, Lars Lienenlüke, Lukas Gründel, Robotic Milling. Journal of Manufacturing Processes 29:50–61.
[35] Chen S-F, Kao I (2000) Conservative Congruence Transformation for Joint and
Thomas Lepper, Benjamin Bergmann, and Dirk Niermann. Cartesian Stiffness Matrices of Robotic Hands and Fingers. The International
Journal of Robotics Research 19(9):835–847.
[36] Lienenlüke L, Gründel L, Storms S, Brecher C (2018) Model-Based Process
Planning for Milling Operations Using Industrial Robots. 3rd Conference on
Control and Robotics Engineering (ICCRE), .
References [37] Mejri S, Gagnol V, Le T-P, Sabourin L, Ray P, Paultre P (2016) Dynamic
Characterization of Machining Robot and Stability Analysis. The International
[1] IFR Statistical Department (2018) World Robotics 2018 Industrial Robots. Journal of Advanced Manufacturing Technology 82:351–359.
[2] https://ifr.org/ifr-press-releases/news/ [38] Karim A, Hitzer J, Lechler A, Verl A (2017) Analysis of the Dynamic Behaviour
robots-double-worldwide-by-2020. Retreived 21 September 2018. of a Six-Axis Industrial Robot within the Entire Workspace in Respect of
[3] Corke P (2011) Robotics, Vision and Control: Fundamental Algorithms in Machining Tasks. 2017 IEEE International Conference on Advanced Intelligent
MATLAB. Volume 73 of Springer Tracts in Advanced Robotics, Springer Science & Mechatronics (AIM), .
Business Media. [39] Kim J, Kim SR, Kim SJ, Kim DH (2010) A Practical Approach for Minimum-Time
[4] Altintas Y (2012) Manufacturing Automation: Metal Cutting Mechanics, Ma- Trajectory Planning for Industrial Robots. Industrial Robot: An International
chine Tool Vibrations, and CNC Design, 2nd edn. Cambridge U. Press. Journal 37(1):51–61.
[5] Appleton E, Williams DJ (1987) Machining with Robots, in Industrial Robot [40] Constantinescu D, Croft EA (2000) Smooth and Time-Optimal Trajectory
Applications, Dordrecht, Springer Netherlands. p. 170. Planning for Industrial Manipulators Along Specified Paths. Journal of Robotic
[6] Devol Jr. GC, “Programmed Article Transfer” U.S. Patent 2988237, issued Jun. Systems 17(5):233–249.
31, 1961. [41] Piazzi A, Visioli A (2000) Global Minimum-Jerk Trajectory Planning of Robot
[7] International Organization for Standardization (ISO) (2012) Robots and Ro- Manipulators. IEEE Transactions on Industrial Electronics (1982) 47(no.
botic Devices: Vocabulary, 2nd edn. . 25.040.30, 01.040.25, 2012(ISO 8373:). 1):140–149.
https://www.iso.org/standard/55890.html. Accessed 7 August 2018. [42] Gasparetto A, Lanzutti A, Vidoni R, Zanotto V (2012) Experimental Validation
[8] Spong MW, Hutchinson S, Vidyasagar M (2006) Robot Modeling and Control, and Comparative Analysis of Optimal Time-Jerk Algorithms for Trajectory
Wiley, Hoboken, NJ op. Planning. Robotics and Computer-integrated Manufacturing 28(2):164–181.
[9] Murray RM, Li Z, Sastry SS (1994) A Mathematical Introduction to Robotic [43] Gasparetto A, Zanotto V (2007) A New Method for Smooth Trajectory Plan-
Manipulation, CRC Press, Boca Raton, Fla. ning of Robot Manipulators. Mechanism and Machine Theory 42(4):455–471.
[10] Jubien A, Gautier M, Janot A (2014) Dynamic Identification of the Kuka LWR [44] Huang P, Chen K, Yuan J, Xu Y (2007) Motion Trajectory Planning of Space
Robot Using Motor Torques and Joint Torque Sensors Data. IFAC Proceedings 47 Manipulator for Joint Jerk Minimization. Proc. 2007 IEEE Int. Conf. Mecha-
(3):8391–8396. tronics Autom. ICMA 2007, 3543–3548.
[11] Janot A, Vandanjon PO, Gautier M (2014) A Generic Instrumental Variable [45] Zha XF (2002) Optimal Pose Trajectory Planning for Robot Manipulators.
Approach for Industrial Robot Identification. IEEE Transactions on Control Mechanism and Machine Theory 37(10):1063–1086.
Systems Technology 22(1):132–145. [46] Perumaal S, Jawahar N (2012) Synchronized Trigonometric S-Curve Trajecto-
[12] Gautier M, Janor A, Vandanjon PO (2008) DIDIM. A New Method for the ry for Jerk-Bounded Time-Optimal Pick and Place Operation. International
Dynamic Identification of Robots from Only Torque Data. 2008 IEEE Interna- Journal of Robotics and Automation 27(4).
tional Conference on Robotics and Automation (ICRA) 2122–2127. [47] Macfarlane S, Croft EA (2003) Jerk-Bounded Manipulator Trajectory Planning:
[13] Wu J, Wang J, You Z (2010) An Overview of Dynamic Parameter Identification Design For Real-Time Applications. IEEE Transactions on Robotics and Auto-
of Robots. Robotics and Computer-Integrated Manufacturing 26(5):414–419. mation 19(1):42–52.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

22 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

[48] Nguyen KD, Ng TC, Chen IM (2008) On Algorithms for Planning S-Curve [79] Alici G, Shirinzadeh B (2005) Enhanced Stiffness Modeling, Identification, and
Motion Profiles. International Journal of Advanced Robotic Systems 5 Characterization For Robot Manipulators. IEEE Transactions on Robotics 21
(1):99–106. (4):554–564.
[49] Porawagama CD, Munasinghe SR (2014) Reduced Jerk Joint Space Trajectory [80] Dumas C, Caro S, Garnier S, Furet B (2011) Joint Stiffness Identification of Six-
Planning Method Using 5-3-5 Spline for Robot Manipulators. 7th International Revolute Industrial Serial Robots. Robotics and Computer-Integrated
Conference on Information and Automation for Sustainability, 1–6. Manufacturing 27:881–888.
[50] Rossi C, Savino S (2013) Robot Trajectory Planning by Assigning Positions and [81] Lo CC, Hsiao CY (1998) A Method of Tool Path Compensation For Repeated
Tangential Velocities. Robotics and Computer-Integrated Manufacturing 29 Machining Process. International Journal of Machine Tools and Manufacture 38
(1):139–156. (3):205–213.
[51] Avram O, Valente A (2016) Trajectory Planning for Reconfigurable Industrial [82] Belchior J, Guillo M, Courteille E, Maurine P, Leotoing L, Guines D (2013) Off-
Robots Designed to Operate in a High Precision Manufacturing Industry. 49th Line Compensation of the Tool Path Deviations on Robotic Machining:
CIRP Conference on Manufacturing Systems (CIRP-CMS 2016), . Application to Incremental Sheet Forming. Robotics and Computer-integrated
[52] Valente A, Baraldo S, Carpanzano E (2017) Smooth Trajectory Generation for Manufacturing 29(4):58–69.
Industrial Robots Performing High Precision Assembly Processes. CIRP Annals [83] Barnfather JD, Goodfellow MJ, Abram T (2016) Development and Testing of an
Manufacturing Technology 66(1):17–20. Error Compensation Algorithm For Photogrammetry Assisted Robotic Ma-
[53] Baraldo S, Valente A (2017) Smooth Joint Motion Planning for High Precision chining. Measurement 94:561–577.
Reconfigurable Robot Manipulators. 2017 IEEE International Conference on [84] Zäh MF, Rösch O (2014) Improvement of the Machining Accuracy of Milling
Robotics and Automation (ICRA), 845–850. Robots. Production Engineering 8(6):737–744.
[54] Saboori A, et al (2018) The Capacity of Cold Spray Additive Manufacturing [85] Tunc LT, Shaw J (2016) Experimental Study on Investigation of Dynamics of
Technology for Metallic Part Repairing. Euro PM2018, . Submitted. Hexapod Robot For Mobile Machining. The International Journal of Advanced
[55] Mekaouche A, Chapelle F, Balandraud X (2015) FEM-Based Generation of Manufacturing Technology 84(5-8):817–830.
Stiffness Maps. IEEE Transactions on Robotics 31(1):217–222. [86] Tunc LT, Shaw J (2016) Investigation of the Effects of Stewart Platform-Type
[56] Nof SY, (Ed.) (1999), Handbook of Industrial Robotics. 2nd ed. ISBN: 978-0-471- Industrial Robot On Stability Of Robotic Milling. The International Journal of
17783-8. Advanced Manufacturing Technology 87(1-4):189–199.
[57] Asada H, Slotine JJE (1992) Robot Analysis and Control, John Wiley & Sons, Inc., [87] Zhang H, Wang J, Zhang G, Gan Z, Pan Z, Cui H, Zhu Z (2005) Machining With
New York, NY, USA. ISBN:0471830291. Flexible Manipulator: Toward Improving Robotic Machining Performance.
[58] Bi S, Liang J (2011) Robotic Drilling System for Titanium Structures. The IEEE/ASME (AIM) International Conference on Advanced Intelligent Mechatro-
International Journal of Advanced Manufacturing Technology 54(May (5– nics, 1127–1132.
8)):767–774. [88] Shi X, Zhang F, Qu X, Liu B (2016) An Online Real-Time Path Compensation
[59] Zhan Q, Wang X (2012) Hand–Eye Calibration and Positioning for a Robot System for Industrial Robots Based on Laser Tracker. International Journal of
Drilling System. The International Journal of Advanced Manufacturing Technol- Advanced Robotic Systems 13(5). 1729881416663366.
ogy 61(July(5–8)):691–701. [89] Zaghbani I, Songmene V, Bonev I (2013) An Experimental Study on the Vibration
[60] Garniera S, Subrina K, Waiyaganb K (2017) Modelling of Robotic Drilling. Response of a Robotic Machining System. Proceedings of the Institution of
Procedia CIRP 58:416–421. Mechanical Engineers Part B: Journal of Engineering Manufacture 227(6):866–880.
[61] Klimchik A, Wu Y, Caro S, Furet B, Pashkevich A (2014) Accuracy Improvement [90] Maurotto A, Tunc LT (2017) Effects of Chattering on Surface Integrity in
of Robot-Based Milling Using an Enhanced Manipulator Model. Advances on Robotic Milling of Alloy 690. Pressure Vessels and Piping Conference ASME
Theory and Practice of Robots and Manipulators Mechanisms and Machine 2017. pp. V06AT06A004, 6 pages.
Science, vol. 22. 73–81. [91] Blomdell A, Bolmsjö G, Cederberg P, Isaksson M, Johansson R, Haage M,
[62] Klimchik A, Bondarenko D, Paskkevich A, Briot S, Furet B, et al (2014) Nilsson K, Olsson M, Olsson T, Robertsson A, Wang J (2005) Extending an
Compliance Error Compensation in Robotic-Based Milling. Ferrier JL, (Ed.) Industrial Robot Controller. IEEE Robotics & Automation Magazine 85–94.
Informatics in Control Automation and Robotics, Lecture Notes in Electrical [92] Bottero A, Gerio G, Perna V, Gagliano A (2014) Adaptive Control Techniques
Engineering 283. 197–214. and Feed Forward Compensation of Periodic Disturbances in Industrial
[63] Altintas Y (2000) Manufacturing Automation, Cambridge University Press. Manipulators. IEEE/ASME 10th International Conference on Mechatronic and
[64] Schneider U, Drust M, Ansaloni M, Lehmann C, Pellicciari M, Leali F, Gunnink Embedded Systems and Applications (MESA) 1–7.
JW, Verl A (2016) Improving Robotic Machining Accuracy through Experi- [93] KUKA Robot Sensor Interface 3.2, KUKA Roboter GmbH.
mental Error Investigation and Modular Compensation. The International [94] Olsson T, Haage M, Kihlman H, Johansson R, Nilsson K, Robertsson A, Björk-
Journal of Advanced Manufacturing Technology 85:3–15. man M, Isaksson R, Ossbahr G, Brogårdh T (2010) Cost-Efficient Drilling Using
[65] Schneider U, Ansaloni M, Drust M, Leali F, Verl A (2013) Experimental Industrial Robots with High-Bandwidth Force Feedback. Robotics and Com-
Investigation of Sources of Error in Robot Machining. Neto P, Moreira AP, puter-integrated Manufacturing 26:24–38.
(Eds.) Robotics in Smart Manufacturing, Communications in Computer and [95] Pan Z, Zhang H (2008) Robotic Machining from Programming to Process
Information Science 371. 14–26. Control: A Complete Solution By Force Control. Industrial Robot an Interna-
[66] Diaz Posada JR, Schneider U, Sridhar A, Verl A (2017) Automatic Motion tional Journal 35(5):400–409.
Generation for Robotic Milling Optimizing Stiffness with Sample-Based [96] Schneider U, Diaz Posada JR, Drust M, Verl A (2013) Position Control of an
Planning. Machines 5(3):1–26. http://dx.doi.org/10.3390/machines5010003. Industrial Robot Using an Optical Measurement System for Machining Pur-
[67] Slavkovic NR, Milutinovic DS, Glavonjic MM (2014) A Method for Off-Line poses. Proceedings of the 11th International Conference on Manufacturing
Compensation of Cutting Force-Induced Errors in Robotic Machining by Tool Research (ICMR2013) 307–312.
Path Modification. The International Journal of Advanced Manufacturing Tech- [97] Schneider U, Olofsson B, Drust M, Robertsson A, Hägele M, Johansson R (2014)
nology 70:2083–2096. Integrated Approach to Robotic Machining With Macro/Micro-Actuation.
[68] Tyapin I, Kaldestad KB, Hovland G (2015) Off-Line Path Correction of Robotic Robotics and Computer-integrated Manufacturing 30:636–647.
Face Milling Using Static Tool Force and Robot Stiffness. IEEE/RSJ International [98] Diaz Posada JR, Schneider U, Pidan S, Geravand M, Stelzer P, Verl A (2016) High
Conference on Intelligent Robots and Systems 5506–5511. Accurate Robotic Drilling With External Sensor and Compliance Model-Based
[69] Brüning J, Denkena B, Dittrich MA, Park HS (2016) Simulation Based Planning of Compensation. Proceedings of the IEEE International Conference on Robotics and
Machining Processes with Industrial Robots. Procedia Manufacturing 6:17–24. Automation (ICRA) 3901–3907.
[70] Cordes M, Hintze W (2017) Offline Simulation of Path Deviation due to Joint [99] Moeller C, Schmidt HC, Koch P, Boehlmann C, Kothe S, Wollnack J, Hintze W
Compliance and Hysteresis For Robot Machining. The International Journal of (2017) Real Time Pose Control of an Industrial Robotic System for Machining
Advanced Manufacturing Technology 90:1075–1083. of Large Scale Components in Aerospace Industry Using Laser Tracker System.
[71] Hollerbach JM, Suh KI-C (1987) Redundancy Resolution of Manipulators SAE International Journal of Aerospace 10(2).
through Torque Optimization. IEEE Robotics & Automation Magazine RA-3 [100] Cen L, Melkote SN (2017) A Wireless Force-Sensing and Model-Based Ap-
(4):308–316. proach for Enhancement of Machining Accuracy in Robotic Milling. IEEE/
[72] Susemihl H, Brillinger C, Stürmer SP, Hansen S, Boehlmann C, Kothe S (2017) ASME Transactions on Mechatronics 21(5):2227–2235.
Referencing Strategies for High Accuracy Machining of Large Aircraft Com- [101] Bongardt T (2004) Methode zur Kompensation betriebsabhängiger Einflüsse auf
ponents with Mobile Robotic Systems. SAE Technical Paper 2017-01-2166. die Absolutgenauigkeit von Industrierobotern, Dissert, TU München, Utz.
http://dx.doi.org/10.4271/2017-01-2166. [102] Abele E, Polle W, Ehm A, Troure M (2013) Spanende Bearbeitung mit indus-
[73] Andres J, Gracia L, Tornero J (2011) Calibration and Control of a Redundant trierobotern - thermische einflüsse auf die bearbeitungsgenauigkeit. Werk-
Robotic Workcell For Milling Tasks. International Journal of Computer Inte- stattstechnik. online, no. 9.
grated Manufacturing 24(6):561–573. [103] Ehm A (2016) Einsatz von Industrierobotern für die Bohrbearbeitung an
[74] Xiao W, Huan J (2012) Redundancy and Optimization of a 6R Robot for Five- automobilen Strukturbauteilen unter Berücksichtigung des thermischen
Axis Milling Applications: Singularity, Joint Limits and Collision. Production Verlagerungsverhaltens und der Prozessinteraktion, Schriftreihe des PTW
Engineering Research and Development 6:287–296. “Innovation Fertigungstechnik”. Darmstadt.
[75] Sabourin L, Subrin K, Cousturier R, Gogu G, Mezouar Y (2016) Redundancy- [104] Reinhart G, Gräser R-G, Klingel R (1998) Qualification of Standard Industrial
Based Optimization Approach to Optimize Robotic Cell Behaviour: Applica- Robots to Cope With Sophisticated Assembly Tasks. CIRP Annals Manufactur-
tion to Robotic Machining. Industrial Robot: An International Journal 42 ing Technology 47(1). http://dx.doi.org/10.1016/S0007-8506(07)62772-3.
(2):156–166. [105] Du G, Zhang P (2013) Online Robot Calibration Based on Vision Measurement.
[76] Mousavi S, Gagnol V, Bouzgarrou B, Ray P (2017) Control of a Multi Degrees Robotics and Computer-Integrated Manufacturing 29(6). http://dx.doi.org/
Functional Redundancies Robotic Cell for Optimization of the Machining 10.1016/j.rcim.2013.05.003.
Stability. Procedia CIRP 58:269–274. [106] Wiest U (2009) Kinematische Kalibrierung von Industrierobotern. Dissertation,
[77] Tobias SA (1965) Machine-Tool Vibration, Blackie & Son Ltd., London. Universität Karlsruhe, Karlsruhe.
[78] Yuan L (2017) A Study of Chatter in Robotic Machining and A Semi-Active Chatter [107] Boochs F, Schutze R, Simon C, Marzani F, Wirth H, Meier J (2010) Increasing
Suppression Method Using Magnetorheological Elastomers (MREs). M. Eng. the Accuracy of Untaught Robot Positions by Means of a Multi-Camera
Thesis, School of Mechanical, Material, Mechatronic and Biomedical Engi- System. International Conference on Indoor Positioning and Indoor Navigation
neering, University of Wollongong, Australia. (IPIN).

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx 23

[108] Santolaria J, Conte J, Ginés M (2013) Laser Tracker-Based Kinematic Parame- [143] Buff B, Magnus C, Zhu J, Meier H (2013) Robot-Based Incremental Sheet Metal
ter Calibration of Industrial Robots by Improved CPA Method and Active Forming — Increasing the Geometrical Complexity and Accuracy. Key Engi-
Retroreflector. The International Journal of Advanced Manufacturing Technology neering Materials. Proceedings of the 15th International Conference on Sheet
66(9):2087–2106. Metal — SHEMET 2013 549:149–155.
[109] Yin S, et al (2014) Real-time Thermal Error Compensation Method for Robotic [144] Störkle DD, Möllensiep D, Thyssen L, Kuhlenkötter B (2018) Geometry-
Visual Inspection System. The International Journal of Advanced Manufacturing Dependent Parameterization of Local Support in Robot-Based Incremental
Technology 75(5-8):933–946. Sheet Forming. Procedia Manufacturing 15:1164–1169.
[110] Elbestawi MA, Bone GM, Tern PW (1992) An Automated Planning, Control, [145] Meier H, Zhu J, Buff B, Laurischkat R (2012) CAx Process Chain for Two Robots
and Inspection System For Robotic Deburring. CIRP Annals Manufacturing Based Incremental Sheet Metal Forming. Procedia CIRP 3:37–42.
Technology 397. Manuf Tech. [146] Laurischkat R (2012) Kompensation prozesskraftbedingter Bahnfehler bei der
[111] Möller C, Schmidt H, et al (2016) Enhanced Absolute Accuracy of an Industrial roboterbasierten inkrementellen Blechumformung, Dissertation, Ruhr-Universi-
Milling Robot Using Stereo Camera System. Procedia Technology 389–398. tät Bochum, Lehrstuhl für Produktionssysteme.
http://dx.doi.org/10.1016/j.protcy.2016.08.050. [147] Li Y, Liu Z, Lu H, Daniel WJT, Liu S, Meehan PA (2014) Efficient Force Prediction
[112] Karim A, Verl A (2013) Challenges and Obstacles in Robot-Machining. 44th for Incremental Sheet Forming and Experimental Validation. The International
IEEE International Symposium on Robotics (ISR) 224–228. Journal of Advanced Manufacturing Technology 73(1-4):571–587.
[113] Bongardt T (2003) Methode Zur Kompensation Betriebsabhängiger Einflüsse Auf [148] Wang J, Nair M, Zhang Y (2017) An Efficient Force Prediction Strategy for
Die Absolutgenauigkeit Von Industrierobotern. Dissertation, TU München. Single Point Incremental Sheet Forming. The International Journal of Advanced
[114] Gossel O (1996) Steigerung der Genauigkeit von Industrierobotern basierend auf Manufacturing Technology 92(9-12):3931–3939.
einer durchgängigen Genauigkeitsanalyse, Univ. Hamburg. [149] Abele E, Bauer J, Hemker T, Laurischkat R, Meier H, Reese S, von Stryk O (2011)
[115] Roos E (1998) Anwendungsorientierte Me- und Berechnungsverfahren zur Kali- Comparison and Validation of Implementations of a Flexible Joint Multibody
brierung offline programmierter Roboterapplikationen. Diss.Hamburg . Dynamics System Model for an Industrial Robot. CIRP Journal of Manufactur-
[116] Bauer J (2011) Methoden Der Offline-Bahnkorrektur Für Die Spanende Bearbei- ing Science and Technology 4(1):38–43.
tung Mit Industrierobotern. Dissertation, TU Darmstadt. [150] Hands Off Dies and Molds. The Open Industrial & Manufacturing Engineering
[117] Ergebnisbericht zum Verbundvorhaben. Hochproduktive generative Produkt- Journal 14(October).
herstellung durch laserbasiertes, hybrides Fertigungskonzept. PROGEN. [151] Bogue R (2009) Finishing Robots: A Review of Technologies and Applications.
Fraunhofer IRB Verlag. Dresden. September 2018. Industrial Robot an International Journal 36(1):6–12.
[118] Rösch O. Steigerung der Arbeitsgenauigkeit bei der Fräsbearbeitung metal- [152] Takeuchi Y, Asakawa N, Ge D (1993) Automation of Polishing Work By an
lischer Werkstoffe mit Industrierobotern. Zugl.: München, Techn. Univ., Fak. Industrial Robot: System of Polishing Robot. JSME International Journal Series
für Maschinenwesen. Dissertation, 2014. München: Utz, 2015. C Dynamics Control Robotics Design and Manufacturing 36(4):556–561.
[119] Lienenlüke L, Gründel L, Storms S, Brecher C (2018) Model-Based Process [153] Mizugaki Y, Sakamoto M, Kamijo K, Taniguchi N (1990) Development of
Planning for Milling Operations Using Industrial Robots. 3rd International Metal-Mold Polishing Robot System With Contact Pressure Control Using
Conference on Control and Robotics Engineering. 978-1-5386-6662-3 (to be CAD/CAM Data. CIRP Annals-Manufacturing Technology 39(1):523–526.
published end of 2018). [154] Saito K, Miyoshi T, Sasaki T (1993) Automation of Polishing Process for a
[120] Robotic Counter Sinking Cell. https://bit.ly/2rcvCUT. Accessed 17 May 2018. Cavity Surface on Dies and Molds by Using an Expert System. CIRP Annals-
[121] Möller C, Schmidt HC, Koch P, Böhlmann C, Kothe SM, Wollnack J, Hintze W Manufacturing Technology 42(1):553–556.
(2017) Machining of Large Scaled CFRP-Parts with Mobile CNC-Based Robotic [155] Mizugaki Y, Sakamoto M, Sata T (1992) Fractal Path Generation for a Metal-
System in Aerospace Industry. Procedia Manufacturing 14:17–29. Mold Polishing Robot System and Its Evaluation by the Operability. CIRP
[122] Cognitively enhanced robot for flexible manufacturing of metal and composite Annals Manufacturing Technology 41(1):531–534.
parts (COROMA). https://www.coroma-project.eu/. Accessed 23 March 2018. [156] Tam HY, Lui OCH, Mok AC (1999) Robotic Polishing of Free-Form Surfaces
[123] Susemihl H, Moeller C, Kothe S, Schmidt H, et al (2016) High Accuracy Mobile Using Scanning Paths. Journal of Materials Processing Technology 95(1-
Robotic System for Machining of Large Aircraft Components. SAE International 3):191–200.
Journal of Aerospace 9(2):231–238. http://dx.doi.org/10.4271/2016-01-2139. [157] Márquez JJ, Pérez JM, R??ios J, Vizán A (2005) Process Modeling for Robotic
[124] Leali F, Pellicciari M, Pini F, Berselli G, Vergnano A (2013) An Offline Pro- Polishing. Journal of Materials Processing Technology 159(1):69–82.
gramming Method for the Robotic Deburring of Aerospace Components. Neto [158] Tsai MJ, Fang JJ, Chang JL (2004) Robotic Path Planning for an Automatic
P, Moreira AP, (Eds.) Robotics in Smart Manufacturing Communications in Mold Polishing System. International Journal of Robotics and Automation 19
Computer a Information Science, vol. 371. Springer, Berlin Heidelberg. (2):81–90.
[125] Mohammad M, Babriya V, Sobh T (2008) Modeling a Deburring Process, Using [159] Tsai MJ, Huang JF, Kao WL (2009) Robotic Polishing of Precision Molds with
DELMIA V5. Journal of Engineering and Applied Sciences 3(11):835–847. Uniform Material Removal Control. International Journal of Machine Tools and
[126] Princely L, Selvaraj T (2014) Vision Assisted Robotic Deburring of Edge Burrs Manufacture 49(11):885–895.
in Cast Parts. Procedia Engineering 97:1906–1914. [160] Yang Z, Xi F, Wu B (2005) A Shape Adaptive Motion Control System with
[127] Chung J (2007) Modeling and Control of a New Robotic Deburring System. Application to Robotic Polishing. Robotics and Computer-integrated
Journal of Manufacturing Science and Engineering 129(5):965–972. Manufacturing 21(4-5):355–367.
[128] Jayaweera N, Webb P (2011) Robotic Edge Profiling of Complex Components. [161] Sabourin M, Paquet F, Hazel B, Côté J, Mongenot P (2010) Robotic Approach to
Industrial Robot an International Journal 38:38–47. Improve Turbine Surface Finish. Applied Robotics for the Power Industry
[129] Zhang H, Chen H, Xi N, Zhang G, He J (2006) On-Line Path Generation for Robotic (CARPI) 2010 1st International Conference 1–6.
Deburring of Cast Aluminium Wheels. IEEE Conf on Intel Rob and Syst 2006. [162] Ryuh BS, Park SM, Pennock GR (2006) An Automatic Tool Changer and
[130] Song H, Song J (2013) Precision Robotic Deburring Based on Force Control for Integrated Software for a Robotic Die Polishing Station. Mechanism and
Arbitrarily Shaped Workpiece Using CAD Model Matching. International Machine Theory 41(4):415–432.
Journal of Precision Engineering and Manufacturing 14(1):85–91. [163] Tian F, Lv C, Li Z, Liu G (2016) Modeling and Control of Robotic Automatic
[131] Villagrossi E, Pedrocchi N, Beschi M, Tosatti LM (2018) A Human Mimicking Polishing For Curved Surfaces. CIRP Journal of Manufacturing Science and
Control Strategy For Robotic Deburring of Hard Materials. International Technology 14:55–64.
Journal of Computer Integrated Manufacturing. [164] Denkena B, Lepper T (2015) Enabling an Industrial Robot for Metal Cutting
[132] Dietz T, Schneider U, Barho M, Oberer-Treitz S, Drust M, Hollmann R, Hägele M Operations. Procedia CIRP 35:79–84.
(2012) Programming System for Efficient Use of Industrial Robots for Debur- [165] Denkena B, Brüning J, Windel L, Euhus D, Kirsch S, Overbeck D, Lepper T
ring in SME Environments. 7th German Conference on Robotics. ROBOTIK, . (2017) Holistic Process Planning Chain For Robot Machining. Production
[133] Kuss A, Drust M, Verl A (2016) Detection of Workpiece Shape Deviations for Engineering 111(9):515.
Tool Path Adaptation in Robotic Deburring Systems. 49th CIRP Conference on [166] Iglesias I, Sebastián MA, Ares JE (2015) Overview of the State of Robotic
Manufacturing Systems, . Machining: Current Situation and Future Potential. Procedia Engineering
[134] Diaz Posada JR, Kumar S, Kuss A, Schneider U, Drust M, Dietz T, Verl A (2016) 132:911–917.
Automatic Programming and Control for Robotic Deburring. 47th Internation- [167] Puzik A, Meyer C, Verl A (2010) Robot Machining With Additional 3D-Piezo-
al Symposium on Robotics ISR 2016, . Actuation Mechanism for Error Compensation. 41st International Symposium
[135] Breun F, Wohnig W, Siegert K (1990) Werkzeugbau für prototypen, Klein- und on Robotics (ISR) & 6th German Conference on Robotics (ROBOTIK) 415–421.
mittelserien. Neuere Entwicklungen in der Blechumformung 227–243. [168] Furtado LFF, Villani E, Trabasso LG, Sutério R (2017) A Method to Improve the
[136] SCHULER GmbH (1998) Metal Forming Handbook, Springer, Berlin, Heidelberg. Use of 6-dof Robots as Machine Tools. The International Journal of Advanced
[137] Allwood JM, King GPF, Duflou J (2005) A Structured Search for Applications of Manufacturing Technology 92:2487–2502.
the Incremental Sheet-Forming Process by Product Segmentation. Journal of [169] Abele E, Weigold M, Kulok M (2005) Spanende Bearbeitung mit Industrie-
Engineering Manufacture 219(2):239–244. robotern: Statische und dynamische Analyse der vollseriellen Kinematik95 .
[138] Amino H, Lu Y, Maki T, Osawa S, Fukuda K, Dieless NC, (Ed.) (2002), Forming, [170] Abele E, Weigold M (2007) Steigerung der Genauigkeit und Bauteilqualität bei
Prototype of Automotive Service Parts. der spanenden Bearbeitung mit Industrierobotern.
[139] Bramley AN, Vollersten F, Kleiner M (2001) Incremental Sheet Forming [171] Kurze M (2008) Modellbasierte Regelung von Robotern mit elastischen Gelenken
Process for Small Batch and Prototype Parts. Idee-Vision-Innovation Verlag ohne abtriebsseitige SensorikMünchen .
Meisenbach Bamberg 95–102. [172] Weigold M (2008) Kompensation der Werkzeugabdrängung bei der spanenden
[140] Jeswiet J (2005) Asymmetric Incremental Sheet Forming. Advanced Materials Bearbeitung mit IndustrieroboternShaker, Darmstadt .
Research. Proceedings of the 11th International Conference on Sheet Metal — [173] Karim A, Waibel A, Lechler A (2016) Experimentelle Posenanalyse am Bearbei-
SheMet 6-8:35–58. tungsroboter: Ermittlung der Richtungs- sowie Posenabhängigkeit von Indus-
[141] Meier H, Buff B, Laurischkat R, Smukala V (2009) Increasing the Part Accuracy trierobotern für die Fräsbearbeitung 106p. 347 .
in Dieless Robot-Based Incremental Sheet Metal Forming. CIRP Annals [174] Brecher C, Lienenlüke L, Obdenbusch M (2017) Spanende Bearbeitung mit
Manufacturing Technology 58:233–238. Industrierobotern: Modellbasierte Planungsassistenz für die roboterbasierte
[142] Smukala V (2012) Systematische untersuchung der druckspannungs-überla- Bearbeitung 107346–351.
gerung in der roboterbasierten inkrementellen blechumformung, disserta- [175] Abele E, Bauer J, Stelzer M, von Stryk O (2008) Wechselwirkung von Fräsprozess
tion, Ruhr-Universität Bochum. Lehrstuhl für Produktionssysteme. und Maschinenstruktur am Beispiel des Industrieroboters733–737.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009
CIRP 2028 1–24

24 A. Verl et al. / CIRP Annals - Manufacturing Technology xxx (2019) xxx–xxx

[176] Ozturk E, Barrios A, Sun C, Rajabi S, Munoa J (2018) Robotic Assisted Milling [204] Qiao G, Song G, Zhang J, Sun H, Ge J, Wang W (2012) Design of a Self-
for Increased Productivity. CIRP Annals Manufacturing Technology. http://dx. Reconfigurable Wireless Network System for Modular Self-Reconfigurable
doi.org/10.1016/j.cirp.2018.04.031. Robots. 2012 IEEE International Conference on Robotics and Biomimetics, ROBIO
[177] Denkena B, Bergmann B, Lepper T (2017) Design and Optimization of a 2012 — Conference Digest, 1337–1342.
Machining Robot. Procedia Manufacturing 14:89–96. [205] Bordignon M, Stoy K, Schultz UP (2011) Generalized Programming of Modular
[178] Denkena B, Immel J, Schönherr M (2011) Adaptronic Systems in Robot Robots Through Kinematic Configurations. IEEE International Conference on
Manufacturing. Advanced Materials Research ;(383-390)1013–1018. Intelligent Robots and Systems, 3659–3666.
[179] Denkena B, Litwinski K, Schönherr M (2013) Innovative Drive Concept for [206] Castro S, Koehler S, Kress-Gazit H (2011) High-Level Control of Modular
Machining Robots. Procedia CIRP 9:67–72. Robots. International Conference on Intelligent Robots and Systems, 3120–3125.
[180] Baier C (2017) Process and Software Chain for Combined Laser Cladding and [207] Hofbaur M, Brandstötter M, Jantscher S, Schörghuber C (2010) Modular Re-
Milling in Robot Systems. 9th International Laser Symposium. Configurable Robot Drives. 2010 IEEE Conference on Robotics, Automation and
[181] Abele (2017) Besser qualitätsschleifen. Form Werkzeug Carl Hanser Verlag Mechatronics, RAM 2010, 150–155.
München 26(4):62–63. ISSN 1439-667X. [208] Vonasek V, Saska M, Kosnar K, Preucil L (2013) Global Motion Planning for
[182] König M (2016) Sinumerik-gesteuerte Montage und Zerspanung von Motoren Modular Robots With Local Motion Primitives. Proceedings - IEEE Interna-
bei Siemens, Technologietag “Zerspanen mit Industrierobotern”. Darmstadt. tional Conference on Robotics and Automation, 2465–2470.
[183] Kohkonen KE, Potdar N (1998) Composite Machining. in Peters ST, (Ed.) [209] Hou F, Shen WM (2010) On the Complexity of Optimal Reconfiguration
Handbook of Composites., Springer, Boston, MA. Planning for Modular Reconfigurable Robots. Proceedings - IEEE International
[184] Yim M, et al (2007) Modular Self-Reconfigurable Robot Systems [Grand Conference on Robotics and Automation, 2791–2796.
Challenges of Robotics]. IEEE Robotics & Automation Magazine 14(1):43–52. [210] Lin Y, Xi F, Mohamed RP, Tu X-W (2010) Calibration of Modular Reconfigurable
[185] Fukuda T, Nakagawa S, Kawauchi Y, Buss M (1989) Structure Decision Method Robots Based on a Hybrid Search Method. Journal of Manufacturing Science
for Self Organising Robots Based on Cell Structures-CEBOT. Proceedings of the and Engineering ASME 132(no. 6):610021–610028.
1989 International Conference on Robotics and Automation, 695–700. [211] Mohamed RP, Xi Jr F, Finistauri AD (2010) Module-Based Static Structural
[186] Yim M (1994) New Locomotion Gaits. Proceedings of the 1994 IEEE Interna- Design of a Modular Reconfigurable Robot. Journal of Mechanical Design 132
tional Conference on Robotics and Automation, 2508–2514. (1):014501.
[187] Murata S, Yoshida E, Kamimura A, Kurokawa H, Tomita K, Kokaji S (2002) M- [212] Valente A (2016) Reconfigurable Industrial Robots: A Stochastic Programming
TRAN: Self-Reconfigurable Modular Robotic System. IEEE/ASME Transactions Approach For Designing and Assembling Robotic Arms. Robotics and Comput-
on Mechatronics: A Joint Publication of the IEEE Industrial Electronics Society and er-integrated Manufacturing 41:115–126.
the ASME Dynamic Systems and Control Division 7(4):431–441. [213] Galar Pascual D (2015) Artificial Intelligence Tools — Decision Support Systems in
[188] Castano A, Behar A, Will PM (2002) The Conro Modules for Reconfigurable Condition Monitoring and Diagnosis, CRC Press.
Robots. IEEE/ASME Transactions on Mechatronics: A Joint Publication of the IEEE [214] Jaber AA, Bicker R (2014) The State Of The Art In Research Into The Condition
Industrial Electronics Society and the ASME Dynamic Systems and Control Monitoring Of Industrial Machinery. International Journal of Current Engineer-
Division 7(4):403–409. ing and Technology 4(3):1986–2001.
[189] Jørgensen MW, Østergaard EH, Lund HH (2004) Modular ATRON: Modules for [215] Elosegui P (1992) Measurement of the Dynamic Model of a Puma 560 Ro-
a Self-Reconfigurable Robot. Proceedings of 2004 IEEE/RSJ International Con- bot Using Experimental Modal Analysis. Journal of Mechanical Design 116
ference on Intelligent Robots and Systems, . (1):75–79.
[190] Lyder A, Garcia RFM, Stoy K (2008) Mechanical Design of Odin, an Extendable [216] Pan MC, Van Brussel H, Sas P (1998) Intelligent Joint Fault Diagnosis of
Heterogeneous Deformable Modular Robot. 2008 IEEE/RSJ International Con- Industrial Robots. Mechanical Systems and Signal Processing 12(4):571–588.
ference on Intelligent Robots and Systems 883–888. [217] Liu H, Wei T, Wang X (2009) Signal Decomposition and Fault Diagnosis of a
[191] Sproewitz A, Billard A, Dillenbourg P, Ijspeert AJ (2009) Roombots-Mechani- Scara Robot Based Only on Tip Acceleration Measurement. IEEE International
cal Design Of Self-Reconfiguring Modular Robots for Adaptive Furniture. Conference on Mechatronics and Automation ICMA, 4811–4816.
2009 IEEE International Conference on Robotics and Automation ;(c)4259–4264. [218] Trendafilova I, Van Brussel H (2003) Condition Monitoring of Robot Joints
[192] Davey J, Kwok N, Yim M (2012) Emulating Self-Reconfigurable Robots - Using Statistical and Nonlinear Dynamics Tools. Meccanica 38(2):283–295.
Design of the SMORES System. IEEE International Conference on Intelligent [219] Valente A, Carpanzano E (2011) Development of Multi-Level Adaptive Control
Robots and Systems, 4464–4469. And Scheduling Solutions For Shop-Floor Automation In Reconfigurable
[193] Gilpin K, Kotay K, Rus D, Vasilescu I (2008) Miche: Modular Shape Formation Manufacturing Systems. CIRP Annals — Manufacturing Technology 60
By Self-Disassembly. The International Journal of Robotics Research 27(3– (1):449–452.
4):345–372. [220] Deisenroth MP, Neumann G, Peters J (2011) A Survey On Policy Search For
[194] Yim M, Duff D, Roufas K (2000) Polybot: a Modular Reconfigurable Robot. IEEE Robotics. Foundations and Trends in Robotics 2(1–2):1–142.
International Conference on Robotics and Automation, 514–520. [221] Levine S, Finn C, Darrell T, Abbeel P (2016) “End-to-End Training of Deep
[195] Liedke J, Wörn H (2011) CoBoLD — A Bonding Mechanism For Modular Self- Visuomotor Policies. Journal of Machine Learning Research 17(1):1334–1373.
Reconfigurable Mobile Robots. 2011 IEEE International Conference on Robotics [222] Chebotar Y, Kalakrishnan M, Yahya A, Li A, Schaal S, Levine S (2017) Path
and Biomimetics ROBIO 2011 2025–2030. Integral Guided Policy Search. Proceedings — IEEE International Conference on
[196] Thorne CE, Skorodinski N, Tipton H, Van Schoyck T, Yim M (2010) Brake Robotics and Automation, 3381–3388.
Design for Dynamic Modular Robots. Proceedings — IEEE International Con- [223] Peters J, Schaal S (2008) Reinforcement Learning of Motor Skills With Policy
ference on Robotics and Automation, 3135–3140. Gradients. Neural Networks 21(4):682–697.
[197] Vespignani M, Senft E, Bonardi S, Moeckel R, Ijspeert AJ (2013) An Experi- [224] Riedmiller M (2005) Neural Fitted Q Iteration — First Experiences With a Data
mental Study on the Role of Compliant Elements on the Locomotion of the Efficient Neural Reinforcement Learning Method. Lecture Notes in Computer
Self-Reconfigurable Modular Robots Roombots. IEEE International Conference Science (including subseries Lecture Notes in Artificial Intelligence and Lecture
on Intelligent Robots and Systems, 4308–4313. Notes in Bioinformatics) 3720:317–328. LNAI.
[198] Catalano MG, et al (2011) VSA-CubeBot: A Modular variable Stiffness Platform [225] Mnih V, et al (2016) Asynchronous Methods for Deep Reinforcement Learn-
for Multiple Degrees of Freedom Robots. Proceedings — IEEE International ing. International Conference on Machine Learning, 1928–1937.
Conference on Robotics and Automation, 5090–5095. [226] Nikolaidis S, Lasota P, Rossano G, Martinez C, Fuhlbrigge T, Shah J (2013)
[199] Aghili F, Parsa K (2009) A Reconfigurable Robot with Lockable Cylindrical Human-Robot Collaboration in Manufacturing: Quantitative Evaluation of
Joints. IEEE Transactions on Robotics: A Publication of the IEEE Robotics and Predictable, Convergent Joint Action. 44th International Symposium on
Automation Society 25(4):785–797. Robotics.
[200] Pan X, Wang H, Jiang Y, Yu C (2011) Research on Kinematics of Modular [227] Cencen A, Verlinden JC, Geraedts J (2016) Qualifying the Performance of
Reconfigurable Robots. 2011 IEEE International Conference on Cyber Technology Human Robot Coproduction at a Relabling Station. TMCE 2016.
in Automation, Control, and Intelligent Systems, CYBER 2011, 91–96. [228] Chhim P, Chinnam RB, Sadawi N (2017) Product Design and Manufacturing
[201] Huang T, Li M, Zhao XM, Mei JP, Chetwynd DG, Hu SJ (2005) Conceptual Process Based Ontology For Manufacturing Knowledge Reuse. Journal of
Design And Dimensional Synthesis for a 3-DOF Module of the TriVariant — A Intelligent Manufacturing 1–12.
Novel 5-DOF Reconfigurable Hybrid Robot. IEEE Transactions on Robotics: A [229] Zhang Y, Huang GQ, Sun S, Yang T (2014) Multi-agent Based Real-Time
Publication of the IEEE Robotics and Automation Society 21(3):449–456. Production Scheduling Method For Radio Frequency Identification Enabled
[202] Wu W, et al (2013) Task-Oriented Inverse Kinematics of Modular Reconfigur- Ubiquitous Shopfloor Environment. Computers & Industrial Engineering 76
able Robots. 2013 IEEE/ASME International Conference on Advanced Intelligent (1):89–97.
Mechatronics: Mechatronics for Human Wellbeing 1187–1192. [230] Wu D, Rosen DW, Wang L, Schaefer D (2015) Cloud-Based Design and
[203] Köker R (2013) A Genetic Algorithm Approach to a Neural-Network-Based Manufacturing: A New Paradigm in Digital Manufacturing and Design Inno-
Inverse Kinematics Solution of Robotic Manipulators Based on Error Mini- vation. Computer-Aided Design 59:1–14.
mization. Information Sciences 222:528–543.

Please cite this article in press as: A. Verl, et al., Robots in machining, (2019), https://doi.org/10.1016/j.cirp.2019.05.009