CILAP-Architecture for Simultaneous Position- and Force-Control in

Constrained Manufacturing Tasks

Sophie Klecker, Bassem Hichri and Peter Plapper

Faculty of Science, Technology and Communication, University of Luxembourg,

6, rue Richard Coudenhove-Kalergi, L-1359 Luxembourg, Luxembourg

Keywords: Parallel Control, Simultaneous Position- and Force-Control, Constrained Manufacturing, Bio-inspired.

Abstract: This paper presents a parallel control concept for automated constrained manufacturing tasks, i.e. for

simultaneous position- and force-control of industrial robotic manipulators. The manipulator’s interaction

with its environment results in a constrained non-linear switched system. In combination with internal and

external uncertainties and in the presence of friction, the stable system performance is impaired. The aim is

to mimic a human worker’s behaviour encoded as lists of successive desired positions and forces obtained

from the records of a human performing the considered task operating the lightweight robot arm in gravity

compensation mode. The suggested parallel control concept combines a model-free position- and a model-

free torque-controller. These separate controllers combine conventional PID- and PI-control with adaptive

neuro-inspired algorithms. The latter use concepts of a reward-like incentive, a learning system and an

actuator-inhibitor-interplay. The elements Conventional controller, Incentive, Learning system and Actuator-

Preventer interaction form the CILAP-concept. The main contribution of this work is a biologically inspired

parallel control architecture for simultaneous position- and force-control of continuous in contrast to discrete

manufacturing tasks without having recourse to visual inputs. The proposed control-method is validated on a

surface finishing process-simulation. It is shown that it outperforms a conventional combination of PID- and

PI-controllers.

1 INTRODUCTION

Automation of contact-based manufacturing

processes is of significant interest to the industrial as

well as to the scientific community. Humans being

highly proficient at manufacturing tasks requiring

compliance and force control, high number of

research works in the field aim to mimic the human

workers’ behaviour and to translate its capabilities

into robot skills.

(Rozo et al., 2013) used Programming by

Demonstration, PbD which teaches a robot by

showing the desired behaviour rather than by writing

commands in a programming language. Based on

Gaussian mixture theory, a single model encompasses

both, desired positions and forces. (Abu-Dakka et al.,

2015) presented a concept for learning and adaptation

of contact-based manipulation tasks. The authors

suggested a scheme for online modifications to match

desired reference position- and force-profiles. The

latter were obtained from programming by

demonstration and encoded with dynamic movement

primitives. (Oba et al., 2016) discussed the

acquisition and replication of polishing skills of a

human worker represented as tool trajectory, tool

posture and polishing force. These variables which

were to be controlled independently and

simultaneously formed the input to the controller.

As far as the considered processes are concerned,

most manufacturing tasks require the robotic

manipulator to interact with its environment which

results in contact situations and constrained

movements, i.e. the robot arm cannot move freely in

all directions. Constraints include natural constraints

due to the specificities of the environment as well as

artificial constraints due to and characteristic of the

desired task. Varying or switching constraints are due

to successive discrete or continuous phases in a task.

By their nature, the control of constrained tasks

requires the simultaneous control of pose and force.

Pure position control cannot cope with these complex

tasks because already slight deviations from the

desired trajectory can lead to errors in the desired

forces and torques (Abu-Dakka et al., 2015). Pure

244

Klecker, S., Hichri, B. and Plapper, P.

CILAP-Architecture for Simultaneous Position- and Force-Control in Constrained Manufacturing Tasks.

DOI: 10.5220/0006828902440251

In Proceedings of the 15th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2018) - Volume 2, pages 244-251

ISBN: 978-989-758-321-6

force control on the other hand can lead to contact

instabilities at an increased speed (Newman et al.,

1999). A promising approach for simultaneous

position- and force-control is parallel control, i.e. two

controllers acting in parallel. Both independent

controllers yield control torque commands which are

summed up. In contrast to other state-of-the-art

hybrid control methods, parallel control allows for the

simultaneous and independent control of position-

and force-signals. Parallel control has been the

subject of repeated research efforts over the past

decades. Based on the interactions of controller, robot

arm and environment, (Chiaverini and Sciavicco,

1993) developed a dynamic parallel force-/position-

control for constrained motions with an elastic

environment. (Ferguene et al., 2009) extended a

conventional parallel force-/position-controller with a

3-layer feed-forward neural network to compensate

for uncertain or varying robot dynamics and

environments. The intended application areas ranged

from elastic environments over curved surfaces to

unknown environmental stiffness. (Karayiannidis and

Doulgeri, 2010) suggested adaptive concepts for

position-/force-control in compliant and frictional

contacts in the presence of uncertainties in models,

end effector-orientation and environment. (Yin et al.,

2012) based his tracking controller on a human

analogy, i.e. on the human’s approach to finger

tracking in the absence of visual feedback. For the

tracking of an unknown surface, the authors relied on

the concepts of moving frames and vector-variations.

(Lange et al., 2013) presented a parallel position-

based force-/torque-control scheme taking into

account couplings between forces and torques,

constrained configurations, compliances in robot,

sensor and environment as well as the effects due to

impact forces. An experimental validation completed

the work. The cited contributions present some

drawbacks for the here considered automated

manufacturing task. The majority of the state-of-the-

art controllers are model-based and based only on

conventional concepts, not taking advantage of gains

in robustness and adaptability offered by intelligent

control-extensions.

Over the past decades, some research has also

been done separately on control-algorithms. Since

their introduction in 1940, model-free PID-

controllers have been predominant in industrial

settings (Adar and Kozan, 2016). This is due to their

key-advantages: robustness and simple design. Their

constant fixed parameters as well as their linearity

however, make it hard to cope with either nonlinear,

time-varying systems or disturbances. The lack of

flexible adaptability and the impossibility to increase

gains arbitrarily due to actuator limitations as well as

the occurrence of instabilities and noise sensitivity

(Kuc and Han, 2000) (Siciliano and Khatib, 2008)

limit the application areas. Conventional controllers

are therefore not suited for controlling manufacturing

processes automated with highly nonlinear, coupled

robotic systems (Adar and Kozan, 2016) (Kuc and

Han, 2000) (Siciliano and Khatib, 2008).

With the aim to take advantage of the

conventional controller’s trumps while overcoming

its drawbacks, biomimetic extensions imitating the

learning behaviour of the human brain are presented.

Due to the complex nature of this biological system

only a concise selection of its key-aspects has been

retained for the development of control concepts.

(Lucas et al., 2004) presented BELBIC (Brain

Emotional Learning-Based Intelligent Control) on the

base of the work by (Balkenius and Morén, 2001) on

a computational model of the abstracted human

amygdalo-orbitofrontal cortex system. The suggested

controller mimics the natural interplay of actuating

amygdala and inhibiting orbitofrontal cortex. The

implementation of an emotional signal can be

interpreted as a reward or incentive to guide the

system’s learning behaviour. (Yi, 2015) combined

robust sliding mode control with an intelligent control

element comprising an actuator and a preventer

inspired on the mammalian limbic system. (Frank et

al., 2014)’s work focussed on reinforcement learning

allowing an agent to learn a policy with the goal to

maximize a reward-signal. The authors combined a

low-level, reactive controller with a high-level

curious agent. Artificial curiosity contributes to the

learning process by guiding exploration to areas

where the agent can efficiently learn. The work was

validated by a real-time motion planning task on a

humanoid robot. (Merrick, 2012) implemented a

goal-lifecycle and introspection for reinforcement

learning. The aim was to make the system aware of

when to learn what as well as of which acquired skills

to keep either active, ignored or erased.

The aim of this work is to combine freeform

trajectory tracking with force control, i.e. to develop

a model-free control strategy enabling an industrial

robot-arm to follow a desired freeform-path and

simultaneously apply specified adequate joint-

torques at the appropriate moment and position. The

desired position- and force-signals are to be learned

from kinesthetic teaching and introduced as

independent lists of successive joint-angles and -

torques. This work combines elements of PbD,

parallel control and neuro-inspired control-

extensions. Compared with related work cited above,

the main differences in this paper are: 1) no visual

CILAP-Architecture for Simultaneous Position- and Force-Control in Constrained Manufacturing Tasks

245

Figure 1: Inspiration for the CILAP-architecture.

information, i.e. no camera is used and 2) the

considered tasks are continuous movements instead

of discrete contact state formations. The input for the

controller is a list of desired successive joint-specific

positions and torques which are obtained from the

records of a human performing the considered task

operating the lightweight robot arm in gravity

compensation mode. 3) Further, a parallel control

concept composed of a model-free position-controller

and a model-free force-controller is designed. These

separate controllers combine conventional PID- or

PI-control with adaptive neuro-inspired algorithms.

The latter make use of an incentive and a learning

system as well as of the interaction of an actuator and

a preventer to improve the controller performance.

The elements Conventional controller, Incentive,

Learning system and Actuator-Preventer interplay

form the CILAP-concept. The suggested method is

validated on a manufacturing process simulation. For

the control objectives, the main focus is put on

precision. As the considered tasks are not time

critical, minimal mean errors of position- and force-

signals are the objective.

The rest of the paper is structured as follows: It

follows the description of the challenge, i.e. section 2

‘Problem Statement’. Section 3 describes the used

concepts. In Section 4, the suggested parallel control

concept CILAP (conventional-incentive-learning-

actuator-preventer) is developed and in section 5 the

results of the simulation are presented and discussed.

The paper ends with a conclusion.

2 PROBLEM STATEMENT

The robot-arm considered in this work has n links and

its dynamics in the presence of uncertainties,

disturbances and switching constraints are expressed:



(



)





+

(

,



)





+

(



)

=++



+ (1)

with ,



,



∈



link position, velocity and

acceleration with index  for the desired reference

values. 

(



)

∈



is the inertia matrix, 

(

,



)

∈





the centripetal/Coriolis terms, 

(



)

∈



the

gravitational torque-vector. External disturbances are

represented by the bounded term ∈



while

internal uncertainties are implemented as variations

in 

(



)

, (,



) and (). ∈



stands for the

friction between end-effector and environment or

surface. The friction is a function of the applied

torque and the robot link velocity: =













with  a constant factor. 



∈



the global constraint force, 



=



()





()

where ()∈



is the manipulator’s Jacobian, ∈





is the vector of Lagrange multipliers and 



(



)





()



is the gradient of the task space constraints with

∅



(



)

∈



the 

kinematic constraint. ∈



stands for the Cartesian pose and =1,2,…

denotes the index of constraints for the case of

multiple switching constraints with  the total

number of constraints.

The considered application is a contact-based

manufacturing task, i.e. freeform trajectory tracking

with the application of specified forces at specific

positions. Manual work is current state-of-the-art for

these tasks. Not only the fact that these processes

were designed by and for humans but also humans’

capabilities make them the most appropriate

performers for these complex tasks. The challenge in

this work is to mimic the human’s approach to

perform the considered task by translating his

capabilities into robot skills and by including the

worker’s expertise in the control algorithm. The input

for the controller is a .csv-file with desired successive

positions and torques which are obtained from the

records of a human operating a lightweight robot arm

in gravity compensation mode. ∈



is the applied

input torque. It is the sum of the outputs of the pose-

controller and the force-controller, i.e. 



+



. The

control action consists in adapting both robot joint

positions and the applied forces to match the desired

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

246

poses and forces. The goal is to keep the error

between desired and measured signals minimal at all

times, i.e. the mean error signal over the whole

process-period should be minimized.

3 USED CONCEPTS

The concepts used in this work are biologically

inspired, similar to the related work cited in section 1.

Despite the control concepts being inspired on the

functioning of the human brain, they do not attempt

to accurately model its structure. Rather than

presenting a true-to-life computational model of the

mammalian learning behaviour, the aim is to improve

conventional model-free PID-control through the

implementation of neuro-inspired concepts.

An incentive system transforms sensory

information into an incentive, i.e. a reward-based

extrinsic motivational stimulus. Depending on the

environment, the stimulus to the agent, i.e. how to

maximize the reward for the system is changed. This

adaptive incentive then forms the input to a learning

system which feeds both an actuator and a preventer.

The interplay of the latter is inspired on the interplay

of the amygdala and the orbitofrontal cortex in the

mammalian brain during emotional learning. While

the actuator establishes stimulus-action associations,

the preventer erases associations which are no longer

needed. The removal of no longer relevant stimulus-

action association is essential for a successful

learning and to reduce the amount of data in the

system. The latter is similar to the phenomenon of

synaptic plasticity in the human brain. The described

structure is schematically represented in Figure 1.

4 THE PARALLEL

CILAP-ARCHITECTURE

A parallel control concept is developed to

simultaneously control joint angular positions and

torques. The complex constrained control task is

broken down into two independent subsystems. The

suggested concept is composed of a model-free

position-controller and a model-free force-controller.

Both independent controllers consist of a

conventional model-free controller and a model-free

controller extension. The former C combines with the

Incentive-Learning-Actuator-Preventer to form the

CILAP-architecture. The suggested method attempts

to combine robustness, simplicity and intuitiveness

and is depicted in Figure 2.

The input to the controller is a .csv-file, i.e. a list

containing a succession of desired joint positions 



and joint torques 



4.1 Position-Control

The conventional PID-controller-output 





as in

Equation (2)







=







+









+









 (2)

with the error-signal, i.e. the difference between

measured and desired signal as in Equation (3)





=−



(3)

where 



,



and 



are constant gain factors.

For the controller extension, the appreciation, i.e.

the value of the current state is defined as the error-

signal (Equation (4)). As the only way to collect

information about the environment is to interact with

it, a feedback-loop is implemented in this controller-

part.





=−



=



(4)

The reward, i.e. incentive ∈



is defined in

Equation (5).

=(

(





)



∙





)(



−) (5)

where ′ indicates the vector-transpose. ∈



represents the interplay of actuator and preventer. 

is defined as the difference between their respective

outputs (Equations (8) and (9)) which guarantees only

relevant connections are kept. Mimicking synaptic

plasticity, this law allows to limit the number of

active learned connections.

The incentive is the input to the learning system.

Its outputs are the learning rates for both the actuator

(Equation (6)) and the preventer (Equation (7)).





=





∙max

(

,

)

(6)





=





∙(−) (7)

with >0 a constant factor.

The main part of this half of the control algorithm

consists in the interaction between an actuator and an

inhibitor. The actuator-output ∈



and the

preventer-output ∈



are defined in Equations

(8) and (9).

CILAP-Architecture for Simultaneous Position- and Force-Control in Constrained Manufacturing Tasks

247

Figure 2: CILAP-architecture.

=







(8)

=







(9)

The controller-extension-output is defined in

Equation (10), the integration over time mimicking

experience.







=



−



 (10)

with >0 being a constant gain-factor.

The final position-controller output combines the

outputs of the conventional controller and of the

extension.





=





−





(11)

4.2 Force-Control

Parallel to the position-controller, a force-controller

is implemented to make sure the desired force-torques

from the reference .csv-file are applied.

The conventional PI-controller-output 





defined in Equation (12)







=







+









(12)

where 



and 



are constant gain-matrices and the

error-signal





=−



(13)

with =



+.

For the controller-extension the state-value





is defined as the error-signal (Equation (13)).

The incentive is defined as follows

=(

(





)



∙





)(



−) (14)

For the output of the actuator-preventer-system, its

constituents and their learning rates the formulas

defined in Equations (6)-(9) apply. The controller-

extension-output follows Equation (15).







=−





 (15)

The final force-torque-controller output combines

the outputs of the conventional controller and of the

extension.





=





−





(16)

5 RESULTS

5.1 Simulation

The proposed control concept is validated on a

surface finishing task, i.e. the robot-arm successively

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

248

follows desired positions and applies desired torques

at specified positions. The controller is implemented

on a 2D-RR planar robot in the Matlab/Simulink-

environment. The parameters of the robotic arm with

two rotational joints are described in set of Equation

(17).



=





+





















cos(



−



)













cos(



−



)



+











=

0−











sin

(





−



)













sin

(





−



)



=









cos

(





)

+







cos(



)









cos(



)



(17)

with link masses 



=



=1, link lengths 







=1, gravitational acceleration =9.8







distances from the link source end to its centre of

mass 



=



and link moments of inertia 



=



The system-inputs are extracted from .csv-files

containing a succession of desired reference joint

angular positions and torques. The controller-

parameters introduced in Equations (2)-(16) are

chosen and optimized by trial-and-error-procedure as

follows: =5, =30,=1,



=−5,



−20,



=−20,



=5,



=35.

The performance-results are illustrated in Figures

3-8. While Figures 3 and 4 show the trajectory-

tracking performance of the suggested controller

scheme, Figures 5 and 6 depict its velocity-tracking

and Figures 7 and 8 show the force-tracking of link 1

and 2, respectively.

Figure 3: Trajectory tracking of link 1.

Figure 4: Trajectory tracking of link 2.

Figure 5: Velocity tracking of link 1.

Figure 6: Velocity tracking of link 2.

Figure 7: Force tracking of link 1.

CILAP-Architecture for Simultaneous Position- and Force-Control in Constrained Manufacturing Tasks

249

Figure 8: Force tracking of link 2.

5.2 Analysis and Discussion

The simulation-results present good tracking results

for all considered signals: joint torques, angular joint

positions and velocities. This is illustrated

qualitatively in Figures 3-8 as well as quantitatively

in Table 1 and Table 2.

The added value of the controller-extension is

illustrated in Tables 1-3. In Table 3, the improvement

from the parallel controller without the neuro-

inspired extension to the suggested CILAP-

architecture is given in percentages. It is proven that

the latter outperforms a parallel controller with PID-

position- and PI-force-control. The main focus in this

work was put on the mean error signals.

The industrial applications of the presented work

are contact-based manufacturing in general and

surface finishing in specific. Manual work being

current industrial state-of-the-art, surface finishing

processes are the bottleneck of the concerned industry

due to their time- and cost-intensive nature. Different

studies suggest shares of up to 30-50% of the entire

manufacturing time and up to 40% of the total cost.

(Dardouri et al., 2017) (Dieste et al., 2013) (Pagilla

and Yu, 2001) (Robertsson et al., 2006) (Roswell et

al., 2006) (Wilbert et al., 2012). From the scientific

point of view, the automation of these contact-based

manufacturing processes is highly complex as it

requires tackling freeform trajectory tracking and

force control simultaneously while mimicking the

robust and adaptive behaviour a human provides on a

nonlinear system. Additional challenges arising in

industrial practice and taken into account here are the

absence of visual information from a camera and the

application involving continuous movements, e.g.

path following rather than discrete contact state

formations, e.g. gripping.

As the shown work presents promising results for

the automation of the considered continuous

manufacturing processes, future work will involve the

experimental validation on a KUKA LWR4+ robot-

arm with a variety of processes, e.g. grinding or

polishing tasks. Comparisons with state-of-the-art

controllers will be performed to demonstrate the

outperformance of the suggested concept. Also, its

real-time capabilities will be proven.

Table 1: Maximum, minimum and mean absolute positional

errors for both manipulator-links [rad] for the experiment

with (a) the parallel architecture with PID- and PI-

controllers and (b) the CILAP-architecture.

























Max 3.16 0.82 1.18 0.82

Min 9.70e-4 2.25e-5 2.07e-4 2.95e-5

Mean 1.64 0.09 0.13 0.08

Table 2: Maximum, minimum and mean absolute force-

errors for both manipulator-links [N/m

] for the experiment

with (a) the parallel architecture with PID- and PI-

controllers and (b) the CILAP-architecture.

























Max 2.41 2.45 0.05 0.05

Min 0.003 0.001 0.002 1.41e-5

Mean 1.56 1.62 0.02 0.01

Table 3: Average improvement-rate [%] from the parallel

architecture with PID- and PI-controllers to the CILAP-

architecture for position- and force-tracking.

Position-Control 35 %

Force-Control 87 %

6 CONCLUDING REMARKS

In this paper, the control problem of automated

constrained manufacturing tasks was addressed. A

parallel control concept composed of two model-free

controller-halves is developed. One half controls the

position while the other half controls the applied

torque of the robot manipulator performing a

freeform trajectory tracking application with the

application of manufacturing-forces at specified

positions in the presence of uncertainties and friction.

Both controller-halves combine conventional control

with biomimetic adaptive control. The latter is

inspired on the human learning behaviour making use

of an actuator-inhibitor-system and a reward-like

incentive. The elements Conventional controller,

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

250

Incentive, Learning system and Actuator-Preventer

interplay form the CILAP-concept. The developed

model-free control concept combines PbD, parallel

control and neuro-inspired control-extensions. A

surface finishing application-simulation illustrates

the suggested scheme outperforms a combination of

conventional PID- and PI-controllers.

ACKNOWLEDGEMENTS

This work has been done in the framework of the

European Union supported INTERREG GR-project

“ROBOTIX-Academy”. (http://robotix.academy/)

REFERENCES

Abu-Dakka, F., Nemee, B., Jorgensen, J., Savarimuthu, T.,

Krüger, N., Ude, A., 2015. Adaptation of manipulation

skills in physical contact with the environment to

reference force profiles In Autonomous Robots, 39-2,

pp. 199-217.

Adar, N., Kozan, R., 2016. Comparison between Real Time

PID and 2-DOF PID Controller for 6-DOF Robot Arm.

In Acta Physica Polonica A, 130-1, pp. 269-271.

Balkenius, C., Morén, J., 2001. Emotional Learning: a

Computational Model of the Amygdala. In Cybernetics

and Systems, 32-6, pp. 611-636.

Chiaverini, S., Sciavicco, L., 1993. The Parallel Approach

to Force/Position Control of Robotic Manipulators In

IEEE Transactions on Robotics and Automation, 9-4,

pp. 361-373.

Dardouri, F., Abba, G., Seemann, W., 2017. Parallel Robot

Structure Optimizations for a Friction Stir Welding

Application. In ICINCO 2017-14

International

Conference on Informatics in Control, Automation and

Robotics.

Dieste, J., Fernández, A., Roba, D., Gonzalvo, B., Lucas,

P., 2013. Automatic grinding and polishing using

Spherical Robot. In Procedia Engineering, 63, pp. 938–

946.

Ferguene, F., Achour, N., Toumi, R., 2009. Neural network

parallel force/position control of robot manipulators

under environment and robot dynamics uncertainties. In

Archives of Control Sciences, 19-1, pp. 93-121.

Frank, M., Leitner, J. Stollenga, M., Förster, A.,

Schmidhuber, J., 2014. Curiosity driven reinforcement

learning for motion planning on humanoids. In

Frontiers in Neurorobotics, 7, pp. 1-15.

Karayiannidis, Y., Doulgeri, Z., 2010. Robot contact tasks

in the presence of control target distortions. In Robotics

and Autonomous Systems, 58-5, pp. 596-606.

Kuc, T., Han, W., 2000. An adaptive PID learning control

of robot manipulators. In Automatica, 36, pp. 717-725.

Lange, F., Bertleff, W., Suppa, M., 2013. Force and

Trajectory Control of Industrial Robots in Stiff Contact.

In IEEE International Conference on Robotics and

Automation (ICRA), pp. 2927-2934.

Lucas, C., Shahmirzadi, D., Sheikholeslami, N., 2004.

Introducing BELBIC: brain emotional learning based

intelligent controller. In Intelligent Automation and Soft

Computing, 10-1, pp. 11-22.

Merrick, K., 2012. Intrinsic Motivation and Introspection in

Reinforcement Learning. In IEEE Transactions on

Autonomous Mental Development, 4-4, pp. 315-329.

Newman, W., Branicky, M., Podgurski, H., Chhatpar, S.,

Huang, L., Swaminathan, J., Zhang, H., 1999. Force-

responsive robotic assembly of transmission

components. In IEEE International Conference on

Robotics and Automation (ICRA), 3, pp. 2096-2102.

Oba, Y., Yamada, Y., Igarashi, K., Katsura, S., Kakinuma,

Y., 2016. Replication of skilled polishing technique

with serial-parallel mechanism polishing machine In

Precision Engineering, 45, pp. 292-300.

Pagilla, P., Yu, B., 2001. Adaptive control of robotic

surface finishing processes. In Proceedings of the 2001

American Control Conference. (Cat. No.01CH37148),

1, pp. 630–635.

Robertsson, A., Olsson, T., Johansson, R., Blomdell, A.,

Nilsson, K., Haage, M., Lauwers, B., de Baerdemaeker,

H., Brogardh, T., Brantmark, H., 2006. Implementation

of Industrial Robot Force Control Case Study: High

Power Stub Grinding and Deburring. In 2006 IEEE/RSJ

International Conference on Intelligent Robots and

Systems, pp. 2743–2748.

Roswell, A., Jeff, F., Liu, G., 2006. Modelling and analysis

of contact stress for automated polishing. In

International Journal of Machine Tools &

Manufacture, 46, pp. 424–435.

Rozo, L., Jiménez, P., Torras, C., 2013. A robot learning

from demonstration framework to perform force-based

manipulation tasks. In Intelligent Service Robotics, 6,

pp. 33–51.

Siciliano, B., Khatib, O., 2008.

Springer Handbook of

Robotics, Springer-Verlag. Berlin Heidelberg.

Wilbert, A., Behrens, B., Dambon, O., Klocke, F., 2012.

Robot Assisted Manufacturing Systme for High Gloss

Finishing of Steel Molds State of the Art. In ICIRA, 1,

LNAI 7506, pp. 673–685.

Yi, H., 2015. A Sliding Mode Control Using Brain Limbic

System Control Strategy for a Robotic Manipulator. In

International Journal of Advanced Robotic Systems,

12-158.

Yin, Y., Xu, Y., Jiang, Z., Wang, Q., 2012. Tracking and

Understanding Unknown Surface With High Speed by

Force Sensing and Control for Robot In IEEE Sensors

Journal, 12-9, pp. 2910-2916.

CILAP-Architecture for Simultaneous Position- and Force-Control in Constrained Manufacturing Tasks

251