High-level robot programming based on CAD: 
dealing with unpredictable environments 
 
 
1Pedro Neto, 1Nuno Mendes, 1Ricardo Araújo, 1J. Norberto Pires 
2A. Paulo Moreira 
 
1Department of Mechanical Engineering (CEMUC), University of Coimbra, Coimbra, Portugal 
2Institute for Systems and Computer Engineering of Porto (INESC-Porto), Porto, Portugal 
 
 
Abstract: 
Purpose – The purpose of this paper is to present a CAD-based human-robot interface 
that allows non-expert users to teach a robot in a manner similar to that used by humans 
to teach each other. Another important issue addressed here has to do with how robots 
deal with uncertainty.  
 
Design/methodology/approach – Intuitive robot programming is achieved by using 
CAD drawings to generate robot programs off-line. Sensory feedback allows 
minimization of the effects of uncertainty, providing information to adjust the robot 
paths during robot operation. 
 
Findings – It was found that it is possible to generate a robot program from a common 
CAD drawing and run it without any major concerns about calibration or CAD model 
accuracy. 
 
Research limitations/implications – A limitation of the proposed system has to do 
with the fact that it was designed to be used for particular technological applications. 
 
Practical implications – Since today most of the manufacturing companies have CAD 
packages in their facilities, CAD-based robot programming may be a good option to 
program robots without the need for skilled robot programmers.  
 
Originality/value – A new CAD-based robot programming system is proposed. Robot 
programs are directly generated from a CAD drawing “running” on a commonly 
available 3D CAD package and not from a commercial computer aided robotics 
software, making it a simple CAD integrated solution. This is a low-cost and low-setup 
time system where no advanced robot programming skills are required to operate it.  
 
Keywords: CAD, Industrial Robotics, High-Level Programming, Sensory Feedback, 
Unpredictable Environments. 
 
 
 
 
 
 
 
 
 
 
 
 
 
1. Introduction 
1.1. Motivation 
Increasingly, companies are changing and reinventing their production systems. 
Traditional manufacturing systems (often based on fixed automation and manual work) 
are being replaced by flexible and intelligent manufacturing systems, enabling 
companies to continue to be competitive in the global market (Kopacek, 1999). This 
competitiveness is reflected in the companies’ capacity to respond/react quickly to 
market demands, producing more and better quality products at competitive prices. 
Owing to its flexibility, programmability and efficiency, industrial robots are seen as 
a fundamental element of modern flexible manufacturing systems. Nevertheless, there 
are still some problems that hinder the utilization of robots in industry, especially in 
small and medium-sized enterprises (SMEs). SMEs have difficulty finding skilled 
workers capable of operating with robots. Therefore, new and more intuitive ways for 
people to interact with robots are required to make robot programming more accessible, 
easier and faster. The goal is that the instructor can teach a robot in a manner similar to 
that used by humans to teach each other, for example using CAD drawings, gestures or 
through verbal explanation (Neto et al., 2010a). 
 
1.2. Objectives 
Robot programming through the typical teaching method (using the teach pendant) is a 
tedious and time-consuming task that requires technical expertise. The goal is to 
develop methodologies that help users to program a robot in an intuitive way, quickly, 
with a high-level of abstraction from the robot specific language, and, if possible, 
without speeding too much money. 
In this paper, a CAD-based system to program a robot from a 3D CAD drawing, 
allowing users with basic skills in CAD and robot programming to generate robot 
programs off-line, is presented. In addition, the 3D CAD package, Autodesk Inventor, 
that interfaces with the user is a well-known generic CAD package, widespread on the 
market at a relative low-cost. Starting from the CAD model of the robotic cell in study, 
the way the user generates a robot program is as simple as “drawing” the desired robot 
paths in the CAD environment. Later, the information needed is automatically extracted 
from the CAD environment, analysed and converted into robot programs. Note that the 
robot programs are not extracted neither from a computer aided manufacturing (CAM) 
software nor from a computer aided robotics (CAR) software. It means that we are 
proposing a simple CAD integrated solution for the robotics field. 
Commercial CAR packages are powerful tools, which enable modelling, simulation 
and robot programming. Nevertheless, they have some disadvantages that hinder their 
use in companies, especially in SMEs. By comparing commercial CAR packages with a 
CAD-based robot programming system similar to that presented in this paper (Neto et 
al., 2010b), it was found that the CAD-based system has some relative advantages: 
 
• Low-cost. Since the construction of CAD models and the robot programming task 
are performed in the same environment/platform (Autodesk Inventor) the programming 
task becomes easier and cheaper; 
• Short learning curve; 
• Simplicity of use. The most time consuming task, the construction of the CAD 
model, is present in both systems.  
 
CAD-based robot programming approaches work well if the environment of the 
robot tasks is well defined. In the other hand, there are situations which are likely to 
create errors or impede the normal operation of the robot: 
 
• The CAD models do not reproduce correctly the geometry of the real scenario; 
• Inaccuracies created in the robot calibration process; 
• Inefficient fixtures that do not ensure the static character of the workpieces; 
• A “foreign” object is introduced in the real environment. 
  
In these cases, we can say that we are in the presence of a dynamic and unpredictable 
environment.  
To perform successful manipulation robots depend on precise information about 
objects in their surrounding. In an unpredictable environment, such information cannot 
be given to the robot a priori, robots have to autonomously and continuously acquire 
information about their surrounding environment to support their decision making and 
react to unanticipated events. Sensory feedback allows a robot to recognize your work 
environment for itself, for example producing corrections (on-line) in pre-programmed 
robot paths (Figure 1). In fact, the integration of sensors into robotic platforms reduces 
the setup time, the need for accurate robot trajectory programming and promotes 
flexibility and the autonomous behaviour of robotic systems (Bolmsjö and Olsson, 
2005; Johansson et al., 2004).  
 
P1
P1
P1
P2
P2
P2
Planned 
path
a)
b)
c)
Robot path
Robot 
path
Collision
 
Figure 1 (a) – planned path for a specific environment; (b) – a “foreign” object is introduced into the 
environment and collision occurs; (c) – sensory feedback is introduced, helping the robot to deal with the 
unpredictable environment (robot path is adjusted) 
 
We validate our methods with two different real-world experiments for two different 
tasks, seam tracking and for applications that require the robot to follow a geometric 
profile while maintaining a contact force. 
 
2. Related Work  
In recent years, CAD technology has become economically attractive and easy to work 
with so that today millions of SMEs worldwide are using it to design and model their 
products. Already in the 80’s, CAD was seen as a technology that could help in the 
development of robotics (Bhanu, 1987). Since then, a variety of research has been 
conducted in the field of CAD-based robot planning and programming. 
Pires et al. (2004) proposes to extract robot motion information from a CAD DXF 
file and converting it into robot commands for welding purposes. A review of CAD-
based robot path planning for spray painting is presented by Chen et al. (2009). Another 
study presents a method to generate 3D robot working paths for a robotic adhesive 
spray system for shoe outsoles and uppers (Kim, 2004). Nagata et al. (2007) proposes a 
robotic sanding platform where the robot paths are generated by CAD/CAM software. 
An example of a novel process that benefits from the robots and CAD versatility is the 
so-called incremental forming process of metal sheets (Schaefer and Schraft, 2005). 
Feng-yun and Tian-sheng, (2005) presents a robot path generator for the polishing 
process, where the cutter location data is generated from the postprocessor of a CAD 
system. As we have seen above, a variety of research has been done in the area of CAD-
based robot planning and programming. However, none of the studies so far deals with 
a “global” solution for this problem.  
Unpredictable environments pose a significant challenge because of their complexity 
and inherent uncertainty. Over the last few years, important studies have been carried 
out to deal with uncertainty in the robotics field: using models of “ideal” environments, 
sensory feedback, and implementing reasoning methods into robotic platforms 
(Bruyninckx et al., 1991; Nayak and Ray, 1990). These concepts have evolved and 
recently, researchers have been successful in developing skills that can handle the 
complexity of dynamic and predictable environments (Kenney et al. 2009; Mendes et 
al., 2010). A number of authors have devoted attention to sensor simulation, trying to 
mimic as closely as possible the behaviour of a real sensor, and thus integrating it (the 
virtual sensor) within a CAR platform (Cederberg et al., 2002; Brink et al., 1997; 
Bolmsjö and Olsson, 2005). Moreover, sensor information has been used to update 
robotic cell models in real-time, allowing to avoid problems such as collisions, 
kinematic singularities and exceeding of joint limits (Brink et al., 1997; Johansson et 
al., 2004). 
 
The concept of seam tracking applied to robotic welding has been studied over the 
last two decades (Nayak and Ray, 1990). Recently, important work has been carried out 
in the integration of sensors to assist the robotic arc welding process (Fridenfalk and 
Bolmsjö, 2002; Bolmsjö and Olsson, 2005). 
 
3. Robot Programming from CAD  
Starting from a 3D CAD model of the robotic cell in study, the way the user generates a 
robot program can be as simple as “drawing” the desired robot paths in the CAD 
environment. Furthermore, to define the robot end-effector pose (position and 
orientation), it is necessary to know, not only the robot path positions but also the end-
effector orientations in space. Therefore, after drawing the robot paths, simplified tool 
models should be placed along the paths. These models will define the orientation of the 
robot end-effector in each segment of the path (Figure 2). 
 
Xn-1
Xn
Xn+1
Yn-1
Zn
Yn
Yn+1
Zn+1
Zn-1
Simplified tool model
Robot path
 
Figure 2 simplified tool models defining the end-effector orientation 
 
 The information needed to program the robot will be extracted from the CAD 
environment by using an application programming interface (API) provided by 
Autodesk. This API allows the extraction of the points that characterize each of the 
different lines used to define a robot path; straight lines, splines and arcs. Moreover, the 
API also gives information about the transformation matrix of each part model 
represented in the CAD environment. The transformation matrix contains the rotation 
matrix and the position of the origin of the part model to which it refers, both in relation 
to the origin of the CAD assembly model. Later, the information extracted from the 
CAD is converted into robot programs (Video 1, 2010). A diagram with the procedure 
to extract 3D data from CAD and their conversion into a robot program is presented in 
Figure 3. 
 
Start
Open a 3D CAD Drawing
Draw Robot Paths
Select Robot Paths
Extract Data from CAD
(Position and Orientation)
Define Robot Tool Orientation
Define Robot Parameters
Generate Robot Program
End
Execute Interface Software
 
Figure 3 extracting 3D data from CAD  
 
3.1. Application programming interface 
The Autodesk Inventor API shows the Inventor’s functionalities in an object-oriented 
manner, allowing developers to interact with Autodesk Inventor using current 
programming languages; Visual Basic, Visual C#, Visual C++. In our proposed system, 
a standalone application was used to extract information from the CAD and the 
Autodesk Apprentice Server was used to display the CAD models on the screen, Figure 
4. A flow chart, containing the method to automatically extract information about a 
straight line drawn in CAD, is shown in Figure 5. 
 
Add-In
(EXE)
Standalone
EXE
Add-In
(DLL)
VBA
Inventor
Application
Apprentice Server
Inventor
Data
 
Figure 4 accessing the Autodesk Inventor’s API 
 
Extract 3D Data from CAD
Select Item(s)
Is the selected item 
a SketchLine3D?
Yes
No
Define a Line3D as a 
ScketchLine3D
Get the starting point of 
the Line3D
Get the endpoint of the 
Line3D
 
Figure 5 extracting data from CAD (straight line) 
 
 
3.2. Position and orientation in space 
In order to off-line generate a robot program from a CAD environment and put it 
running in a real environment, the CAD cell should match with the real one. In other 
words, it is necessary to have all robot end-effector positions and orientations with 
respect to one or more reference frames known a priori by the robot. These frames are 
made known to the robot through a calibration process. Generally, this is a simple and 
non-time consuming process where the user needs to define the frame(s) within the 
CAD environment and then to teach the real robot about that frame(s)’ pose in the real 
scenario (off-line to on-line mapping). When there are a significant number of frames to 
define, the calibration process can be lengthy and prone to error. 
The API gives all the information (transformation matrices and path lines data) with 
respect to the origin of the CAD assembly model, the universe coordinate system {U}. 
Considering that a frame {B} is defined relative to {U} during the calibration process, 
from the API we have the transformation matrix of {B} relative to {U}, U
BT . This 
means that frame {B} “makes the link” between the virtual and real world. Note that, as 
mentioned above, it is possible to define more than one frame if necessary, as the 
process is similar. 
Since Autodesk Inventor considers the robot path lines drawn as a constituent of a 
single CAD part model (.ipt file) contained in the CAD assembly model (.iam file), the 
transformation matrix (relative to {U}) of that single part model defines the pose of the 
path lines. For the general case presented in Figure 6, the path line is part of the table 
top model in which the origin and orientation is defined by frame {E}. However, it is 
not necessary to know the orientation of the path lines as the API gives all the necessary 
points to define the path lines relative to {U}, for example the initial path point U
ini
P  
(Figure 6). So it is necessary to achieve the path line points relative to frame {B}. In 
terms of establishing the robot end-effector orientation, frames {C} and {D} help to 
define the origin and orientation of simplified tool models in Figure 6. As mentioned, 
the API gives the transformation matrix of these models relative to {U}, U
CT  and U
DT . 
However, for robot programming purposes we wish to express frame {C} and {D} in 
terms of frame {B}, B
CT  and B
DT . For the case of B
CT  we have: 
 
B
B
U
C
U
C
T
T
T


 
(1) 
 
To find B
UT , we must compute the rotation matrix that defines frame {U} relative to 
{B}, B
UR , and the vector that locates the origin of frame {U} relative to {B}, B
Uorg
P
. 
So, we know that: 
 
0
0
0
1
B
B
B
U
Uorg
U
R
P
T








 
(2) 
 
Given the characteristics of a rotation matrix, B
U
T
U
B
R= R , and as we know U
BT , the next 
step is to calculate B
Uorg
P
. Considering a generic vector/point defined in {U},  UP; if 
we wish to express this point in space in terms of frame {B} we must compute:    
 
 
B
B
U
B
U
Uorg
P
R
P
P



 
(3) 
 
For the specific case of the initial path point in Figure 6, 
ini
P , since the API gives U
ini
P , 
from (3) we can write 
ini
P  relative to {B}: 
 
 
B
B
U
B
ini
U
ini
Uorg
P
R
P
P



 
(4) 
 
Rewriting (3): 
 


 
B
U
B
U
B
Borg
U
Borg
Uorg
P
R
P
P



 
(5) 
 
The left side of (4) must be zero, so, from (4) we have: 
 
 
 
B
B
U
U
T
U
Uorg
U
Borg
B
Borg
P
R
P
R
P




 
(6) 
 
From (2) and (5) we can write:  
 
0
0
0
1
U
T
U
T
U
B
B
B
Borg
U
R
R
P
T










 
(7) 
 
Now, we can rewrite (1) and achieve B
CT . The same methodology can be used to 
achieve B
DT  and any other transformation.  
 
XB
ZB
YB
YC
XC
ZC
XU
ZU
YU
Robot Path
Tool Model
Borg
U P
ZD
XD
YD
XE
ZE
YE
ini
P
 
Figure 6 system frames  
 
3.3. Position and orientation interpolation 
When an industrial robot is performing a pre-programmed movement and this one 
requires abrupt end-effector orientation changes, we must take special care because it 
can come into a situation where no one has total control over the end-effector 
orientation. This is particularly true when robot programs are generated off-line. The 
proposed solution to circumvent this problem is based on the implementation of linear 
smooth interpolation of end-effector positions and orientations (Feng-yun and Tian-
sheng, 2005). The process involves the following steps: 
 
 Identification of risk areas (paths). This is done by analyzing the CAD model and 
manually defining those areas in the drawing. 
 Discretization of the risk path in equally spaced intervals. 
 Calculation of end-effector orientations for each interpolated path point. The new 
path is smoother than the initial (Figure 7). 
 
a)
b)
jP
2
jP 
jP
2
jP 
1

jP
 
Figure 7 (a) – end-effector pose before interpolation; (b) – end-effector pose after interpolation 
 
Consider 
T
x
y
z
r(k)= r (k)
r (k)
r (k)



 a generic end-effector position generated at the 
discrete time k and defined in 
j
j+2
P
P



, (Figure 7). 
jP , 
j+1
P
 and 
j+2
P
 are known end-
effector poses, extracted from the CAD drawing (see section 4.1.2). For the profile in 
Figure 7 (possible area of risk) we will separate the interpolation in two sections, 
1S  and 
2
S ; 
1
j
j+1
S
P
P



 and 
2
j+1
j+2
S
P
P



. The calculations are presented for section 
1S  
but for other sections the procedure is the same. So, r(k)  is calculated using both the 
known data points from CAD (
jP , 
j+1
P
) and the profiling velocity v(k) : 
 
( )
( )
( )
( )
T
x
y
z
v k
v k
v k
v k



 
(8) 
 
It is assumed that the magnitude of v(k) , v(k) , is a constant. Considering 
j
j+1
r(k)
P
P



, a direction vector W  can be defined as: 
 
1
j
j
W
P
P



 
(9) 
 
From (8) and (9), each directional velocity profile is obtained by: 
 
( )
( )
, (
, , )
i
i
W
v k
v k
i
x y z
W



 
(10) 
 
From (10), using a sampling width Δt , the interpolated position r(k)  is given by: 
 
,
,
,
(0)
T
j
j x
j y
j z
r
P
P
P
P




 
(11) 
 
1
1,
1,
1,
( )
T
j
j
x
j
y
j
z
r n
P
P
P
P








 
(12) 
 
(
, , )
( )
(0)
( )
,
(
1,...,
1)
i
i
i
i
x y z
r k
r
v k
k
t
k
n









 
(13) 
 
Note that n  represents the number of interpolated points. 
A quaternion interpolation algorithm (spherical linear interpolation – Slerp) to 
interpolate smoothly a sequence of end-effector orientations was used. For the profile in 
Figure 7 we will interpolate end-effector orientations between 
jP  and
j+2
P
. Given two 
known unit quaternions, 
0
Q  (from 
jP ) and 
n
Q  (from
j+2
P
), with parameter k  moving 
from 1 to n-1, the interpolated end-effector orientation 
k
Q  can be obtained as follows: 
 


0
1
1
sin
1
sin
1
1
,
1
1
sin
sin
k
n
k
k
n
n
Q
Q
Q
k
n



































 
(14) 
 
Where:  
 


1
0
cos
n
Q Q




 
(15) 
 
3.4. Robot program generation 
Using the information extracted from the CAD environment, the system presented here 
is able to generate robot programs for specific robotic applications. The code generation 
process is divided into two distinct phases: 
 
 Definition and parameterization of robot positions/orientations, reference frames, 
tools, etc. The end-effector positions and orientations extracted from CAD are used 
to define the robot path target poses (16). When confronted with risk areas the 
interpolation algorithms automatically generate the appropriate end-effector poses 
for these areas. From (3) we have the end-effector positions BP ; from (1) the 
transformation matrix 

B
CT  containing the rotation matrix, which in turn is used to 
calculate the end-effector orientation in the form of quaternions or Euler angles; from 
(13) the interpolated positions r(k) ; and finally from (14) the interpolated 
orientations (quaternions) 
k
Q . 
 
( )
,
,
,
1,
2,
3,
4
B
B
i
C
k
P and r k
T and Q
P
x
y
z
q
q
q
q

 
(16) 
 
 Body of the program. A robot program contains predominantly robot motion 
instructions (linear, joint, circular or spline robot movement). These movement 
instructions are selected according to the type of lines used in the CAD drawing to 
define the robot paths. 
 
4. Experiments 
Two different experiments are discussed, and in both cases, robot programs are 
generated off-line from a CAD drawing. In the first experiment, seam tracking, robot 
paths are adjusted with the information received from a laser camera attached to the 
robot. In the second experiment, a robot follows a geometric profile while maintaining a 
contact force, robot paths are adjusted with the information received from a force/torque 
(F/T) sensor attached to the robot wrist. 
To better visualize the robot path adjustments provided by sensory feedback, the 
robotic space was forced to become a more “viewable” unpredictable environment by 
purposely making a rough calibration process. Often, calibration errors arise from the 
little time and attention devoted to the robot calibration process. 
 
4.1. Seam tracking 
4.1.1. Experimental setup 
The experimental setup of the robotic platform (Figure 8) is the following: 
 
 An industrial robot ABB IRB 2400 equipped with a S4C+/M2000 controller. 
 A computer running Microsoft Windows Xp. 
 A laser camera DIGI-I/S from Servo Robot. 
 
The computer is running a CAD package (Autodesk Inventor) and the developed 
software interface, which receives data from CAD, interprets the data received and 
generates robot programs. The robot can be remotely controlled and managed by the 
software interface, which uses an ActiveX named PcRob for such purposes. The laser 
camera is connected with the robot controller via serial port. 
  
PcRob
3D CAD 
Package
 Software 
Interface
Welding Machine
Camera 
Controller
 
Figure 8 system architecture 
 
4.1.2. CAD model 
The CAD assembly model from which a robot program will be generated does not need 
to accurately represent the real cell in all its aspects (Figure 9). On the contrary, it can 
be a simplified model containing the “important” information. As an example, the robot 
tool length, robot paths and relative positioning of CAD models should represent the 
real scenario, however, the models appearance do not need to be exactly equal to the 
real objects.  
For this particular experiment, the CAD assembly model should contain the 
workpieces to be welded, the robot paths and the robot tools with the desired torch 
orientation for each path segment. In terms of risk areas, there is only one abrupt tool 
orientation change (Figure 9). 
 
Risk Path
X
Y
Z
 
Figure 9 CAD assembly model of the workpieces to be welded (butt joint). Note: a robot program will be 
generated from this model 
 
4.1.3. Path adjustment 
Analyzing the incoming data from the laser camera, the implemented control system 
decides which end-effector adjustments should be applied to the main paths extracted 
from CAD. The system modus operandi is relatively simple:  
 
1 Definition/calibration of the robot tool to match with the robot reference frame. 
2 The laser camera is configured with information about the welding joint and the 
desired vertical and/or horizontal distances (tool standoff) that the torch must 
maintain to the welding joint. 
3 Features from the workpiece profile are extracted and matched against the predefined 
joint templates and tolerances. 
4 The automatic end-effector adjustment is achieved by a closed loop position control 
that promotes compensation of the errors in Y and Z directions. Correction data are 
acquired with a sample rate of 5 Hz. 
 
4.1.4. Results and discussion 
Results showed that the CAD-based robot programming system is easy to use and 
within minutes an untrained user can generate a robot program for welding purposes. 
However, in the real scenario (Figure 10) we have a dynamic environment where robot 
path adjustments are required. Figure 11 shows the robot path adjustments/corrections 
(in the Y direction) made by the robot during the seam tracking process (Video 2, 
2010).  
As the robot only allows path adjustments at a frequency of 5 Hz, for higher welding 
speeds the path correction does not appear so smooth. Another limitation is the low 
robot resolution (0.01 mm), making the path adjustment process more abrupt.  
 
 
Figure 10 robotic cell 
 
0
10
20
30
40
50
60
70
80
0
2
4
6
Distance (mm)
Corrected path (mm)
“Ideal” Situation (gray)
Out of Line (dashed)
 
Figure 11 path adjustments in Y direction (robot velocity 10 mm/s) 
 
4.2. Profile following 
4.2.1. Experimental setup and features 
The experimental setup of the robotic platform (Figure 12) is the following: 
 
 An industrial robot Motoman HP6 equipped with the NX100 controller. 
 A computer running Microsoft Windows Xp. 
 A six degrees of freedom (DOF) F/T sensor from JR3. 
 A local area network (LAN), Ethernet and TCP/IP based, used for robot-computer 
communication (100 Mbps). 
 
The computer is running Autodesk Inventor and the developed software interface. This 
interface generates robot programs from CAD and manages the force control system, 
acquiring data from the F/T sensor and sending motion commands (adjustments) to the 
robot. The software interface communicates with the robot using a software component 
named MotomanLib. The ActiveX component JR3PCI is used to acquire force and 
torque data from the F/T sensor. The robot pose is adjusted with a sample rate of 20 Hz.  
As in the previous experiment, the robot program is generated from a CAD drawing 
(Figure 13). The real work environment is an unpredictable environment due to the 
“uncertainty” that comes from an inaccurate calibration process and due to the surface 
roughness of the workpiece. The robot tool should follow a geometric profile while 
maintaining a contact force. In order to facilitate the analysis of experimental results, a 
ball-shaped tool was mounted on the robot’s end-effector. 
 
MotomanLib
JR3PCI
3D CAD 
Package
 Software 
Interface
 
Figure 12 system architecture 
 
 
Figure 13 CAD assembly model of the working profile. 
 
4.2.2. Results and discussion 
Regarding the generation of the robot program from a CAD drawing, this experiment 
showed similar results to those of section 4.1.4. From the incoming data from the F/T 
sensor, the implemented force and robot displacement control system (Fuzzy-PI and PI 
reasoning) decides which displacements should be applied to the robot end-effector to 
achieve satisfactory performance (Mendes et al., 2010; Video 3, 2010). The force 
control system ensures that the contact forces are maintained at a constant value, 
adjusting the pre-programmed robot paths extracted from CAD (Figure 14 and Figure 
15). The graphs of Figure 14 show some force fluctuation due to the roughness of the 
surface and the noise of F/T data. 
 
0
50
100
150
200
-30
-20
-10
0
Force (N)
 
 
Axial Force
Set Point -15 (N)
0
50
100
150
200
-30
-20
-10
0
Force (N)
0
50
100
150
200
-5
0
5
10
Time (sec)
Position (mm)
0
50
100
150
200
-5
0
5
10
Position (mm)
0
50
100
150
200
-5
0
5
10
Time (sec)
Position (mm)
0
50
100
150
200
-5
0
5
10
Time (sec)
Position (mm)
 
Figure 14 experimental results by using a Fuzzy-PI controller (at left) and PI controller (at right) 
 
Axial Force
 
Figure 15 robot tool in contact with the real workpiece 
 
4.3. Overall results 
Some problems can occur when external sensors are used to on-line adjust robot 
motion:  
 
 Collisions between the external sensor and the surrounding workspace; 
 Situations in which the robot arm is sent to a location outside of the robot working 
area; 
 Kinematic singularities; 
 Poor choice of process parameters; 
 The communications delay between the external sensor and the robot controller can 
produce a negative effect on the proper definition of the robotic task. 
 
In order to avoid the above mentioned problems the operator should ensure that the 
workpieces are inside the working area of the robot, no collisions occur and kinematic 
singularities are identified.  
 
During effective robot operation, if a failure or malfunctioning is detected, two 
different situations can be considered: task abortion or activation of a reactive task. 
After aborting the process, the restarting of the system can be a complicated issue, 
depending on the type of robotic task. For example, for an arc welding application, 
restarting the system requires at least placing the torch at the point where the robot 
stopped. 
 
5. Conclusion and future work 
A CAD-based robot programming system with capacity to deal with dynamic and 
unpredictable environments was presented. Results showed that the proposed platform 
opens new possibilities for intuitive robot programming. It means that an untrained 
operator can generate a robot program for a specific task within minutes. Moreover, 
since the construction of the CAD models and robot programming task are performed in 
the same platform the entire robot programming process becomes easier and cheaper. 
This is very important for SMEs that produce small batches of products and need to 
constantly reprogram the robotic cells. In addition, sensory feedback enables the robot 
to be more flexible when confronted with product changeover. By adding sensory 
feedback to the robotic platforms we ensure that the robot manoeuvres in an 
unpredictable environment, damping possible impacts and increasing the tolerance to 
positioning errors from the calibration process or from the construction of the CAD 
models. 
Future work will be required to proceed with the development of methodologies 
which would facilitate sensor integration in robotic platforms, especially for when 
robots are programmed off-line. 
 
References 
Bhanu, B. (1987), “CAD-based robot vision”, IEEE Computer, Vol. 20 No. 8, pp. 12-
16. 
Bolmsjö, G. and Olsson, M. (2005), “Sensors in robotic arc welding to support small 
series production” Industrial Robot, Vol. 32 No. 4, pp. 341-5. 
Brink, K., Olsson, M., and Bolmsjö, G. (1997), “Increased autonomy in industrial 
robotic systems: a framework” Journal of Intelligent and Robotic Systems, Vol. 19, 
pp. 357-73. 
Bruyninckx, H., De Schutter, J. and Allotta, B. (1991), “Model-based constrained 
motion: a. modelling, specification and control” IEEE 5th International Conference 
on Advanced Robotics, Pisa, pp. 976-81. 
Cederberg, P., Olsson, M. and Bolmsjö, G. (2002), “Virtual triangulation sensor 
development, behavior simulation and CAR integration applied to robotic arc-
welding” Journal of Intelligent and Robotic Systems, Vol. 35, pp. 365-79. 
Chen, H., Fuhlbrigge, T. and Li, X. (2009), “A review of CAD-based robot path 
planning for spray painting”, Industrial Robot, Vol. 36 No. 1, pp. 45-50. 
Feng-yun, L. and Tian-sheng, L. (2005), “Development of a robot system for complex 
surfaces polishing based on CL data”, The International Journal of Advanced 
Manufacturing Technology, Vol. 26, pp. 1132-7. 
Fridenfalk, M., and Bolmsjö, G. (2002), “Design and validation of a sensor guided robot 
control system for welding in shipbuilding” International Journal for the Joining of 
Materials, Vol. 14 No. 3/4, pp. 44-55. 
Johansson, R., Robertsson, A., Nilsson, K., Brogardh, T., Cederberg, P., Olsson, M., 
Olsson, T. and Bolmsjö, G. (2004), “Sensor integration in task-level programming 
and industrial robotic task execution control” Industrial Robot, Vol. 31 No. 3, pp. 
284-96. 
Kenney, J., Buckley, T. and Brock, O. (2009), “Interactive Segmentation for 
Manipulation in Unstructured Environments” IEEE International Conference on 
Robotics and Automation, Kobe, Japan, pp. 1337-82. 
Kim, J.Y. (2004), “CAD-based automated robot programming in adhesive spray 
systems for shoe outsoles and uppers” Journal of Robotic Systems, Vol. 21 No. 11, 
pp. 625-34. 
Kopacek, P. (1999), “Intelligent manufacturing: present state and future trends” Journal 
of Intelligent and Robotic Systems, Vol. 26, pp. 217-29. 
Mendes, N., Neto, P., Pires, J.N. and Moreira, A.P. (2010), “Fuzzy-PI force control for 
industrial robotics”, in Vadakkepat et al. (Ed.), Trends in Intelligent Robotics, 
Springer-Verlag, Berlin Heidelberg, pp. 322-9. 
Nagata, F., Kusumoto, Y., Fujimoto, Y. and Watanabe, K. (2007), “Robotic sanding 
system for new designed furniture with free-formed surface”, Robotics and 
Computer-Integrated Manufacturing, Vol. 23 No. 4, pp. 371-9. 
Nayak, N. and Ray, A. (1990), “An integrated system for intelligent seam tracking in 
robotic welding: part I – conceptual and analytical development” IEEE International 
Conference on Robotics and Automation, pp. 1892-7. 
Neto, P., Pires, J.N. and Moreira, A.P. (2010a), “High-level programming and control 
for industrial robotics: using a hand-held accelerometer-based input device for 
gesture and posture recognition” Industrial Robot, Vol. 37 No. 2, pp. 137-47. 
Neto, P., Pires, J.N. and Moreira, A.P. (2010b), “CAD-based off-line robot 
programming” IEEE International Conference on Robotics, Automation and 
Mechatronics, Singapore, pp. 516-21. 
Pires, J.N., Godinho, T. and Ferreira, P. (2004), “CAD interface for automatic robot 
welding programming” Industrial Robot, Vol. 31 No. 1, pp. 71-6. 
Schaefer, T. and Schraft, D. (2005), “Incremental sheet metal forming by industrial 
robot” Rapid Prototyping Journal, Vol. 11 No. 5, pp. 278-86. 
Video 1 (2010), “Robot program generation from CAD virtual paths”, available at: 
http://robotics.dem.uc.pt/pedro.neto/GS3.html (accessed 15 December 2010) 
Video 2 (2010), “Robot path adjustment – laser camera”, available at: 
http://robotics.dem.uc.pt/pedro.neto/GS4.html (accessed 15 December 2010) 
Video 3 (2010), “Robot path adjustment – force sensor”, available at: 
http://robotics.dem.uc.pt/pedro.neto/GS5.html (accessed 15 December 2010) 
 
High-level robot programming based on CAD: 
dealing with unpredictable environments 
 
 
1Pedro Neto, 1Nuno Mendes, 1Ricardo Araújo, 1J. Norberto Pires 
2A. Paulo Moreira 
 
1Department of Mechanical Engineering (CEMUC), University of Coimbra, Coimbra, Portugal 
2Institute for Systems and Computer Engineering of Porto (INESC-Porto), Porto, Portugal 
 
 
Abstract: 
Purpose – The purpose of this paper is to present a CAD-based human-robot interface 
that allows non-expert users to teach a robot in a manner similar to that used by humans 
to teach each other. Another important issue addressed here has to do with how robots 
deal with uncertainty.  
 
Design/methodology/approach – Intuitive robot programming is achieved by using 
CAD drawings to generate robot programs off-line. Sensory feedback allows 
minimization of the effects of uncertainty, providing information to adjust the robot 
paths during robot operation. 
 
Findings – It was found that it is possible to generate a robot program from a common 
CAD drawing and run it without any major concerns about calibration or CAD model 
accuracy. 
 
Research limitations/implications – A limitation of the proposed system has to do 
with the fact that it was designed to be used for particular technological applications. 
 
Practical implications – Since today most of the manufacturing companies have CAD 
packages in their facilities, CAD-based robot programming may be a good option to 
program robots without the need for skilled robot programmers.  
 
Originality/value – A new CAD-based robot programming system is proposed. Robot 
programs are directly generated from a CAD drawing “running” on a commonly 
available 3D CAD package and not from a commercial computer aided robotics 
software, making it a simple CAD integrated solution. This is a low-cost and low-setup 
time system where no advanced robot programming skills are required to operate it.  
 
Keywords: CAD, Industrial Robotics, High-Level Programming, Sensory Feedback, 
Unpredictable Environments. 
 
 
 
 
 
 
 
 
 
 
 
 
 
1. Introduction 
1.1. Motivation 
Increasingly, companies are changing and reinventing their production systems. 
Traditional manufacturing systems (often based on fixed automation and manual work) 
are being replaced by flexible and intelligent manufacturing systems, enabling 
companies to continue to be competitive in the global market (Kopacek, 1999). This 
competitiveness is reflected in the companies’ capacity to respond/react quickly to 
market demands, producing more and better quality products at competitive prices. 
Owing to its flexibility, programmability and efficiency, industrial robots are seen as 
a fundamental element of modern flexible manufacturing systems. Nevertheless, there 
are still some problems that hinder the utilization of robots in industry, especially in 
small and medium-sized enterprises (SMEs). SMEs have difficulty finding skilled 
workers capable of operating with robots. Therefore, new and more intuitive ways for 
people to interact with robots are required to make robot programming more accessible, 
easier and faster. The goal is that the instructor can teach a robot in a manner similar to 
that used by humans to teach each other, for example using CAD drawings, gestures or 
through verbal explanation (Neto et al., 2010a). 
 
1.2. Objectives 
Robot programming through the typical teaching method (using the teach pendant) is a 
tedious and time-consuming task that requires technical expertise. The goal is to 
develop methodologies that help users to program a robot in an intuitive way, quickly, 
with a high-level of abstraction from the robot specific language, and, if possible, 
without speeding too much money. 
In this paper, a CAD-based system to program a robot from a 3D CAD drawing, 
allowing users with basic skills in CAD and robot programming to generate robot 
programs off-line, is presented. In addition, the 3D CAD package, Autodesk Inventor, 
that interfaces with the user is a well-known generic CAD package, widespread on the 
market at a relative low-cost. Starting from the CAD model of the robotic cell in study, 
the way the user generates a robot program is as simple as “drawing” the desired robot 
paths in the CAD environment. Later, the information needed is automatically extracted 
from the CAD environment, analysed and converted into robot programs. Note that the 
robot programs are not extracted neither from a computer aided manufacturing (CAM) 
software nor from a computer aided robotics (CAR) software. It means that we are 
proposing a simple CAD integrated solution for the robotics field. 
Commercial CAR packages are powerful tools, which enable modelling, simulation 
and robot programming. Nevertheless, they have some disadvantages that hinder their 
use in companies, especially in SMEs. By comparing commercial CAR packages with a 
CAD-based robot programming system similar to that presented in this paper (Neto et 
al., 2010b), it was found that the CAD-based system has some relative advantages: 
 
• Low-cost. Since the construction of CAD models and the robot programming task 
are performed in the same environment/platform (Autodesk Inventor) the programming 
task becomes easier and cheaper; 
• Short learning curve; 
• Simplicity of use. The most time consuming task, the construction of the CAD 
model, is present in both systems.  
 
CAD-based robot programming approaches work well if the environment of the 
robot tasks is well defined. In the other hand, there are situations which are likely to 
create errors or impede the normal operation of the robot: 
 
• The CAD models do not reproduce correctly the geometry of the real scenario; 
• Inaccuracies created in the robot calibration process; 
• Inefficient fixtures that do not ensure the static character of the workpieces; 
• A “foreign” object is introduced in the real environment. 
  
In these cases, we can say that we are in the presence of a dynamic and unpredictable 
environment.  
To perform successful manipulation robots depend on precise information about 
objects in their surrounding. In an unpredictable environment, such information cannot 
be given to the robot a priori, robots have to autonomously and continuously acquire 
information about their surrounding environment to support their decision making and 
react to unanticipated events. Sensory feedback allows a robot to recognize your work 
environment for itself, for example producing corrections (on-line) in pre-programmed 
robot paths (Figure 1). In fact, the integration of sensors into robotic platforms reduces 
the setup time, the need for accurate robot trajectory programming and promotes 
flexibility and the autonomous behaviour of robotic systems (Bolmsjö and Olsson, 
2005; Johansson et al., 2004).  
 
P1
P1
P1
P2
P2
P2
Planned 
path
a)
b)
c)
Robot path
Robot 
path
Collision
 
Figure 1 (a) – planned path for a specific environment; (b) – a “foreign” object is introduced into the 
environment and collision occurs; (c) – sensory feedback is introduced, helping the robot to deal with the 
unpredictable environment (robot path is adjusted) 
 
We validate our methods with two different real-world experiments for two different 
tasks, seam tracking and for applications that require the robot to follow a geometric 
profile while maintaining a contact force. 
 
2. Related Work  
In recent years, CAD technology has become economically attractive and easy to work 
with so that today millions of SMEs worldwide are using it to design and model their 
products. Already in the 80’s, CAD was seen as a technology that could help in the 
development of robotics (Bhanu, 1987). Since then, a variety of research has been 
conducted in the field of CAD-based robot planning and programming. 
Pires et al. (2004) proposes to extract robot motion information from a CAD DXF 
file and converting it into robot commands for welding purposes. A review of CAD-
based robot path planning for spray painting is presented by Chen et al. (2009). Another 
study presents a method to generate 3D robot working paths for a robotic adhesive 
spray system for shoe outsoles and uppers (Kim, 2004). Nagata et al. (2007) proposes a 
robotic sanding platform where the robot paths are generated by CAD/CAM software. 
An example of a novel process that benefits from the robots and CAD versatility is the 
so-called incremental forming process of metal sheets (Schaefer and Schraft, 2005). 
Feng-yun and Tian-sheng, (2005) presents a robot path generator for the polishing 
process, where the cutter location data is generated from the postprocessor of a CAD 
system. As we have seen above, a variety of research has been done in the area of CAD-
based robot planning and programming. However, none of the studies so far deals with 
a “global” solution for this problem.  
Unpredictable environments pose a significant challenge because of their complexity 
and inherent uncertainty. Over the last few years, important studies have been carried 
out to deal with uncertainty in the robotics field: using models of “ideal” environments, 
sensory feedback, and implementing reasoning methods into robotic platforms 
(Bruyninckx et al., 1991; Nayak and Ray, 1990). These concepts have evolved and 
recently, researchers have been successful in developing skills that can handle the 
complexity of dynamic and predictable environments (Kenney et al. 2009; Mendes et 
al., 2010). A number of authors have devoted attention to sensor simulation, trying to 
mimic as closely as possible the behaviour of a real sensor, and thus integrating it (the 
virtual sensor) within a CAR platform (Cederberg et al., 2002; Brink et al., 1997; 
Bolmsjö and Olsson, 2005). Moreover, sensor information has been used to update 
robotic cell models in real-time, allowing to avoid problems such as collisions, 
kinematic singularities and exceeding of joint limits (Brink et al., 1997; Johansson et 
al., 2004). 
 
The concept of seam tracking applied to robotic welding has been studied over the 
last two decades (Nayak and Ray, 1990). Recently, important work has been carried out 
in the integration of sensors to assist the robotic arc welding process (Fridenfalk and 
Bolmsjö, 2002; Bolmsjö and Olsson, 2005). 
 
3. Robot Programming from CAD  
Starting from a 3D CAD model of the robotic cell in study, the way the user generates a 
robot program can be as simple as “drawing” the desired robot paths in the CAD 
environment. Furthermore, to define the robot end-effector pose (position and 
orientation), it is necessary to know, not only the robot path positions but also the end-
effector orientations in space. Therefore, after drawing the robot paths, simplified tool 
models should be placed along the paths. These models will define the orientation of the 
robot end-effector in each segment of the path (Figure 2). 
 
Xn-1
Xn
Xn+1
Yn-1
Zn
Yn
Yn+1
Zn+1
Zn-1
Simplified tool model
Robot path
 
Figure 2 simplified tool models defining the end-effector orientation 
 
 The information needed to program the robot will be extracted from the CAD 
environment by using an application programming interface (API) provided by 
Autodesk. This API allows the extraction of the points that characterize each of the 
different lines used to define a robot path; straight lines, splines and arcs. Moreover, the 
API also gives information about the transformation matrix of each part model 
represented in the CAD environment. The transformation matrix contains the rotation 
matrix and the position of the origin of the part model to which it refers, both in relation 
to the origin of the CAD assembly model. Later, the information extracted from the 
CAD is converted into robot programs (Video 1, 2010). A diagram with the procedure 
to extract 3D data from CAD and their conversion into a robot program is presented in 
Figure 3. 
 
Start
Open a 3D CAD Drawing
Draw Robot Paths
Select Robot Paths
Extract Data from CAD
(Position and Orientation)
Define Robot Tool Orientation
Define Robot Parameters
Generate Robot Program
End
Execute Interface Software
 
Figure 3 extracting 3D data from CAD  
 
3.1. Application programming interface 
The Autodesk Inventor API shows the Inventor’s functionalities in an object-oriented 
manner, allowing developers to interact with Autodesk Inventor using current 
programming languages; Visual Basic, Visual C#, Visual C++. In our proposed system, 
a standalone application was used to extract information from the CAD and the 
Autodesk Apprentice Server was used to display the CAD models on the screen, Figure 
4. A flow chart, containing the method to automatically extract information about a 
straight line drawn in CAD, is shown in Figure 5. 
 
Add-In
(EXE)
Standalone
EXE
Add-In
(DLL)
VBA
Inventor
Application
Apprentice Server
Inventor
Data
 
Figure 4 accessing the Autodesk Inventor’s API 
 
Extract 3D Data from CAD
Select Item(s)
Is the selected item 
a SketchLine3D?
Yes
No
Define a Line3D as a 
ScketchLine3D
Get the starting point of 
the Line3D
Get the endpoint of the 
Line3D
 
Figure 5 extracting data from CAD (straight line) 
 
 
3.2. Position and orientation in space 
In order to off-line generate a robot program from a CAD environment and put it 
running in a real environment, the CAD cell should match with the real one. In other 
words, it is necessary to have all robot end-effector positions and orientations with 
respect to one or more reference frames known a priori by the robot. These frames are 
made known to the robot through a calibration process. Generally, this is a simple and 
non-time consuming process where the user needs to define the frame(s) within the 
CAD environment and then to teach the real robot about that frame(s)’ pose in the real 
scenario (off-line to on-line mapping). When there are a significant number of frames to 
define, the calibration process can be lengthy and prone to error. 
The API gives all the information (transformation matrices and path lines data) with 
respect to the origin of the CAD assembly model, the universe coordinate system {U}. 
Considering that a frame {B} is defined relative to {U} during the calibration process, 
from the API we have the transformation matrix of {B} relative to {U}, U
BT . This 
means that frame {B} “makes the link” between the virtual and real world. Note that, as 
mentioned above, it is possible to define more than one frame if necessary, as the 
process is similar. 
Since Autodesk Inventor considers the robot path lines drawn as a constituent of a 
single CAD part model (.ipt file) contained in the CAD assembly model (.iam file), the 
transformation matrix (relative to {U}) of that single part model defines the pose of the 
path lines. For the general case presented in Figure 6, the path line is part of the table 
top model in which the origin and orientation is defined by frame {E}. However, it is 
not necessary to know the orientation of the path lines as the API gives all the necessary 
points to define the path lines relative to {U}, for example the initial path point U
ini
P  
(Figure 6). So it is necessary to achieve the path line points relative to frame {B}. In 
terms of establishing the robot end-effector orientation, frames {C} and {D} help to 
define the origin and orientation of simplified tool models in Figure 6. As mentioned, 
the API gives the transformation matrix of these models relative to {U}, U
CT  and U
DT . 
However, for robot programming purposes we wish to express frame {C} and {D} in 
terms of frame {B}, B
CT  and B
DT . For the case of B
CT  we have: 
 
B
B
U
C
U
C
T
T
T


 
(1) 
 
To find B
UT , we must compute the rotation matrix that defines frame {U} relative to 
{B}, B
UR , and the vector that locates the origin of frame {U} relative to {B}, B
Uorg
P
. 
So, we know that: 
 
0
0
0
1
B
B
B
U
Uorg
U
R
P
T








 
(2) 
 
Given the characteristics of a rotation matrix, B
U
T
U
B
R= R , and as we know U
BT , the next 
step is to calculate B
Uorg
P
. Considering a generic vector/point defined in {U},  UP; if 
we wish to express this point in space in terms of frame {B} we must compute:    
 
 
B
B
U
B
U
Uorg
P
R
P
P



 
(3) 
 
For the specific case of the initial path point in Figure 6, 
ini
P , since the API gives U
ini
P , 
from (3) we can write 
ini
P  relative to {B}: 
 
 
B
B
U
B
ini
U
ini
Uorg
P
R
P
P



 
(4) 
 
Rewriting (3): 
 


 
B
U
B
U
B
Borg
U
Borg
Uorg
P
R
P
P



 
(5) 
 
The left side of (4) must be zero, so, from (4) we have: 
 
 
 
B
B
U
U
T
U
Uorg
U
Borg
B
Borg
P
R
P
R
P




 
(6) 
 
From (2) and (5) we can write:  
 
0
0
0
1
U
T
U
T
U
B
B
B
Borg
U
R
R
P
T










 
(7) 
 
Now, we can rewrite (1) and achieve B
CT . The same methodology can be used to 
achieve B
DT  and any other transformation.  
 
XB
ZB
YB
YC
XC
ZC
XU
ZU
YU
Robot Path
Tool Model
Borg
U P
ZD
XD
YD
XE
ZE
YE
ini
P
 
Figure 6 system frames  
 
3.3. Position and orientation interpolation 
When an industrial robot is performing a pre-programmed movement and this one 
requires abrupt end-effector orientation changes, we must take special care because it 
can come into a situation where no one has total control over the end-effector 
orientation. This is particularly true when robot programs are generated off-line. The 
proposed solution to circumvent this problem is based on the implementation of linear 
smooth interpolation of end-effector positions and orientations (Feng-yun and Tian-
sheng, 2005). The process involves the following steps: 
 
 Identification of risk areas (paths). This is done by analyzing the CAD model and 
manually defining those areas in the drawing. 
 Discretization of the risk path in equally spaced intervals. 
 Calculation of end-effector orientations for each interpolated path point. The new 
path is smoother than the initial (Figure 7). 
 
a)
b)
jP
2
jP 
jP
2
jP 
1

jP
 
Figure 7 (a) – end-effector pose before interpolation; (b) – end-effector pose after interpolation 
 
Consider 
T
x
y
z
r(k)= r (k)
r (k)
r (k)



 a generic end-effector position generated at the 
discrete time k and defined in 
j
j+2
P
P



, (Figure 7). 
jP , 
j+1
P
 and 
j+2
P
 are known end-
effector poses, extracted from the CAD drawing (see section 4.1.2). For the profile in 
Figure 7 (possible area of risk) we will separate the interpolation in two sections, 
1S  and 
2
S ; 
1
j
j+1
S
P
P



 and 
2
j+1
j+2
S
P
P



. The calculations are presented for section 
1S  
but for other sections the procedure is the same. So, r(k)  is calculated using both the 
known data points from CAD (
jP , 
j+1
P
) and the profiling velocity v(k) : 
 
( )
( )
( )
( )
T
x
y
z
v k
v k
v k
v k



 
(8) 
 
It is assumed that the magnitude of v(k) , v(k) , is a constant. Considering 
j
j+1
r(k)
P
P



, a direction vector W  can be defined as: 
 
1
j
j
W
P
P



 
(9) 
 
From (8) and (9), each directional velocity profile is obtained by: 
 
( )
( )
, (
, , )
i
i
W
v k
v k
i
x y z
W



 
(10) 
 
From (10), using a sampling width Δt , the interpolated position r(k)  is given by: 
 
,
,
,
(0)
T
j
j x
j y
j z
r
P
P
P
P




 
(11) 
 
1
1,
1,
1,
( )
T
j
j
x
j
y
j
z
r n
P
P
P
P








 
(12) 
 
(
, , )
( )
(0)
( )
,
(
1,...,
1)
i
i
i
i
x y z
r k
r
v k
k
t
k
n









 
(13) 
 
Note that n  represents the number of interpolated points. 
A quaternion interpolation algorithm (spherical linear interpolation – Slerp) to 
interpolate smoothly a sequence of end-effector orientations was used. For the profile in 
Figure 7 we will interpolate end-effector orientations between 
jP  and
j+2
P
. Given two 
known unit quaternions, 
0
Q  (from 
jP ) and 
n
Q  (from
j+2
P
), with parameter k  moving 
from 1 to n-1, the interpolated end-effector orientation 
k
Q  can be obtained as follows: 
 


0
1
1
sin
1
sin
1
1
,
1
1
sin
sin
k
n
k
k
n
n
Q
Q
Q
k
n



































 
(14) 
 
Where:  
 


1
0
cos
n
Q Q




 
(15) 
 
3.4. Robot program generation 
Using the information extracted from the CAD environment, the system presented here 
is able to generate robot programs for specific robotic applications. The code generation 
process is divided into two distinct phases: 
 
 Definition and parameterization of robot positions/orientations, reference frames, 
tools, etc. The end-effector positions and orientations extracted from CAD are used 
to define the robot path target poses (16). When confronted with risk areas the 
interpolation algorithms automatically generate the appropriate end-effector poses 
for these areas. From (3) we have the end-effector positions BP ; from (1) the 
transformation matrix 

B
CT  containing the rotation matrix, which in turn is used to 
calculate the end-effector orientation in the form of quaternions or Euler angles; from 
(13) the interpolated positions r(k) ; and finally from (14) the interpolated 
orientations (quaternions) 
k
Q . 
 
( )
,
,
,
1,
2,
3,
4
B
B
i
C
k
P and r k
T and Q
P
x
y
z
q
q
q
q

 
(16) 
 
 Body of the program. A robot program contains predominantly robot motion 
instructions (linear, joint, circular or spline robot movement). These movement 
instructions are selected according to the type of lines used in the CAD drawing to 
define the robot paths. 
 
4. Experiments 
Two different experiments are discussed, and in both cases, robot programs are 
generated off-line from a CAD drawing. In the first experiment, seam tracking, robot 
paths are adjusted with the information received from a laser camera attached to the 
robot. In the second experiment, a robot follows a geometric profile while maintaining a 
contact force, robot paths are adjusted with the information received from a force/torque 
(F/T) sensor attached to the robot wrist. 
To better visualize the robot path adjustments provided by sensory feedback, the 
robotic space was forced to become a more “viewable” unpredictable environment by 
purposely making a rough calibration process. Often, calibration errors arise from the 
little time and attention devoted to the robot calibration process. 
 
4.1. Seam tracking 
4.1.1. Experimental setup 
The experimental setup of the robotic platform (Figure 8) is the following: 
 
 An industrial robot ABB IRB 2400 equipped with a S4C+/M2000 controller. 
 A computer running Microsoft Windows Xp. 
 A laser camera DIGI-I/S from Servo Robot. 
 
The computer is running a CAD package (Autodesk Inventor) and the developed 
software interface, which receives data from CAD, interprets the data received and 
generates robot programs. The robot can be remotely controlled and managed by the 
software interface, which uses an ActiveX named PcRob for such purposes. The laser 
camera is connected with the robot controller via serial port. 
  
PcRob
3D CAD 
Package
 Software 
Interface
Welding Machine
Camera 
Controller
 
Figure 8 system architecture 
 
4.1.2. CAD model 
The CAD assembly model from which a robot program will be generated does not need 
to accurately represent the real cell in all its aspects (Figure 9). On the contrary, it can 
be a simplified model containing the “important” information. As an example, the robot 
tool length, robot paths and relative positioning of CAD models should represent the 
real scenario, however, the models appearance do not need to be exactly equal to the 
real objects.  
For this particular experiment, the CAD assembly model should contain the 
workpieces to be welded, the robot paths and the robot tools with the desired torch 
orientation for each path segment. In terms of risk areas, there is only one abrupt tool 
orientation change (Figure 9). 
 
Risk Path
X
Y
Z
 
Figure 9 CAD assembly model of the workpieces to be welded (butt joint). Note: a robot program will be 
generated from this model 
 
4.1.3. Path adjustment 
Analyzing the incoming data from the laser camera, the implemented control system 
decides which end-effector adjustments should be applied to the main paths extracted 
from CAD. The system modus operandi is relatively simple:  
 
1 Definition/calibration of the robot tool to match with the robot reference frame. 
2 The laser camera is configured with information about the welding joint and the 
desired vertical and/or horizontal distances (tool standoff) that the torch must 
maintain to the welding joint. 
3 Features from the workpiece profile are extracted and matched against the predefined 
joint templates and tolerances. 
4 The automatic end-effector adjustment is achieved by a closed loop position control 
that promotes compensation of the errors in Y and Z directions. Correction data are 
acquired with a sample rate of 5 Hz. 
 
4.1.4. Results and discussion 
Results showed that the CAD-based robot programming system is easy to use and 
within minutes an untrained user can generate a robot program for welding purposes. 
However, in the real scenario (Figure 10) we have a dynamic environment where robot 
path adjustments are required. Figure 11 shows the robot path adjustments/corrections 
(in the Y direction) made by the robot during the seam tracking process (Video 2, 
2010).  
As the robot only allows path adjustments at a frequency of 5 Hz, for higher welding 
speeds the path correction does not appear so smooth. Another limitation is the low 
robot resolution (0.01 mm), making the path adjustment process more abrupt.  
 
 
Figure 10 robotic cell 
 
0
10
20
30
40
50
60
70
80
0
2
4
6
Distance (mm)
Corrected path (mm)
“Ideal” Situation (gray)
Out of Line (dashed)
 
Figure 11 path adjustments in Y direction (robot velocity 10 mm/s) 
 
4.2. Profile following 
4.2.1. Experimental setup and features 
The experimental setup of the robotic platform (Figure 12) is the following: 
 
 An industrial robot Motoman HP6 equipped with the NX100 controller. 
 A computer running Microsoft Windows Xp. 
 A six degrees of freedom (DOF) F/T sensor from JR3. 
 A local area network (LAN), Ethernet and TCP/IP based, used for robot-computer 
communication (100 Mbps). 
 
The computer is running Autodesk Inventor and the developed software interface. This 
interface generates robot programs from CAD and manages the force control system, 
acquiring data from the F/T sensor and sending motion commands (adjustments) to the 
robot. The software interface communicates with the robot using a software component 
named MotomanLib. The ActiveX component JR3PCI is used to acquire force and 
torque data from the F/T sensor. The robot pose is adjusted with a sample rate of 20 Hz.  
As in the previous experiment, the robot program is generated from a CAD drawing 
(Figure 13). The real work environment is an unpredictable environment due to the 
“uncertainty” that comes from an inaccurate calibration process and due to the surface 
roughness of the workpiece. The robot tool should follow a geometric profile while 
maintaining a contact force. In order to facilitate the analysis of experimental results, a 
ball-shaped tool was mounted on the robot’s end-effector. 
 
MotomanLib
JR3PCI
3D CAD 
Package
 Software 
Interface
 
Figure 12 system architecture 
 
 
Figure 13 CAD assembly model of the working profile. 
 
4.2.2. Results and discussion 
Regarding the generation of the robot program from a CAD drawing, this experiment 
showed similar results to those of section 4.1.4. From the incoming data from the F/T 
sensor, the implemented force and robot displacement control system (Fuzzy-PI and PI 
reasoning) decides which displacements should be applied to the robot end-effector to 
achieve satisfactory performance (Mendes et al., 2010; Video 3, 2010). The force 
control system ensures that the contact forces are maintained at a constant value, 
adjusting the pre-programmed robot paths extracted from CAD (Figure 14 and Figure 
15). The graphs of Figure 14 show some force fluctuation due to the roughness of the 
surface and the noise of F/T data. 
 
0
50
100
150
200
-30
-20
-10
0
Force (N)
 
 
Axial Force
Set Point -15 (N)
0
50
100
150
200
-30
-20
-10
0
Force (N)
0
50
100
150
200
-5
0
5
10
Time (sec)
Position (mm)
0
50
100
150
200
-5
0
5
10
Position (mm)
0
50
100
150
200
-5
0
5
10
Time (sec)
Position (mm)
0
50
100
150
200
-5
0
5
10
Time (sec)
Position (mm)
 
Figure 14 experimental results by using a Fuzzy-PI controller (at left) and PI controller (at right) 
 
Axial Force
 
Figure 15 robot tool in contact with the real workpiece 
 
4.3. Overall results 
Some problems can occur when external sensors are used to on-line adjust robot 
motion:  
 
 Collisions between the external sensor and the surrounding workspace; 
 Situations in which the robot arm is sent to a location outside of the robot working 
area; 
 Kinematic singularities; 
 Poor choice of process parameters; 
 The communications delay between the external sensor and the robot controller can 
produce a negative effect on the proper definition of the robotic task. 
 
In order to avoid the above mentioned problems the operator should ensure that the 
workpieces are inside the working area of the robot, no collisions occur and kinematic 
singularities are identified.  
 
During effective robot operation, if a failure or malfunctioning is detected, two 
different situations can be considered: task abortion or activation of a reactive task. 
After aborting the process, the restarting of the system can be a complicated issue, 
depending on the type of robotic task. For example, for an arc welding application, 
restarting the system requires at least placing the torch at the point where the robot 
stopped. 
 
5. Conclusion and future work 
A CAD-based robot programming system with capacity to deal with dynamic and 
unpredictable environments was presented. Results showed that the proposed platform 
opens new possibilities for intuitive robot programming. It means that an untrained 
operator can generate a robot program for a specific task within minutes. Moreover, 
since the construction of the CAD models and robot programming task are performed in 
the same platform the entire robot programming process becomes easier and cheaper. 
This is very important for SMEs that produce small batches of products and need to 
constantly reprogram the robotic cells. In addition, sensory feedback enables the robot 
to be more flexible when confronted with product changeover. By adding sensory 
feedback to the robotic platforms we ensure that the robot manoeuvres in an 
unpredictable environment, damping possible impacts and increasing the tolerance to 
positioning errors from the calibration process or from the construction of the CAD 
models. 
Future work will be required to proceed with the development of methodologies 
which would facilitate sensor integration in robotic platforms, especially for when 
robots are programmed off-line. 
 
References 
Bhanu, B. (1987), “CAD-based robot vision”, IEEE Computer, Vol. 20 No. 8, pp. 12-
16. 
Bolmsjö, G. and Olsson, M. (2005), “Sensors in robotic arc welding to support small 
series production” Industrial Robot, Vol. 32 No. 4, pp. 341-5. 
Brink, K., Olsson, M., and Bolmsjö, G. (1997), “Increased autonomy in industrial 
robotic systems: a framework” Journal of Intelligent and Robotic Systems, Vol. 19, 
pp. 357-73. 
Bruyninckx, H., De Schutter, J. and Allotta, B. (1991), “Model-based constrained 
motion: a. modelling, specification and control” IEEE 5th International Conference 
on Advanced Robotics, Pisa, pp. 976-81. 
Cederberg, P., Olsson, M. and Bolmsjö, G. (2002), “Virtual triangulation sensor 
development, behavior simulation and CAR integration applied to robotic arc-
welding” Journal of Intelligent and Robotic Systems, Vol. 35, pp. 365-79. 
Chen, H., Fuhlbrigge, T. and Li, X. (2009), “A review of CAD-based robot path 
planning for spray painting”, Industrial Robot, Vol. 36 No. 1, pp. 45-50. 
Feng-yun, L. and Tian-sheng, L. (2005), “Development of a robot system for complex 
surfaces polishing based on CL data”, The International Journal of Advanced 
Manufacturing Technology, Vol. 26, pp. 1132-7. 
Fridenfalk, M., and Bolmsjö, G. (2002), “Design and validation of a sensor guided robot 
control system for welding in shipbuilding” International Journal for the Joining of 
Materials, Vol. 14 No. 3/4, pp. 44-55. 
Johansson, R., Robertsson, A., Nilsson, K., Brogardh, T., Cederberg, P., Olsson, M., 
Olsson, T. and Bolmsjö, G. (2004), “Sensor integration in task-level programming 
and industrial robotic task execution control” Industrial Robot, Vol. 31 No. 3, pp. 
284-96. 
Kenney, J., Buckley, T. and Brock, O. (2009), “Interactive Segmentation for 
Manipulation in Unstructured Environments” IEEE International Conference on 
Robotics and Automation, Kobe, Japan, pp. 1337-82. 
Kim, J.Y. (2004), “CAD-based automated robot programming in adhesive spray 
systems for shoe outsoles and uppers” Journal of Robotic Systems, Vol. 21 No. 11, 
pp. 625-34. 
Kopacek, P. (1999), “Intelligent manufacturing: present state and future trends” Journal 
of Intelligent and Robotic Systems, Vol. 26, pp. 217-29. 
Mendes, N., Neto, P., Pires, J.N. and Moreira, A.P. (2010), “Fuzzy-PI force control for 
industrial robotics”, in Vadakkepat et al. (Ed.), Trends in Intelligent Robotics, 
Springer-Verlag, Berlin Heidelberg, pp. 322-9. 
Nagata, F., Kusumoto, Y., Fujimoto, Y. and Watanabe, K. (2007), “Robotic sanding 
system for new designed furniture with free-formed surface”, Robotics and 
Computer-Integrated Manufacturing, Vol. 23 No. 4, pp. 371-9. 
Nayak, N. and Ray, A. (1990), “An integrated system for intelligent seam tracking in 
robotic welding: part I – conceptual and analytical development” IEEE International 
Conference on Robotics and Automation, pp. 1892-7. 
Neto, P., Pires, J.N. and Moreira, A.P. (2010a), “High-level programming and control 
for industrial robotics: using a hand-held accelerometer-based input device for 
gesture and posture recognition” Industrial Robot, Vol. 37 No. 2, pp. 137-47. 
Neto, P., Pires, J.N. and Moreira, A.P. (2010b), “CAD-based off-line robot 
programming” IEEE International Conference on Robotics, Automation and 
Mechatronics, Singapore, pp. 516-21. 
Pires, J.N., Godinho, T. and Ferreira, P. (2004), “CAD interface for automatic robot 
welding programming” Industrial Robot, Vol. 31 No. 1, pp. 71-6. 
Schaefer, T. and Schraft, D. (2005), “Incremental sheet metal forming by industrial 
robot” Rapid Prototyping Journal, Vol. 11 No. 5, pp. 278-86. 
Video 1 (2010), “Robot program generation from CAD virtual paths”, available at: 
http://robotics.dem.uc.pt/pedro.neto/GS3.html (accessed 15 December 2010) 
Video 2 (2010), “Robot path adjustment – laser camera”, available at: 
http://robotics.dem.uc.pt/pedro.neto/GS4.html (accessed 15 December 2010) 
Video 3 (2010), “Robot path adjustment – force sensor”, available at: 
http://robotics.dem.uc.pt/pedro.neto/GS5.html (accessed 15 December 2010)