Using Robot Skills for Flexible Reprogramming of Pick Operations in

Industrial Scenarios

Rasmus S. Andersen

, Lazaros Nalpantidis

, Volker Kr¨uger

, Ole Madsen

and Thomas B. Moeslund

Department of Mechanical and Manufacturing Systems, Aalborg University, Aalborg, Denmark

Department of Architecture, Design and Media Technology, Aalborg University, Aalborg, Denmark

Keywords:

Robot Vision, Robotic Skills, Industrial Robots, Tabletop Object Detector.

Abstract:

Traditional robots used in manufacturing are very efﬁcient for solving speciﬁc tasks that are repeated many

times. The robots are, however, difﬁcult to (re-)conﬁgure and (re-)program. This can often only be done by

expert robotic programmers, computer vision experts, etc., and it requires additionally lots of time. In this

paper we present and use a skill based framework for robotic programming. In this framework, we develop a

ﬂexible pick skill, that can easily be reprogrammed to solve new speciﬁc tasks, even by non-experts. Using

the pick skill, a robot can detect rotational symmetric objects on tabletops and pick them up in a user-speciﬁed

manner. The programming itself is primarily done through kinesthetic teaching. We show that the skill

has robustness towards the location and shape of the object to pick, and that objects from a real industrial

production line can be handled. Also, preliminary tests indicate that non-expert users can learn to use the skill

after only a short introduction.

1 INTRODUCTION

In modern manufacturing industries automation us-

ing robots is very widespread, and automation has for

several decades proven its potential to increase pro-

ductivity many times. Still, areas exist where automa-

tion has not gained the same foothold. This is for

instance the case in small to medium sized compa-

nies, where investment in automated production lines

can be too big a risk. Also in large companies, au-

tomated production lines are typically required to run

for several years to justify the investment. This can

be a problem in shifting markets where consumer de-

mand cannot always be predicted accurately, even for

few years into the future. Especially for newproducts,

there is a need for gradually increasing production, in-

stead of constructing fully automated production lines

all at once.

To make industrial robots better suited for these

scenarios, it has been identiﬁed that improved con-

ﬁguration options and human-robot interaction are

core requirements (EUROP, 2009). In short, making

robots more ﬂexible and easier to reconﬁgure and re-

program is a necessity for future manufacturing. In

this paper we address the issue of making robot pro-

gramming fast and ﬂexible by taking advantage of a

skill-based framework, which is described in Section

2.1.

The contribution of this paper is to show how such

a skill-based framework can be used to develop a ﬂex-

ible vision-based pick skill, which fast and easily can

be reprogrammed by non-experts. The developedpick

skill uses a depth sensor to locate rotational symmet-

ric objects, and then picks them up using a small num-

ber of taught parameters.

The paper is organized as follows: In Section 2.1

the concept of robotic skills is introduced. Both re-

lated research and the interpretation used here is pre-

sented. In Section 2.2 basic methods for performing

object detection is described, and on this basis the ap-

proach taken here is presented.

In Section 3 the complete proposed system is de-

scribed, consisting of a robot system (Section 3.1), a

tabletop object detector (Section 3.2), and an integra-

tion of this into a pick skill (Section 3.3). Experimen-

tal results are presented in Section 4; ﬁrst in Sections

4.1 and 4.2 with regards to robustness against varia-

tions in position and shape of the object to pick. In

Section 4.3 the skill is tested on parts from an indus-

trial production line from the Danish pump manufac-

turer Grundfos A/S, and in Section 4.3 reprogram-

ming (teaching) of the pick skill is tested on non-

expert users. Conclusions are ﬁnally drawn in Section

678

S. Andersen R., Nalpantidis L., Krüger V., Madsen O. and Moeslund T..

Using Robot Skills for Flexible Reprogramming of Pick Operations in Industrial Scenarios.

DOI: 10.5220/0004752306780685

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 678-685

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2 BASIC CONCEPTS AND

RELATED WORK

The central concepts for the pick skill presented in

this paper are robotic skills and object detection.

These are covered here.

2.1 Robotic Skills

The concept of robotic skills it not new, and a sig-

niﬁcant amount of literature has attempted to deﬁne

the concept in the most generic, useful way. The pur-

pose of a robotic skill is to encapsulate complicated

knowledge and present itself to a non-expert user in a

way that allows the user to exploit this skill to make

the robot perform new, repeatable operations on ob-

jects. One of the ﬁrst to draw attention to this idea are

(Fikes and Nilsson, 1972) with their STRIPS planner.

Their focus is to make it possible to perform auto-

matic planning of how to get from an initial state to

a goal state by combining a set of simple actions or

skills. In (Archibald and Petriu, 1993), the idea of

skills is further generalized in a framework known as

SKORP. Here, the focus is both on what a skill should

consist of, and also on developing a (at the time) mod-

ern and user friendly interface to enable workers to

program robots faster.

A broader view is taken in (Gat, 1998), where

the complete planning and control structure of robots

is treated. It is argued that a traditional sense-plan-

act (SPA) architecture is insufﬁcient for robots solv-

ing problems in a complex scenario, and he suggests

instead a three-layered architecture, where all layers

hold their own SPA structure. In the lowest layer, real-

time control is placed. In the highest layer, time inde-

pendent processes are run, that are related to the over-

all task of the robot. The middle layer, the Sequencer,

continuously changes the behavior of the lowest level,

in order to complete a goal. The functionality in this

level can be interpreted as a type of skills, which en-

capsulates the functionality of the lower levels, and

works towards a goal by sequencing hardware-near

actions.

More recently, in (Mae et al., 2011) the three lay-

ered structure is augmented by yet another layer on

top, which contains a high level scenario description

presented in a user friendly interface. This is aimed at

solving manipulation tasks for service robots. The fo-

cus here is also on the architecture, where skills con-

stitute the third layer; just above the hardware control

layer. Another approach to skills are taken at Lund

University, where skills are seen as a means to reuse

functionality across multiple platforms (Bjorkelund

et al., 2011).

We interpret a skill as an object-centered ability,

which can easily be parameterized by a non-expert.

Rather than attempting to develop a large architec-

ture, able to handle general problems in real world,

we focus on making skills very easy and fast to re-

conﬁgure by non-expert humans. The focus is there-

fore especially on the human-robotinteraction, and on

the reusability of skills within the domain of tasks that

are frequently required by the manufacturingindustry.

In its essence, a skill in our framework consists of a

teaching part and an execution phase, as illustrated in

Figure 1 (slightly modiﬁed from (Bøgh et al., 2012)).

Figure 1: General structure of a skill. A skill consists of

a teaching phase (top) and an execution phase (bottom).

Both during teaching and execution, the skill transforms the

world state in a way, which allows other skills to proceed.

The execution block in the ﬁgure captures the

ability of the robot to perform a task. The remain-

ing blocks is what makes this ability into a skill.

The teaching phase is what makes the skill repro-

grammable. Here, the user speciﬁes all parameters,

that transform the skill from a generic template into

something, which performs a useful operation. The

phase is divided into an ofﬂine speciﬁcation part and

an online teaching part. The speciﬁcation is typically

done while selecting to use this skill to solve (part of)

a task. This can be done using e.g. a computer or a

tablet. In the online teaching most of the parameters

are speciﬁed, and this can be carried out using a robot

controller, kinesthetic teaching, human demonstration

etc.

The execution phase consists of the execution

block combined with pre- and postcondition checks

as well as prediction and continuous evaluation. The

precondition check determines if the world state lives

up to the requirements of the skill. For a pick oper-

ation this can for instance include checking whether

the gripper is empty, whether the robot is properly cal-

ibrated, etc. If this check is passed execution can be-

gin. During ongoing execution continuous evaluation

ensures that the skill is executed as expected. When

the execution has ﬁnished, a postcondition check de-

termines if the current world state is as predicted. To-

UsingRobotSkillsforFlexibleReprogrammingofPickOperationsinIndustrialScenarios

679

gether, the pre- and postcondition checks of a skill

makes it ﬁt with other skills, and thus allows it to be a

piece in solving larger tasks.

We believe that a framework based on skills has

the potential to increase the speed and ease of the

way humans interact with industrial robots. In this

paper, one such skill based on a tabletop object detec-

tor is proposed, and in Section 3.3 in particular it is

described how it is realized as a skill.

2.2 Object Detection

Object detection, recognition and pose estimation

making it possible to pick certain objects are some of

the most fundamental problems to solve in robot vi-

sion, and many different approaches have been taken.

(Klank et al., 2009) describe some of the most com-

mon methods. A depth sensor is used to detect a plane

surface, and segment point clusters supported by this.

An RGB camera then captures an image, and CAD

models are ﬁtted to the potential objects inside the re-

gions given by the depth sensor. This provides poses

of the objects. Good gripping points have been spec-

iﬁed in advance in accordance with the CAD model,

and one of these is used to pick the object.

Speciﬁcally for detecting objects on a surface, the

ﬁeld has matured enough to methods are available,

e.g. through ROS. In general, these systems how-

ever rely on either predeﬁned CAD models or require

models to be learned before being used. This is a

complicated and often time consuming step, which

is necessary for some objects, but not for all. It is

not always necessary to have a detailed model to be

able to pick, place and handle simple objects. Speciﬁ-

cally for rotational symmetric objects, which are very

common in the industry, a cylindrical model is usu-

ally sufﬁcient for determining their position precisely.

Therefore, it is here chosen to use a containing cylin-

der model to estimate the position of all objects. The

cylinder model is containing in the sense, that the en-

tire object is inside the model. Therefore, the object

can be grasped by opening the gripper more than the

diameter of the cylinder, moving the gripper to the

center point of the cylinder, and closing the gripper

until the force sensors detect contact with the object.

For complicated objects this approach can obviously

cause problems, but for many objects encountered

in industrial manufacturing, it will work satisfactory

without the need for object-speciﬁc CAD models. At

the same time, not using models simpliﬁes the process

signiﬁcantly to human users. A more complicated ob-

ject detector can then be used in situations where the

object is very unsymmetrical.

3 PROPOSED SYSTEM

The proposed system consists of a robot system, a

tabletop object detector, and the integration of this de-

tector into a skill-based structure. These are described

in the following subsections.

3.1 Robot System

The robot system used here is called “Little Helper”

(Hvilshøj and Bøgh, 2011; Hvilshøj et al., 2009), and

it is shown in Figure 2. It consists of a robot arm, a

gripper, and has here a depth sensor mounted on its

end effector.

Figure 2: The robot system consists of a robot arm, gripper,

depth sensor, tabletop, and an object to pick.

In the ﬁgure, the robot is about to grasp an object

placed on a tabletop. The tabletop considered in this

work is part of the robotic platform, although any sur-

face within reach of the robot could be used.

The robot arm is a KUKA LWR, which has 7 de-

grees of freedom and supports force control. The

gripper is a traditional parallel gripper. The depth sen-

sor is a PrimeSense Carmine 1.09. This is similar to

the Microsoft Kinect; only smaller and it functions at

distances down to 35 cm. Is has both a RGB camera

and a depth sensor, but here only the depth sensor is

used. This captures the point clouds used for the table

top object detector.

3.2 Tabletop Object Detector

The tabletop object detector described here has been

developed speciﬁcally for this system. However, the

type of object detector used is not essential for the

skill-based approach. Thus, any available object de-

tector could in principle be used.

The tabletop object detector takes as input a point

cloud, and outputs the objects which are positioned

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

680

on a supporting plane. The object models used as

containing cylinders, in the sense that they enclose all

points belonging to each object. The center position

and diameter of these cylinders can be used directly

for a pick operation. Figure 3 shows an example of a

segmented point cloud.

Figure 3: The segmented point cloud from the tabletop ob-

ject detector seen from above. Points belonging to the de-

tected plane are red, points belonging to the detected object

are pink, and a green containing cylinder is ﬁtted around the

detected object.

The main steps in the object detector are shown in

Figure 4. The ﬁrst step is to actually ﬁnd the dominant

plane in the scene, i.e. the tabletop. This is done using

RANSAC (Fischler and Bolles, 1981), followed by

a least squares ﬁtting to all inliers. The inliers are

colored red in Figure 3.

Figure 4: The main steps in the object detector.

The plane model is often supported by both the

actual tabletop as well as other objects. Therefore, all

inliers are clustered, and the largest cluster is chosen

as the tabletop. Then, a convex hull is ﬁtted around

the tabletop, and it is extended along the norm of the

plane to give a 3D volume. This volume is illustrated

in Figure 3 as gray borders, and all objects supported

by the plane must be within this. The points inside

the volume are clustered, and very small objects as

well as objects “hovering” over the plane are ignored.

Cylinder models orthogonal to the detected plane are

ﬁtted around the rest. In Figure 3 one cylinder model

has been ﬁtted. The cylinders are containing in the

sense that all points are inside or exactly on the bor-

ders of the model.

Finally, each cylinder model is checked for valid-

ity. Cylinders with a center very close to the border

of the table are removed along with cylinders whose

border is partly outside of the plane. This serves both

to remove false objects, and also to remove objects

that are too close to the border to be safely grasped.

3.3 Skill Realization

The purpose of the object detector is to be the core of a

pick skill, which can easily be reprogrammed to solve

new speciﬁc pick operations. This requires a number

of other functions, as indicated in Figure 1. The most

essential is which parameters it will be advantageous

for the user to be able to specify and how to do this

most efﬁciently. For this pick skill the most essential

parameters considered are:

• The camera pose used for detecting objects.

• The orientation of the gripper used for grasping

the object.

• The vectors used for approaching and leaving the

grasping position.

Additionally, it is chosen to include the following pa-

rameters related to safety and robustness:

• The velocity of the robot.

• A via position for the robot, to help the robot

avoid obstacles in the scene when approaching the

object.

• The height of the gripper when grasping the ob-

ject.

• The approximate diameter of the object to pick.

This can be used to verify that the robot has

grabbed the correct object.

The velocity of the robot can be provided most

precisely by the user in the ofﬂine speciﬁcation. The

remaining parameters are provided through kines-

thetic teaching; that is, by the human user manually

moving the robot around. The sequence is illustrated

in Figure 5.

UsingRobotSkillsforFlexibleReprogrammingofPickOperationsinIndustrialScenarios

681

Figure 5: Teaching and execution phase of the pick skill.

In the ofﬂine speciﬁcation, the velocity of the robot is spec-

iﬁed. It is also speciﬁed whether the same approach and

leaving vector will be used. If that is the case, the “ap-

proach pose” is skipped during online teaching. All orange

blocks in the teaching phase is input by the user through

kinesthetic teaching.

During teaching of the camera pose the depth im-

age from the camera is shown to the user. This al-

lows him/her to precisely position the camera, so that

the entire desired tabletop is in view. The pre- and

postcondition checks verify that the gripper is initially

empty and ends up holding an object of the approxi-

mate correct size (diameter). The continuous evalua-

tion (see Figure 1) makes sure that the skill execution

proceeds as planned, including that valid object(s) are

detected.

One important feature to note for any skill is that

the ending state must be the same for the teaching and

execution phases. The same applies for the require-

ments for the starting state. This makes it possible

to perform continuous teaching of several speciﬁed

skills in a row. Since a successful execution of this

pick skill will result in the gripper holding an object,

the teaching phase of the pick skill is ordered to en-

sure, that the gripper also here ends up holding an ob-

ject. This allows the user to continue teaching a skill

that expects the gripper to hold an object - such as a

place skill.

4 EXPERIMENTAL RESULTS

The pick skill was tested in a number of ways which is

described in the following. First the robustness of the

skill was tested with regards to variations in position

and shape compared to the shape used for teaching.

Then the skill is tested on industrial parts from a pump

production line. Finally, the teaching part of the skill

was tested on non-expert users. The objects used for

the tests are shown in Figure 6. Where nothing else is

speciﬁed, the perfect cylinder in Figure 6(a) was used.

(a) Cylindrical and deform objects. (b) Pump parts from an industrial

production line.

Figure 6: Objects used for pick tests.

4.1 Variation in Position

During the teaching phase the object was placed in

the middle of the tabletop shown in Figure 2, and the

camera position was taught so that the object was in

the center of the image. In a realistic scenario, the

position of the object cannot be expected to always

be the same during execution. To test the ability of

the skill to handle this, the execution was executed 33

times. Each time, the object was placed between 0

and 20 cm from the teaching position. The results are

shown in Figure 7.

−20 −10 0 10 20

−25

−20

−15

−10

−5

Relative distance from camera [cm]

Relative transverse distance [cm]

Successes

Failures

Figure 7: Results for pick with variation in position rela-

tive to position used during teaching. The ﬁgure shows the

tabletop from above, and the camera position was to the left.

The pick operation succeeded for all positions with varia-

tions up to 15 cm. For errors of 20 cm, the pick operation

failed in 3 of 8 cases.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

682

Of the 33 executions, the robot succeeded in pick-

ing up the object 30 times and failed 3 times. This

corresponds to a success rate of 91%. Notably, all ex-

ecutions with a position deviation of 15 cm and below

succeeded, while all errors occurred at a deviation of

20 cm. Visual inspection showed that the two errors

at (0,-20) and (0,20) were caused by the object being

very close to the border of the plane (see Figure 7).

This caused the object detector to disregard the point

clusters as valid objects. The last error was caused by

an imprecise position estimation.

The time for each pick operation was on average

19.5 seconds. This includes 2.4 seconds on average

for object detection.

4.2 Variation in Shape

The object detector and pick skill were designed for

rotational symmetric objects, as previously described.

It has, however, some robustness to variations in

shape. This robustness is tested in this section using

the increasingly deformed cylinders shown in Figure

6(a).

The pick skill was ﬁrst taught using the perfect

cylinder shown to the left in Figure 6(a). Then, ex-

ecution was carried out using increasingly deformed

cylinders. It was attempted to pick each of the cylin-

ders three times: Once seen from the wide side, once

seen from the narrow side, and once seen in between

(skewed).

In Figure 8 the most deformed cylinder is seen

from the three different angles. The object detec-

tor has in all cases attempted to ﬁt a cylinder model

around the detected point cloud. For the wide and the

skewed view, the model is ﬁtted correctly. For the nar-

row view in Figure 8(b) the object detector has failed

to classify all points from the cylinder together. In re-

ality, both the green points inside the cylinder model

as well as the white points behind the cylinder model

come from the object. However, only the green points

were used. The reason is, that the depth sensor is poor

at detecting points on the top border of the physical

cylinder as well as inside its narrow hole. Therefore,

a gap arises in the point cloud between the front and

back sides of the cylinder. Unfortunately, the gap is

too large to allow it to be bridged by the clustering

algorithm of the object detector. The effect is that the

position of the object is estimated imprecisely.

Table 1 summarizes the tests of the deformed

cylinders. The pick operation succeeded in all sit-

uations; even for the erroneously ﬁtted narrow view

orientation; in the sense, that the gripper held the ob-

ject when execution had ﬁnished. In two situations

the position of the object in the gripper was a bit off.

(a) Wide. (b) Narrow. (c) Skewed.

Figure 8: Rotations of the most deformed cylinder from

Figure 6(a). For the wide and skewed orientations, a con-

taining cylinder is ﬁtted successfully. For the narrow ori-

entation, the containing cylinder is only ﬁtted to the frontal

part of the physical cylinder.

The ﬁrst of these situations, cylinder 4 seen from nar-

row side, is the one illustrated in Figure 8(b), and the

reason is the imprecise position estimate. The second

is the one illustrated in Figure 8(a). Here, the position

estimate is relatively accurate. However, the position

where the gripper grasps the object was taught, so that

the object here was skewed in relation to the gripper.

When it was grasped, it “slipped” into place in the

gripper.

Table 1: Test results for picking the deformed cylinders

from Figure 6(a). Only one result is given for cylinder 1,

since its shape does not change with orientation. It was

possible to grasp all cylinders with all tested orientations.

In the situations marked as (+), the object was grasped, but

its position in the gripper was not as expected.

No.

Diameter Seen from

Wide Narrow Narrow Wide Skewed

1 4.9 cm 4.9 cm +

5.4 cm 4.4 cm + + +

5.9 cm 3.8 cm + + +

6.9 cm 3.1 cm (+) (+) +

4.3 Application Speciﬁc Tests

The pick skill presented here was designed to function

in industrial scenarios with industrial, rotational sym-

metric objects. Therefore it was tested on the indus-

trial objects shown in Figure 6. All three objects are

from a pump manufacturing line at the Danish pump

manufacturer Grundfos A/S. From left, the objects are

respectively a rotor cap, a rotor core, and an assem-

bled rotor.

Obviously, the optimal place to grasp these ob-

jects is not the same. Therefore the pick skill was

taught/parameterized speciﬁcally for each of the ob-

jects. Using these, the robot was able to successfully

UsingRobotSkillsforFlexibleReprogrammingofPickOperationsinIndustrialScenarios

683

pick each of the objects three times in three attempts.

For the teaching phase of the skill, a preliminary

test was carried out to verify that it is possible for hu-

mans to learn to use the skill fast. Two users with no

previous knowledge about the pick skill were asked

to teach the skill after receiving only a short intro-

duction to the skill (∼5 min). The users succeeded

to teach the skill in the ﬁrst attempt, and the teaching

were in all cases completed in less than 4 minutes.

The parameterized pick skills could afterwards be ex-

ecuted. In one case the parameterized pick skill only

functioned in a very limited area, because the robot

otherwise was asked to move outside its limits after

grasping the object. This was, however, easily visible

and could if desired be corrected by re-teaching the

skill.

5 CONCLUSIONS

In this paper, our interpretation of a ﬂexible, robotic

skill has been presented. We see a skill as an object

centered ability, which encapsulates advanced func-

tionality in a way that allows a non-expert user to eas-

ily program the robot to perform a new task. In this

framework, a pick skill has been developed. The pur-

pose of the skill is to make it possible for a robot to

pick up rotational symmetric objects by using a depth

sensor for detection. The skill features both a teach-

ing and an execution phase. In the teaching phase,

the user parameterizes the skill, mainly though kines-

thetic teaching. In the execution phase, the taught pa-

rameters are used to pick up objects.

The skill uses an object detector which is speciﬁ-

cally designed to detect the position of rotational sym-

metric objects, which it does sufﬁciently accurate to

pick up the objects. Also, experiments have shown

that the skill has robustness to deviations in position

and shape. With regards to position, our robot was

able to pick up 25 of 25 objects placed up to 15 cm of

the location used for teaching the skill. At a distance

of 20 cm it failed for 3 of 8 objects, mainly because

the objects were very close to the border of the table.

With regards to variation in shape, it was possible to

pick up objects with different shapes, including a de-

formed cylinder with a diameter of more than double

on one side compared to the other. Moreover, it was

possible to pick up rotational symmetric objects from

a real production line at Grundfos A/S.

Finally, the teaching phase of the skill was tested

on users with no experience with the skill. With min-

imal introduction, the users were able to complete the

teaching phase in less than 4 minutes.

We believe that with further development of func-

tionality within a ﬂexible skill based structure, there

is a potential for robots to perform tasks, which it has

previously not been proﬁtable to automate. This is es-

pecially the case in small companies and in industries

that manufacture products for rapidly changing mar-

kets. A skill-based approach to robot programming

can make it possible for non-expert users to perform

fast reprogramming of the robots to perform required

tasks when it is required, without the need to call in

robot programmers and other experts.

The pick skill presented in this paper is very easy

to use, and it performs well in many scenarios. The

largest restriction is perhaps that it cannot be guar-

antied to perform well with rotational asymmetric ob-

jects. The future research plan therefore includes in-

vestigation of how a generalization can best be in-

tegrated without complicating the teaching interface

signiﬁcantly to the user.

ACKNOWLEDGEMENTS

This research was partially funded by the European

Union under the Seventh Framework Programme

project 260026 TAPAS - Robotics-enabled logistics

and assistive services for the transformable factory of

the future.

REFERENCES

Archibald, C. and Petriu, E. (1993). Model for skills-

oriented robot programming (skorp). In Optical En-

gineering and Photonics in Aerospace Sensing, pages

392–402. International Society for Optics and Photon-

ics.

Bøgh, S., Nielsen, O., Pedersen, M., Kr¨uger, V., and Mad-

sen, O. (2012). Does your robot have skills? In

Proceedings of the 43rd International Symposium on

Robotics.

Bjorkelund, A., Edstrom, L., Haage, M., Malec, J., Nils-

son, K., Nugues, P., Robertz, S. G., Storkle, D.,

Blomdell, A., Johansson, R., Linderoth, M., Nilsson,

A., Robertsson, A., Stolt, A., and Bruyninckx, H.

(2011). On the integration of skilled robot motions for

productivity in manufacturing. In Assembly and Man-

ufacturing, IEEE International Symposium on, pages

1–9.

EUROP (2009). Robotics visions to 2020 and beyond -

the strategic research agenda for robotics in europe

(SRA). Technical report, EUROP.

Fikes, R. E. and Nilsson, N. J. (1972). Strips: A new ap-

proach to the application of theorem proving to prob-

lem solving. Artiﬁcial intelligence, 2(3):189–208.

Fischler, M. A. and Bolles, R. C. (1981). Random sample

consensus: A paradigm for model ﬁtting with appli-

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

684

cations to image analysis and automated cartography.

Communications of the ACM, 24(6):381–395.

Gat, E. (1998). On three-layer architectures. In Ko-

rtenkamp, D., Bonnasso, P. R., and Murphy, R., ed-

itors, Artiﬁcial Intelligence and Mobile Robots, pages

195–210.

Hvilshøj, M. and Bøgh, S. (2011). ”Little Helper” - an

autonomous industrial mobile manipulator concept.

International Journal of Advanced Robotic Systems,

8(2).

Hvilshøj, M., Bøgh, S., Madsen, O., and Kristiansen, M.

(2009). The mobile robot ”Little Helper”: Concepts,

ideas and working principles. In ETFA, pages 1–4.

Klank, U., Pangercic, D., Rusu, R. B., and Beetz, M.

(2009). Real-time CAD model matching for mobile

manipulation and grasping. In Humanoid Robots,

2009. Humanoids 2009. 9th IEEE-RAS International

Conference on, pages 290–296. IEEE.

Mae, Y., Takahashi, H., Ohara, K., Takubo, T., and Arai,

T. (2011). Component-based robot system design for

grasping tasks. Intelligent Service Robotics, 4(1):91–

98.

UsingRobotSkillsforFlexibleReprogrammingofPickOperationsinIndustrialScenarios

685