Enhance Virtual Reality (VR) for advance personalized learning has been listed as the grand engineering challenge [1]. However, current VR model is not deformable to describe the shape change of manufacturing processes, such as subtractive machining or additive layer manufacturing (LM). Further more, current VR model is not precise enough to simulate the measurement of workpiece, which may go to submicron level. This paper reviewed the current industrial needs of virtual computer numerical control (CNC) training, listed the critical technology issues, traced the 20 years evolution of in-process models of CNC machining, and proposed a new solution. Based on the proposed in-process model, a virtual CNC training system has been developed and used in training courses. Comparisons with conventional on-site manual training and VR based e-learning concluded this new approach could greatly improve safety, increase learning efficiency, and save machine and material.

1. Enhance VR with In-process Model for Advance
Personalized Learning of CNC programming and
Operation
Mr. Liu Peiling, Principal Research Engineer
Dr. Liu Kui, Senior Research Engineer
Mr. Wu Hu, Research Engineer
Mr. Shaw Kah Chuan, Senior Research Officer
Singapore Institute of Manufacturing Technology
71 Nanyang Drive Singapore 638075
Tel: 65-67938356 Mobile: 65-90123598 Fax: 65-67916377
plliu@simtech.a-star.edu.sg http://simtech.a-star.edu.sg
Keywords: Modeling & Simulation, In-process Model, CNC Machining, Virtual Training
ABSTRACT
Enhance Virtual Reality (VR) for advance personalized learning has been listed as the
grand engineering challenge [1]. However, current VR model is not deformable to
describe the shape change of manufacturing processes, such as subtractive machining
or additive layer manufacturing (LM). Further more, current VR model is not precise
enough to simulate the measurement of workpiece, which may go to submicron level.
This paper reviewed the current industrial needs of virtual computer numerical
control (CNC) training, listed the critical technology issues, traced the 20 years
evolution of in-process models of CNC machining, and proposed a new solution. Based
on the proposed in-process model, a virtual CNC training system has been developed
and used in training courses. Comparisons with conventional on-site manual training
and VR based e-learning concluded this new approach could greatly improve safety,
increase learning efficiency, and save machine and material.
NOMENCLATURE
B-Rep: Boundary Representation
CNC: Computer Numerical Control
HSM: High Speed Machining
CSG: Constructive Solid Geometry
IPM: In-process model
LM: Layered manufacturing
MM: Multiple machining
M&S: Modeling and simulation
1. Industrial requirements for virtual training
As the cutting edge of industry, CNC machining produces essential inputs for
virtually all types of manufacturing products by providing injection mould, sheet
metal die, casting die, jigs and fixtures, and other special tools. After analyzing the
whole design and manufacturing procedures, it was found that:
 So called high tech revolution of precision engineering, represented by
pervasive use of computer such as CNC control, CAM, HSM, ultra machining,

2. not only reduce the dependence of precision engineering on unskilled workers
(e.g. manual polisher), but also create a pool of demand on skilled workers
(e.g. CNC machinists), especially those who can operate machining knowledge
intensive HSM and ultra precision machining.
 In today's competitive world, not only the latest technologies are needed, but
most importantly, highly-qualified personnel. This is especially true when it
comes to working with CNC machines. Only when CNC machines have been
perfectly mastered are high productivity and exceptional quality guaranteed.
Skilled workers, especially the machinist who can operate High Speed and
Ultra Precision Machine Tool, are playing a key role in current situation.
 HSM requires the machinist to know not only how to operate the machine tool
but also machining knowledge, in order to plan a successful cutting process.
Lacking of skilled worker, especially skilled machinist who can do high speed
and ultra precision machining, is a major problem for local industry. Skilled
machinists are in high demand in Singapore and Asia, especially in China.
 Presently, trainees acquire their operating skills by observing, referring the
operation manual and then operating under the guidance of an experienced
operator. To make training safer, more economical and more effective, there
is an increasing desire to complete initial training away from the operating
environment. The training of skilled machinists is still a slow and manual
process, which need a lot of machine tools and fixtures etc. The machining
job has a traditional image problem of black smith.
 It is recognized that ITE schools, industrial training centers, public education
facilities, machine tool manufacturers and dealers are having increasing-
complexity of training requirement. They require not only training software
but also more comprehensive turnkey solutions for CNC training system.
2. Current Status of Virtual CNC Training
Comparing with CNC simulation applications which are expensive and mature,
Virtual CNC Training System is still primitive. The NC simulator developers are not
very active to provide training system because the training software market is
logically smaller than production software. More important, NC simulator developers
need to revamp graphics engine or geometry kernel to suit education game use.
None of the leading NC simulators has any CNC training capability. This leaves the
development of Virtual CNC Training System to machine tool venders and schools
who do not have expertise to develop a good graphics engine.
The CNC control vendors developed their own training system. For example,
Siemens developed SinuTrain, which is CNC training software. It runs on PC and
suitable for training purposes and self-study as it is for writing programs and
simulation. It serves for writing and simulating NC programs on a PC, based on the
DIN 66025 programming language as well as the products ShopMill, ShopTurn and
ManualTurn+ and language commands for SINUMERIK® 810D, 840D and 840Di controls,
all are Siemens products. Programs written with this software can be used on real
machines. A prerequisite is that the SinuTrain software is adapted to the SINUMERIK
control on which the program is to be executed. This adaptation must be carried out
by specially qualified personnel, e.g. from Siemens. It is important to stick to
Siemens and the machine-tool manufacturer’s instructions when adapting the
software. No liability is accepted by Siemens if these requirements are not adhered.
We evaluated several vender specific Virtual CNC Training System. They have very
good GUI which has been customized to vender’s CNC controller, some even have

3. touch tablet that simulate operation panel. However, the cutting simulation is rough
and primitive, despite the sound and chip flying animation.
The paper [2] discusses the possibility to use VRML for scientific, educational and
even industrial purposes. As an illustration, it showed how VRML can be used as
inexpensive means for simulation of one of the most interesting but also most time
and resource consuming areas of the computer aided manufacturing (CAM) —
machining of complex parts.
There are many research and publications on Virtual CNC Training in last 10 years.
Most of the published graphics engines are based on VRML and Java 3D. They
explored internet based CNC training and remote NC simulation etc, which are
futuristic but not practical at this moment. The computer hardware is very cheap
now that there is no need to run a training system over internet. The remote
graphics over internet is not necessary [3-4]. Some of them use flash movies such as
micro media to do animation, which could only be used to pre-fixed scene.
In VRML, the realization of dynamic material removal during a machining process
remains a problem. Some commercial software such as Deneb’s virtual NC can export
a VRML animation to describe a machining process. Nevertheless, during the cutting
process, the geometry of a workpiece remains unchanged. The reason is that VRML
does not support set operations among geometric objects such as union, intersection
and difference. This makes it difficult to simulate the change in geometry of a
workpiece under cutting.
In layman’s term, the current virtual CNC training systems are educational games
that lacking the realistic feeling of machining, which is quite different from realistic
simulators, such as the flying simulators that are used to train pilot.
The next generation virtual CNC training is to provide knowledge intensive CNC
training, for the future skilled machinist of precision engineering (PE) industry
through pervasive physics-geometrical modeling and simulation of multiple machining
processes, especially high speed and ultra precision machining. To realize this vision,
we need a new in-process model (IPM) that is deformable and precise.
3. In-process Model (IPM) Review
The in-process model represents the state of the product at each step in the
machining process. It is a 3D geometrical construct that reflects the results of the
machining operations. This model allows the user to visually verify that the
machining operations have been defined accurately and that their sequence is
correct. It can be automatically re-generated when there are changes in the product
design, machining parameters or sequence of the operations [5]. In this section, the
geometric representation techniques of IPMs for traditional manufacturing simulation
are presented.
Boundary Representation (B-rep)
The first choice of IPM should be naturally the geometry model used in commercial
Computer Aided Deign (CAD) system, B-rep. The benefits of using the same geometry
model for CAD as the IPM are obvious. The CAD geometry model is matured and
available through CAD development kit, so there is little need to develop a new
geometry model kernel. Sharing a common geometry model with CAD, the IPM
facilitates seamless integration of CAD-CAPP-CAM.
An automatic forging design and manufacture system was developed by the authors
in 1986, in which pre-form forging IPMs were the same as the CAD system
CV/MUDUSA running on VAX-11/750 computer [6-7]. However, the creation of pre-

4. form forging IPMs took days of calculation and often failed due to Boolean operation
failure.
With a great deal of research effort in the last two decades, the B-rep geometry
model has been improved significant in term of Boolean operation stability, but the
B-rep based IPMs are still limited to 2.5-axis milling (Figure 1). Park reported a
prismatic IPM generation method that employed a polygon extrusion algorithm to
sweep a ball-nose cutter [8].
Figure 1: 2.5-axis IPMs
Section Representation
Since the integrated B-rep IPMs can not be created inside a CAD geometry model, a
new, ad-hoc cross-section-wire-frame based approach was proposed in a forging die
CAD/CAM (Computer Aided Manufacturing) system [6-7]. The aim was to use a series
of paralleled cross-section drawing to represent 3D shapes. Figure 2b depicts the
section representation of circled 3D shapes in Figure 2a.
The cross section IPM is widely used in many commercial CAD/CAM systems. I-DEAS
from SDRC uses water level cross-section as an IPM for generative machining. In a
traditional NC programming environment, a significant amount of time is spent trying
to visualize the in-process stock as it goes through various process stages. With I-
DEAS, the wireframe section in-process stock model can be created for downstream
applications such as toolpath generation, process planning, fixture designing, and
clamp positioning.
Figure 2: Section representation of 3D shape
A part can be sectioned along the X, Y and Z axes (Figure 3). The Z-axis section is
usually called water level section. For 3-axis milling, the water level section could
(a)
(b)

5. have many loops, which causes complications in the set operation between sections.
X and Y sections are single half loops and the Z value is unique for every points. Thus
simplifying the set operation considerably.
Figure 3: Section representation
A working system for using IPM in pre-forging design is described in [6-7]. A drawing
sheet with part sections is first created first using the BACIS command language of
CV/MEDUSA CAD system. Since there are many sections in a drawing sheet, each
section wire-frame is assigned to a different layer according to its Y distance, and a
certain number of sections can be looped through layers. Then each cutter section is
moved to its cutter location and compare with the part sections. The overlap
between the cutter section and the part section will be removed from the part
section. A real milling IPM is obtained from the collection of the result sections.
The display of sections is provided by line segments and can be confusing when there
are too many lines, i.e. there is a need to render the IPM as a realistic 3D image. In
order to calculate the surface normal required for rendering, the section wire frame
is divided along the X direction by the same step over of Y direction. A so-called
regulated section is formed to facilitate the calculation of surface normal and
interpolation of points between the sections. A given node in one section is linked to
a node in the next section. A node’s normal can be calculated from the four
neighboring nodes. Figure 4 shows the regulated section representation.
Figure 4: Regulated section representation
The regulated section can also be used to accelerate set operation between cutter
section and part section. Calculation of intersections and trimming between two
sections are time consuming and the re-ordering of the line segments requires more
computing time. This can be improved with the regulated sections, where the line

6. segments are indexed by both cutter section and part section. Only the line segments
with the same index are compared and trimmed, there is mo need to trim two line
segments. If all the line segments fall on the regulated nodes, there is no need to
trim two line segments. The set operation can be simplified to the comparison of two
Z values, which is very fast and stable. Hence, the Z map representation of IPM
emerges [9].
Z Map
If all the section line segments fall on the nodes, the object surface can be
represented by the Z values of the nodes. A map of Z values represents the object
geometry. In t computer language terms, the Z map can be expressed as a two-
dimension array Z[i, j], where i represents the index in X direction and j represents
the index in Y direction. The XY position of the Z map can be calculated by i or j
times grid size.
Figure 5: Needle bed sample of classic Z Map model
The best analogy for a Z map is a needle bed, where needles are uniformly
distributed over the XY plane (Figure 5). The tip of every needle touches the object
surface that it represents. A milling simulation can be seen as the tool cutting
through the needle bed. These needles can be described in mathematical terms as z-
axis aligned vectors, passing through grid points on the XY plane. A Z map
representation can be used effectively for surfaces that are visible looking
“downwards” on the XY plane. Since 3-axis milling parts are composed of surfaces
visible from the z direction, they can be expressed effectively by the Z map
representation. With a Z map representation, the machining process can be
simulated by cutting the Z map vectors with the cutter.

7. Figure 6 shows an example of 3-axis milling simulation system that was developed by
the first author in 1990. The system was using DOS Extender for Z Map and SVGA for Z
Map rendering. The GUI and NC toolpath wireframe display was coded with High C
graphics library. The GUI and mouse control developments were very hard job and
this was not resolved until the arrival of Windows 95 and OpenGL.
Figure 6: Example of Z Map IPM based milling simulation using DOS Extender and
SVGA
The vectors in a Z map have direction and length and are infinitely thin without
volume. The top of each vector, where the Z map and object meet, is just a point
having no shape. Only at this point the Z map and the object meet with each other. Z
map models cannot provide accurate object geometry outside these points. There
are many ways of interpolating the geometry between grid points in order to render
a Z map model. For example, forming a triangle from three neighboring Z values.
It is obvious that the XY resolution of the Z map grid determines the precision of a Z
map model. A finer grid has greater precision but requires increased memory. For a
part of 1m*1m, the size of the Z map is 1000x1000 if the precision is 1mm, but it
increases to 2000x2000 if the precision is 0.5mm. Reducing the model size and
achieving suitable precision becomes a critical issue in a Z map.
One of the solutions is to balance Z precision and XY precision. We used an integer
array to replace the more common floating array of a Z map, which reduces the Z
map size by half. At the same time, this improves the Boolean operation speed
because the comparison of integers is much faster than the comparison of floats. The
memory requirement of a Z map is halved again by compressing the Z map file
section by section, similar to image compression.
Because of the simplicity of its data structure and fast computation time, the Z map
model is used by most commercial CAM software [10]. However, a Z map can not
approximate vertical wall very well since it always has a slope as showed in Figure 7.
This is not a problem for forging die design since there are always draft angles in
forging parts, but it is a serious problem for milling parts since profiling nearly always
creates vertical walls [11].

8. Figure 7: Classic Z Map model with vertical walls
Extended Z Map
Since the precision of the Z map is determined mainly by XY resolution along the
vertical walls, increasing the resolution along these walls while reducing memory is a
key issue. Fortuitously, one important feature of 3-axis milling can be leveraged.
Viewing from the top, the vertical walls only cover a small percentage of the Z
direction projection, so it should be possible to use finer resolution along the vertical
walls while maintaining a rough resolution in the planar area. This was the initial
idea for an extended Z map.
In our research, at least one grid on a z-map is segregated into sub-cells. Only grids
corresponding to intricate features on the surface of an object are assigned sub-cells
to improve the representation of object features. Figure 8a illustrates the plan view
of the z-map grid with sub-cells 52 the front sectional view, while Figure 8b shows
the sectional view.
(a)
(b)
Figure 8: Extend Z Map with sub-cells along vertical walls

9. The size of the grid can be reduced through using sub-cells. But the precision of the
XY dimension is still limited by the size of sub-cells. For a sub-cell of 0.1mm, the
best precision is 0.1mm in XY plane. There is a need to represent XY dimensions
precisely. Instead of using vectors in the sub-cells, we use sticks in the sub-cells that
have volumes and surface geometry. A B-rep surface model can be represented
precisely using a map of B-rep sticks (Figure 9).
Figure 9: Stick method
Milling simulation with stick method involves Boolean operation between cutter and
stick. Figure 10 shows different stick shapes after cutting. The experiments with B-
rep stick model are very slow and a huge B-rep model is created. To simplify stick
and Boolean operation, a polygon is used instead of real surface in a stick cell. The
data structure of a polygon is much simpler than that of a B-rep which needs a group
of complicated pointers to maintain a double wing data structure.
Figure 10: Different shapes of stick elements

10. The real world objects are not always uniform in the XY plane and can be any shape.
Nodes are used to enhance sub-cell precision in object face representations. For
example, one edge of the sub-cell may have two overlapping nodes to represent a
vertical face. The nodes of a sub-cell may not be uniformly distributed over XY
plane. Figure 11 depicts an exploded plan view of a portion of the z-map grid with
nodes 54. Figure 12 illustrates how stick method represents a circular hole and
vertical walls.
Figure 11: Extended sub-cells to approximate vertical wall
Figure 12: Extend Z map to represents a circular hole
Figure 13 shows shop floor examples of extended Z Map IPM based NC simulation and
verification that was developed in Singapore Institute of Manufacturing Technology
and implemented in precision engineering industry for a decade. The detailed
description of the extended Z map IPM can be found in two patents [12].
Figure 13: Extended Z map IPM based NC simulation examples

11. The limitation of the extended Z map is indeed the ‘stick representation’ of the
workpiece, which can only simulate 3 axis machining in a fixed orientation. In
practice, even for a 3 axis milling machine, there is a need to machine on six sides.
So the stick method could be used to simulate 3 axis machining on one side but
cannot be used to train machinist for machining operation, where flip the workpiece
is a must.
4. Proposed In-process Model
The term voxel represents a volume element in space decomposition geometrical
model schema, just like the term pixel denotes a picture element in raster graphics.
Voxelization is the process of converting a 3D object into a voxel model. Volume
graphics, voxelization and volume rendering have attracted considerable research
works in recent years. However, all of this work is directed at the display of volume
data, mainly for medical applications. In this paper a simplified voxel-based IPM is
proposed to simulate the multi-machining processes.
Further analyzing the voxel model, the authors of this paper believe the voxel-
based volume modeling is a very promising approach to the unified IPM for multiple
machining and layered manufacturing simulation. As a natural clone of the LM
technology, the voxel model of an object and the object fabricated using an LM
closely resemble each other since both are made of layers of small cells [13].
Figure 14: Voxel model and its implementation in machining simulation
The memory requirements of the traditional voxel models are enormous. There is
a need to store the voxel array in compressed form and use algorithms that will
operate directly on the compressed data, especially when the material is
homogenous, where internal voxel could represented by boundary voxel extension. It
is possible to convert the voxel array into some other more compact representations
and reconvert them into voxels when required. The original geometric representation
must be kept and voxelization algorithms will be used when it is necessary. This is
especially valuable since design data are mainly generated from conventional CAD
system.
A voxel-based system should be able to update the display at interactive rates.
Current graphics rendering systems cannot provide a level of rendering performance
on voxel models that is comparable to their polygon-rendering performance. Parallel
algorithms and hardware support for volume rendering are the focus of current

12. research efforts. We only render boundary voxel by a patented color list, which
effectively avoid expensive ray-casting of huge internal voxels. The rendering of
voxel model is easily achieved by rendering a points cloud. However, internal voxel
display is not possible with this method and needs more study.
5. Industry Training Applications
Based on this new in-process model, we developed a VR training sysem for
training CAM programmer and CNC machinists – QuickCNC. The students can simulate
the milling process and save the “machined” model for other downstream machining
process. The virtual CNC simulator also allows for different situations to be tested
during training, which would be costly if done on the machines.
We developed a set of 12 different machining samples that can be used to show
the trainee how the generated toolpath works with cutter under various cutting
parameters, which are set by the virtual controller. The trainee can learn many
setups in a very short time with home PC. The same learning process will take days
and shift in a workshop.
It is quite difficulty to show the effect of a wrong setup or NC code on shop floor.
For example, the dial indicator is used in workpiece alignment. In the conventional
training class, its leg could be easily broken by a beginner. The fast moving cutter
also could hurt trainee. QuickCNC could simulate accident through a virtually broken
leg and sound effect. Everyone is safe except a record of errors.
Figure 15: QuickCNC simulates the remaining of stock and scallop height
Figure 16: Machine frame, dial indicator, and control panel

13. Figure 17: QuickCNC machining example for WSQ course
The precision engineering Worker Skill Qualification (PE WSQ) Specialist Diploma
is a joint initiative by the Singapore Institute of Manufacturing Technology (SIMTech)
and the Singapore Workforce Development Agency (WDA) to provide hands-on
training to equip future PE professionals in cutting-edge precision machining
processing technologies. This programme is conducted through a series of lectures,
lab demonstrations and project attachments in selected industrial applications.
However, due to the limited numbers of CNC machine tool CNC machinist in our lab,
it is a great challenge to conduct hands-on training for all the 40 students on evening
session. This was resolved with the implementation of QuickCNC virtual training
system. During the high speed machining course, the WSQ student learned to use
simulator to do virtual machining. 12 sets of high speed machining examples virtually
explained different high speed machining strategy. The student also can learn from
home PC without spending a lot of time at our lab. Furthermore, the student can
control the simulation speed to see the details at certain corner, which is not
possible on the real machine. Using QuickCNC effectively reduced CNC training time
from weeks long to one evening session.
6. The limitation of the system
Lacking of the dynamic simulation limited the effect of virtual CNC training.
During the real machining, cutter always vibrates under different cutting conditions.
Long tool may chatter under too deep cuts, which will generate marks on workpiece
surface. Cutter wears off during long time of cutting.
Lacking of the fun factor in virtual training is still a problem for promoting
machining to video game generation, which may become future machinist and CAM
programmer.
7. Conclusion and Further Development Plan
The 3D representation of IPMs is essential for the integration of various activities
related to manufacturing process planning, toolpath generation, and machine
inspection. With nearly twenty-year research & development on the IPMs for
machining process, various M&S methods including B-rep, section method, Z map,
and patented extended Z map have been studied, which laid a solid foundation of in-
process geometrical model to enhance VR for training. The proposed model is
implemented in QuickCNC CNC simulator, which has been used in WDA WSQ High
Speed machining training course for quick virtual learning of CAM and CNC
operations. QuickCNC can reduce training time and learning curve.
We are developing a realistic machining training that can simulate tool chatter,
cutter deflection, and tool wear.
Author information
Mr. Liu Peiling has been a CAD/CAM researcher and developer for 25 years. His
research interests include geometry modeling, 3D NC tool path generation, high
speed machining simulation, CAM optimization, and mould design application. He

14. filed two patents about geometrical representation and mould cavity design method.
He took part in many industrial projects and one of his developments QuickCNC has
been widely used in Singapore PE industry. He developed Virtual CNC Training Lab
which has been implemented in Singapore Precision Engineering and Tooling
Association (SPETA) for WDA WSQ CNC machining course.

15. List of references
[1] http://www.engineeringchallenges.org
[2] Nikolay Avgoustinov, "VRML as means of expressive 4D illustration in CAM
education", Future Generation Computer Systems 17 (2000) 39–48.
[3 ] Z. M. Qiu1, Y. P. Chen1, Z. D. Zhou1, S. K. Ong2 and A. Y. C. Nee, “Multi-User NC
Machining Simulation over the WWW”, Int J Adv Manuf Technol (2001) 18:1–6 2001
Springer-Verlag London Limited
[4] S. K. Ong, M. A. Mannon, “Virtual reality simulations and animations in a web-
based interactive manufacturing engineering module”, Computers & Education
archive Volume 43 , Issue 4 (December 2004) table of contents, Pages: 361 – 382,
Year of Publication: 2004 ISSN:0360-1315
[5] IMTI, Inc. 2002, “Unit Process Modeling for Affordable Manufacturing”, White
Paper, www.imti21.org
[6] Liu P.L et al, “A New Concept Integrated CAD/CAM System for Complicated Die &
Mold”, Advances in Computer Science Application to Machinery, International
Academic Publisher, 1991.8, ISBN 7-8003-154-3/TH.2, pp.90-95
[7] Liu P.L. et al, “3D Complicated Parts Design Based on the Automatic Shape
Generation”, Chinese Journal of Mechanical Engineering (English Edition), Volume
5 Number 2, 1992, pp88-92.
[8] Sang C. Park, Gopalan Mukundan, Shuxin Gu and Gustav J. Olling, 2003, “In-
process Model Generation for the Process Planning of a Prismatic Part”, Journal of
Advanced Manufacturing Systems, Vol. 2, No. 2, pp. 147–162
[9] Jerard, R. B., Hussaini, S. Z., Drysdale, R. L. and Schaudt, B., “Approximate
methods for simulation and verification of numerically controlled machining
programs”, Visual Computer, 5(4), pp. 329–348, 1989.
[10] Stifter, S., “Simulation of NC machining based on the dexel model: a critical
analysis”, International Journal of Advanced Manufacturing Technology, 10(3),
pp. 149–157, 1995.
[11] Seung Ryol Maenga, Nakhoon Baekc, Sung Yong Shinb, Byoung Kyu Choid, “A Z-
map update method for linearly moving tools”, Computer-Aided Design 35 (2003)
995–1009
[12] Liu P.L et al, 2002, “An object representation method”, WO04032001A1
[13] Chandru, V., Manohar, S., Prakash, C. E., “Voxel-Based Modeling for
Layered Manufacturing”, IEEE Computer Graphics and Applications (1995), v.15 n.6
[14] Jang Donggo, Kim Kwangsoo, and Jung Jungmin ,”Voxel-based Virtual Multi-axis
Machining”, International Journal of Advanced Manufacturing Technology 16(10),
709-713, 2000