micromachines

Review
Analysis of Current Situation, Demand and Development Trend
of Casting Grinding Technology
Haigang Liang 1,2 and Jinwei Qiao 1,2, *

1 School of Mechanicaland Automotive Engineering, Qilu University of Technology (Shandong Academy

of Sciences), Jinan 250353, China
2 Shandong Institute of Mechanical Design and Research, Jinan 250353, China
* Correspondence: [email protected]

Abstract: Although grinding is essential in the production of castings, the casting grinding process in
manufacturing is complicated and there are many difficulties, such as the large amount of noise in the
grinding environment, non-structural casting entities, and the inclination in the overall shape–time
variation. Even in the face of complex technology and a variety of difficulties, modern casting
grinding technology still demands large-batch production, low cost, fast response, thin brittleness,
high precision, etc. The grinding process has a long history. This paper focus on its development
from a human-operated, mechanical job, to an automatic grinding task based on compliant control
theory. However, the methods mentioned above can no longer satisfy the current production
need. In recent years, researchers have proposed intelligent grinding methods to meet the needs
of modern casting production, and provided various strategies and alternatives to the challenges
of machining accuracy, machining efficiency, and surface consistency. The research direction of
casting polishing has mainly focused on online robot detection, material removal prediction, constant
grinding contact force control, and high-precision matching. Although applications for online
detection and constant grinding contact force control exist in industry, there are challenges in material
removal prediction and three-dimensional high-precision matching. This paper also compares and
analyzes the advantages and disadvantages of different grinding methods, and puts forward some
Citation: Liang, H.; Qiao, J. Analysis
of Current Situation, Demand and
research directions for future work, so as to promote more intelligent and efficient grinding of complex
Development Trend of Casting castings in practical application.
Grinding Technology. Micromachines
2022, 13, 1577. https://doi.org/ Keywords: robot grinding; visual control; casting post-processing
10.3390/mi13101577

Academic Editor: Tsz Ho Kwok

Received: 25 August 2022 1. Complexity Description of Casting Grinding Process

Accepted: 16 September 2022 At present, the casting grinding process is faced with many challenges, such as the
Published: 22 September 2022 large amount of noise in the grinding environment, non-structural casting entities, and
Publisher’s Note: MDPI stays neutral the inclination in the overall shape variation, which limit the development of the casting
with regard to jurisdictional claims in grinding process. Therefore, the above problems need to be deeply analyzed.
published maps and institutional affil-
iations. 1.1. Polishing Environment with Large Noise
Casting polishing is the main means of post-processing castings; however, dust gener-
ated in this process cannot be controlled, and large amounts of chromium dust and nickel
particles will cause harm to the environment [1]. Manual grinding causes respiratory tract
Copyright: © 2022 by the authors.
and lung diseases, and even pneumoconiosis, due to dust inhalation. Furthermore, artificial
Licensee MDPI, Basel, Switzerland.
polishing efficiency is low, the consistency of workpiece products is poor, and the scrap rate
This article is an open access article
is high. During mechanical grinding, high-density dust affects the equipment used in the
distributed under the terms and
clamping operation. High-density dust adheres to the equipment, and the dust reduces the
conditions of the Creative Commons
accuracy and stability of the clamping equipment. When an intelligent grinding scheme is
Attribution (CC BY) license (https://
adopted, it is a challenge to seal robots and sensing devices, and thus prevent dust from
creativecommons.org/licenses/by/
4.0/).
entering. In addition, high-density dust blocks the surface of sensing devices, meaning

Micromachines 2022, 13, 1577. https://doi.org/10.3390/mi13101577 https://www.mdpi.com/journal/micromachines

 and thus prevent dust from entering. In addition, high-density dust blocks the su
sensing devices, meaning they cannot accurately make judgements. At the sam
large vibrations have a serious impact on field operations, as shown in Figure
Micromachines 2022, 13, x FOR PEER REVIEW
1.
2 of 25

Micromachines 2022, 13, 1577 2 of 25

and thus prevent dust from entering. In addition, high-density dust blocks the surface of
sensing devices,
they cannot meaning
accurately they cannot
make judgements. Ataccurately make
the same time, largejudgements. At athe
vibrations have same time,
serious
large
impactvibrations have a serious
on field operations, impact
as shown on field
in Figure 1. operations, as shown in Figure 1.

(a) (b)
Figure 1. Casting post-processing site. (a) grinding scene (b) grinding effect (Based on pra
bration can easily occur when grinding the red area of the cavity). During manual grinding
a long period of intense vibration exposure in the grinding process of hand-held workpiece
causes arm vibration syndrome and harm to workers’ physical health [2]. Frequent vibra
also cause (a) harm to the grinding (b) workpiece and tools, because the tools and workpiece are
ble to damage
Figure
Figure 1. Casting
1. Casting during large vibrations.
post-processing
post-processing site.(a)(a)
site. During
grinding
grinding mechanical
scene
scene (b) (b) grinding,
grinding
grinding a large
effect
effect (Based(Based holding forcevi-i
on practice,
on practice,
to fix the
bration
vibration can
cancasting
easily workpiece,
occur
easily occur when and this
whengrinding
grinding thethe force
red red may
areaarea
of theofdamage
the cavity).
cavity). theDuring
During casting part. Agrinding,
manualmanual
grinding, sensing
there devic
is there is
atolong
large
long period
period ofofintense
vibration intense vibration
interference
vibration exposure
cannot
exposure ingrinding
the grinding
inaccurately
the capture
process process
the
of of hand-held
workpiece
hand-held workpieces,
to be
workpieces, which which
polished, and
causes
data that
causes armcontains
arm vibration
vibration a syndrome
syndrome andand
large amount harmharm
oftonoise to workers’
willphysical
workers’ affectphysical
the health
grinding
health [2]. Frequent
accuracy.
[2]. Frequent vibration
Large
vibration will
will vibratio
also
also cause
in acause
largeharmharm to the
to the of
amount grinding
grinding workpiece
noiseworkpiece
for the control and
and tools, tools,
ofbecause because the tools
the tools andand
the end-effector and
workpiece workpiece
have are are
an vulnerable vulnera-
impact on the sti
ble
to to
damagedamage duringduring
large large vibrations.
vibrations. During During mechanical
mechanical grinding,
the equipment. Large vibrations also have a significant impact on the clamping grinding,
a large a large
holding holding
force is forceto
needed is needed
of the wo
to
fix fix
the the casting
casting workpiece,
workpiece, and andforce
this thismayforcedamage
may damage
the thepart.
casting casting
A part. A
sensing sensing
device device
subject to subject
which means the workpiece can easily become loose. Large vibrations can also result in th
to large
large vibration
vibration interference
interference cannot
cannot accurately
accurately capture capture the workpiece
the workpiece to be polished,
to be polished, and acquired
and acquired
ing of hot debris.
data thatcontains
data that contains a large
a large amount
amount of noise
of noise will affect
will affect the grinding
the grinding accuracy.accuracy. Large vibrations
Large vibrations result in result
in a large amount of noise for the control of the end-effector and have an impact on
a large amount of noise for the control of the end-effector and have an impact on the stiffness of thethe stiffness of
the In the
equipment. process
Large of grinding,
vibrations also the
have friction
a between
significant impact the
on grinding
the clamping head
equipment. Large vibrations also have a significant impact on the clamping of the workpiece, which of and
the the
workpiece wo
produces
which
means means
the a the
large
workpiece quantity
workpiece
can easilycan of hot
easily
become debris
become
loose. Large splashing,
loose. Largecan
vibrations as shown
vibrations
also caninin
result theFigure
also result in2.the
splashing ofDuring
splash-
ing
hot of hot
debris. debris.
grinding, a large quantity of thermal debris can block the sight of workers and se
affect
In
theprocess
accuracygrinding,
of the the workpiece. A large quantity of thermal debris having
Inthethe processofof grinding, friction
the between
friction the grinding
between head and
the grinding headtheand
workpiece
the workpiece
temperature
produces aalarge can causeofburns on the exposed skin of workers on site. A large qua
produces largequantity
quantity hot debris
of hot splashing,
debris as shown
splashing, in Figure
as shown 2. During
in Figure manual
2. During manual
uncontrolled
grinding, a largehigh thermal
quantity debris
of thermal can
debris cause
can blockvision
the sightloss of to
grinding, a large quantity of thermal debris can block the sight of workers and seriously
the
workers eyes of workers
and seriously on
affect the accuracy
bris accidentally of the workpiece. A large quantity of thermal debris having a high
affect the accuracyflying of theinto a human
workpiece. eye can
A large seriously
quantity damage
of thermal
temperature can cause burns on the exposed skin of workers on site. A large quantity of
the lens
debris havinganda caus
high
ness [3].
temperature Therefore,
canthermal the
cause burns grinding process is very precise and seriously affects th
uncontrolled high debris oncanthe exposed
cause skintoof
vision loss theworkers on site.onAsite;
eyes of workers large quantity of
debris
efficiency. flying
uncontrolled
accidentally During
highinto mechanical
thermal
a human debris
eye cangrinding,
can the splashing
cause vision
seriously damage loss
theto of
theand
lens eyeshigh-caloric
of workers
cause blindnesson debris wi
[3].site; de-
scalding
bris
Therefore, on
the the
accidentally surface
flying
grinding of athe
into
process is nearby
human
very eyeequipment,
precise canandseriously
seriouslyand
damagean improper
affects thework
the process
causecan
lens efficiency.
and lea
blind-
During
ness [3].mechanical
formation Therefore,
of chip grinding,
the
tumors the
grindingonsplashing
process
the ofis
surface high-caloric
very
of theprecisedebrisand
workpiece. will cause
seriously scalding
affectson
High-caloric the workc
debris
the surface
efficiency.
easily of
damage the nearby
During equipment,
themechanical
outer andthe
layergrinding,
of an improper process of
the splashing
power supply canhigh-caloric
and lead
thetosignal
the formation
debris of the
will
line. For cause
int
chip tumors
scalding on on
the the surface
surface of of
thethe workpiece.
nearby High-caloric
equipment, and debris
an can
improper also easily
process damage
can lead to the
grinding system, high-caloric debris can interfere with the ability of the intelligent
the outer layer of the power supply and the signal line. For the intelligent grinding system,
formation
equipment ofto chip tumors
obtain on the surface
information. aofresult,
the workpiece. High-caloric debris can be also
high-caloric debris can interfere with theAs ability of the the obtained
intelligent information
sensing equipment may to in
easily damage the outer layer of the power supply and the signal line. For the intelligent
and, therefore,
obtain information.not As aapplied.
result, the This
obtainedmay lead to may
information inaccurate pre-judgment
be incorrect and, therefore,and ina
grinding
not system, high-caloric debris can interfere with the abilityplanning
of the intelligent sensing
planning strategies, causing sharp sound pulses at a minor level, orstrategies,
applied. This may lead to inaccurate pre-judgment and inaccurate directly damag
equipment to obtain information. As a result, the obtained
causing sharp sound pulses at a minor level, or directly damaging the grinding equipmentinformation may be incorrect
grinding equipment at a serious level.
and, therefore, not applied. This may lead to inaccurate pre-judgment and inaccurate
at a serious level.
planning strategies, causing sharp sound pulses at a minor level, or directly damaging the
grinding equipment at a serious level.

(a) (b)

Figure2.2.Debris
Figure splash
Debris
(a)
scene.
splash (a) grinding
scene. scene (b)scene
(a) grinding
(b)
grinding
(b)effect (thermal
grinding debris
effect from sanding
(thermal of
debris from s
the red part splashes on the green part).
of the red part splashes on the green part.)
Figure 2. Debris splash scene. (a) grinding scene (b) grinding effect (thermal debris from sanding
of theThe
redmain cause of on
part splashes a sharp sound
the green pulse is excessive grinding. In manual grinding, a
part.)
sudden sharp sound pulse stimulates workers’ pulse
The main cause of a sharp sound is excessive
ears, which can damagegrinding. In cause
hearing and manual grin
sudden
The sharp soundofpulse
main cause stimulates
a sharp workers’
sound pulse ears, which
is excessive canIndamage
grinding. manual hearing
grinding,an
a
hearingsharp
sudden loss,sound
and even
pulsecause perforation
stimulates workers’ofears,
the which
eardrum and deafness,
can damage hearingresulting
and cause
cline in loss,
hearing the quality of cause
and even work.perforation
During mechanical grinding,
of the eardrum the sound
and deafness, can exceed
resulting th
in a de-
cline in the quality of work. During mechanical grinding, the sound can exceed the alarm
Micromachines 2022, 13, 1577 3 of 25
Micromachines 2022, 13, x FOR PEER REVIEW 3 of 25

hearing loss, and even cause perforation of the eardrum and deafness, resulting in a decline
in the
of the equipment,
quality of work.
whichDuring mechanical
can result grinding,
in damage the sound
to workers’ healthcan
andexceed the alarm
equipment of
during
the equipment, which can result in damage to workers’ health and equipment during
the operation of the equipment. For intelligent systems, due to the uncertainty in casting the
aoperation of the
deformation equipment.
structure, For intelligent
the prediction may systems, due to the
not be accurate, anduncertainty in casting
a sudden sharp sounda
deformation structure, the prediction may not be accurate, and a sudden sharp sound
pulse can easily be produced. This can lead to damage to the equipment and the work- pulse
can easily be produced. This can lead to damage to the equipment and the workpiece, and
piece, and serious accidents can be caused.
serious accidents can be caused.
There is an urgent need for industrial robots to realize unmanned casting post-pro-
There is an urgent need for industrial robots to realize unmanned casting post-
cessing because of the large amount of noise in the environment during casting post-pro-
processing because of the large amount of noise in the environment during casting post-
cessing. It is necessary to study and analyze the technology to overcome the challenges of
processing. It is necessary to study and analyze the technology to overcome the challenges
high-density dust, large vibrations, splashing of high-temperature debris, and sharp
of high-density dust, large vibrations, splashing of high-temperature debris, and sharp
sound pulses during the grinding process. In addition to the challenges during the grind-
sound pulses during the grinding process. In addition to the challenges during the grind-
ing process, the non-structural characteristics in the solid design of castings and the time
ing process, the non-structural characteristics in the solid design of castings and the time
variation in the overall inclined shape during the casting process have a serious impact
variation in the overall inclined shape during the casting process have a serious impact on
on the post-processing of castings.
the post-processing of castings.
1.2.
1.2. Non-Structural
Non-Structural Casting
Casting Entities
Entities
Castings
Castings usually have aa three-dimensional
usually have three-dimensional surfacesurface structure,
structure, andand there
there are
are aa large
large
number
numberofofnon-structural
non-structuralconcaveconcave surfaces
surfaces and non-parametric
and non-parametric convex
convexsurfaces. The con-
surfaces. The
cave surface
concave cannot
surface be easily
cannot polished
be easily insideinside
polished by hand.by Due
hand. to the
Duelarge
to thenumber
large of concave
number of
surfaces, during mechanical grinding it is difficult to find the datum
concave surfaces, during mechanical grinding it is difficult to find the datum surface or fix surface or fix the
combination
the combination surface thatthat
surface can can
be clamped.
be clamped. Furthermore,
Furthermore, without
withouta datum
a datum in machining,
in machining, it
is difficult
it is to toensure
difficult ensureaccuracy.
accuracy.InInaddition,
addition,when whenan anintelligent
intelligentgrinding
grinding method
method is is
adopted,
adopted, datadata collection
collectionforfor concave
concave surfaces
surfaces having
having a large
a large curvature
curvature canbeeasily
can easily blocked,be
blocked,
resultingresulting in incomplete
in incomplete data collection.
data collection.
As
As shown
shown in in Figure
Figure 3, 3, the
the casting
casting structure
structure isis complex
complex and and contains
contains aa large
large number
number
of
of features.
features. Among
Amongthese,these,thethesimple
simplefeatures
featuresof ofthe
thecircular
circularsurface
surfacein inthe
thefigure
figure are
are not
not
obvious,
obvious, andand there
there are
are aa large
large number
number of of non-structural
non-structural surfaces
surfaces [4].
[4]. In
Inaddition
addition to tocasting
casting
the convex
the convex outside
outside surface,
surface,therethereare
areaalarge
largenumber
number ofof
internal
internalconcave
concave surfaces, which
surfaces, which are
the the
are most difficult
most structural
difficult concave
structural surface.
concave These
surface. surfaces
These reduce
surfaces the measuring
reduce the measuringaccuracyac-
and theand
curacy grinding accuracyaccuracy
the grinding [5]. The [5].
existing method uses
The existing methodartificial
uses measuring and polishing,
artificial measuring and
and the combination
polishing, of mechanical
and the combination and intelligent
of mechanical methods methods
and intelligent to conduct the processing
to conduct the pro- is
still in the experimental stage.
cessing is still in the experimental stage.

Circular Concave
structure surface
surface
Cnstructured
Convex concave surface
surface

Figure 3. Unstructured cast solid surfaces.

Figure 3. Unstructured cast solid surfaces.
Due to the limitation of the non-structural curved surface casting process, the random
Due to the limitation of the non-structural curved surface casting process, the ran-
interference of casting burrs, flanging, sagging, etc. not only seriously affects the deter-
dom interference
mination of casting
of the datum, butburrs, flanging,
also leads to a sagging, etc. not
sharp change inonly seriously
the load affects
during the de-
the process
termination of the datum, but also leads to a sharp change in the load during
of grinding. Therefore, it is very important to accurately obtain 3D structure information the process
of
of grinding.
machinedTherefore,
parts priorit to
is very important
grinding. to accurately
The obtained obtain 3D
3D structure structure information
information can then be
of machined parts prior to grinding. The obtained 3D structure
processed, and a planning grinding strategy can finally be adopted according toinformation canthe
then be
actual
processed, and awill
situation. There planning
also begrinding strategy
a time change canoverall
in the finally shape
be adopted
of theaccording to the actual
casting tendency in the
situation.
process of forming, which will increase the difficulty in the post-processing oftendency
There will also be a time change in the overall shape of the casting in
the casting.
the process of forming, which will increase the difficulty in the post-processing of the
casting.
 When manual grinding, to ensure the overall shape of the casting subject to tilt time
change, workers need to judge whether deformation exists, and the form of deformation
according to measurements, and then use a means of material removal for post-processing
of the casting. The removal amount and removal method are determined according to the
Micromachines 2022, 13, 1577 workers’ measurement and past experience. Because a certain deformation exists, exces- 4 of 25
sive removal will directly lead to product scrapping. In mechanical grinding, it is also
necessary to judge this situation according to the measured deformation. When removing
materials for deformation correction, parameters of mechanical equipment need to be
1.3. Tendency for the Overall Shape to Change over Time
modified. However, repeated modification of the equipment parameters and the defor-
mationAsjudgment
the number of uses of
of castings sand
will molds
reduce theand other molds
working increases,
efficiency and leadthetobonding
productsurface
scrap-
of the mold and casting will change regularly due to wear or adhesion, causing
ping due to errors [6]. Intelligent methods can accurately match and identify tiny defor- thickness
change, tilt
mations andchange, or overall
reach the accuracyscaling. In turn,
of 0.052 mm this
[7]. results
However, in the loss ofand
judging the original datum
predicting the
and the overall deformation of the casting, as shown in Figure 4. There are
grinding mode based on experience, like that of grinding workers, is still in the experi-clear edges and
cornersand
mental in the (a) initial state
verification stage.mold cavity, and the deformation in (b) wear after a short use
of time will appear with the increase in the number of uses. The deformation in (c) after a
longer period of time and (d) failure status is too large and the shape in (a) is quite different.
When the deformation is small, the casting will not be scrapped directly, but it will pose a
challenge to the post-processing of the casting.

（a） （b） （c） （d）

Figure 4.
Figure The change of the outer
4. The outer dimensions
dimensionsof ofthe
themold
moldwith
withthetheuse
useofoftime
time(a)(a)
initial state
initial (b)(b)
state wear
wear
after aafter a short
short periodperiod
of useof(c)use (c) wear
wear after aafter a longer
longer period
period of useof(d)
use (d) failure
failure state. state

When challenges,
Many manual grinding,
such asto ensure
the large the overall
amount shapeinofthe
of noise thegrinding
casting subject to tilt time
environment, the
change, workers need to judge whether deformation exists, and the form of deformation
time variation in non-structural casting entities, and the inclination in the overall shape,
according
restrict theto measurements,
development andcasting
of the then use a means
grinding of material
process. removal forcastings
Non-structural post-processing
and the
of the casting. The removal amount and removal method are determined
overall tendency in the shape are problems in the casting design and production according
process, to
the workers’ measurement and past experience. Because a certain deformation
and a polishing environment subject to a large amount of noise is a problem in the post- exists,
excessive removal
processing willItdirectly
of castings. lead to
is necessary forproduct scrapping.
researchers to improveIn mechanical grinding,
the detection methods it
is also necessary to judge this situation according to the measured deformation. When
during casting and casting post-processing, and to use advanced industrial robots and
removing materials for deformation correction, parameters of mechanical equipment need
sensors combined with advanced algorithms to replace human detection.
to be modified. However, repeated modification of the equipment parameters and the
deformation judgment of castings will reduce the working efficiency and lead to product
scrapping due to errors [6]. Intelligent methods can accurately match and identify tiny
deformations and reach the accuracy of 0.052 mm [7]. However, judging and predicting the
grinding mode based on experience, like that of grinding workers, is still in the experimental
and verification stage.
Many challenges, such as the large amount of noise in the grinding environment, the
time variation in non-structural casting entities, and the inclination in the overall shape,
restrict the development of the casting grinding process. Non-structural castings and the
overall tendency in the shape are problems in the casting design and production process,
and a polishing environment subject to a large amount of noise is a problem in the post-
processing of castings. It is necessary for researchers to improve the detection methods
during casting and casting post-processing, and to use advanced industrial robots and
sensors combined with advanced algorithms to replace human detection.

2. Development History of Grinding Process

2.1. Initial Manual Polishing
In the Stone Age, stone grinding [8] was mainly used to make various knives, stone
axes, and other tools. In the Bronze Age [9], as the first country to employ copper smelting,
China mastered the advanced casting post-processing technology. The rasp was used to
compensate for casting defects, make the casting surface smooth, and make weapons and
tools sharper and more polished [10]. After entering the Iron Age, rotary grinding tools
appeared [11], which provided a reference for subsequent mechanical grinding. With the
 2.1. Initial Manual Polishing
In the Stone Age, stone grinding [8] was mainly used to make various knives, stone
axes, and other tools. In the Bronze Age [9], as the first country to employ copper smelting,
China mastered the advanced casting post-processing technology. The rasp was used to
Micromachines 2022, 13, 1577
compensate for casting defects, make the casting surface smooth, and make weapons and 5 of 25
tools sharper and more polished [10]. After entering the Iron Age, rotary grinding tools
appeared [11], which provided a reference for subsequent mechanical grinding. With the
emergence of iron tools and rotary tools, and the advent of the first Industrial Revolution
emergence
after of iron of
the emergence tools
theand rotary
steam tools,the
engine, and the advent of materials
manufacturing the first Industrial Revolution
were mainly cast
after the emergence of the steam engine, the manufacturing materials were mainly cast iron.
iron. Although casting products have changed, the grinding method still used manual
Although casting products have changed, the grinding method still used manual grinding.
grinding. Siemens developed the generator in 1866, providing technical support for me-
Siemens developed the generator in 1866, providing technical support for mechanical
chanical grinding. In 1914, sandpaper developed by 3M in the United States resulted in a
grinding. In 1914, sandpaper developed by 3M in the United States resulted in a new
new grinding tool for post-processing for castings. Development subsequently entered a
grinding tool for post-processing for castings. Development subsequently entered a period
period of combining artificial and mechanical grinding in the post-processing of castings,
of combining artificial and mechanical grinding in the post-processing of castings, which
which has continued until now.
has continued until now.
To date, few studies have been conducted on manual grinding technology. In 2021,
To date, few studies have been conducted on manual grinding technology. In 2021,
Zhao Jinghui et al. proposed and developed a hand-held rechargeable grinding tool that
Zhao Jinghui et al. proposed and developed a hand-held rechargeable grinding tool that
can assist personnel in maintaining power equipment subject to corrosion of metal parts
can assist personnel in maintaining power equipment subject to corrosion of metal parts
and oxidation and heating of the contact surface, thus improving the grinding speed. In
and oxidation and heating of the contact surface, thus improving the grinding speed. In
2022, Whitmore,
2022, Whitmore, L et
L al. invented
et al. a precision
invented a precision grinding
grinding tool to be
tool used
to be forfor
used manual
manualgrinding
grinding
of of
samples; the tool can be prepared by a 3D printer [12] and can grind
samples; the tool can be prepared by a 3D printer [12] and can grind samples with samples with a a
surface accuracy
surface accuracyof of
upup to to
10 10
µm.µm.
Manual grinding is highly
Manual grinding is highly dependent
dependent onon people, and
people, andits its
operating
operatingobject is mainly
object is mainly
small-batch samples. Manual grinding is far from satisfying the
small-batch samples. Manual grinding is far from satisfying the requirements requirements of high-vol-
of high-
ume and low-cost
volume workpiece
and low-cost grinding,
workpiece and and
grinding, it is itimpossible to avoid
is impossible thethe
to avoid damage
damage caused
caused
bybynoise, vibration,
noise, and
vibration, andscratches in in
scratches thethe
grinding
grinding process [13].
process [13].

2.2.2.2.
Mainstream
MainstreamMechanical
MechanicalGrinding
GrindingProcess
Process
AsAsa result ofofthe
a result thecombination
combinationof ofthe
the rotary grinding
grinding mode
modewithwithaamotor
motorandand pneu-
pneumatic
matic source,
source, pneumatic
pneumatic grinding
grinding equipment
equipment (e.g.,
(e.g., Figure
Figure 5) 5)
andand electric
electric grinding
grinding equip-
equipment
ment (e.g.,
(e.g., Figure
Figure 6) emerged.
6) emerged. There
There areare
twotwo ways
ways of grinding
of grinding with
with thisthis equipment.
equipment. In In
one,
one,
whenwhen grinding
grinding large
large parts,
parts, the workpiece
the workpiece to betopolished
be polished is fixed,
is fixed, and the and the electric
electric grinding
equipment
grinding movesmoves
equipment relative to the to
relative workpiece surface
the workpiece to complete
surface to completegrinding. In the
grinding. other,
In the
when working on small parts, the electric grinding equipment is
other, when working on small parts, the electric grinding equipment is fixed, and thefixed, and the grinding
is carried
grinding out by out
is carried moving the workpiece
by moving to realize
the workpiece the relative
to realize movement
the relative of the of
movement rotating
the
grinding
rotating head. head.
grinding

(a) (b)
Micromachines 2022, 13, x FOR PEER REVIEW 6 of 25
Figure 5. Pneumatic grinding equipment. (a) Pneumatic polishing machine. (b) Pneumatic debur-
Figure
ring 5. Pneumatic grinding equipment. (a) Pneumatic polishing machine. (b) Pneumatic debur-
tool.
ring tool.

(a) (b)
Figure 6. Electric grinding equipment. (a) Angled polishing machine. (b) Table-top grinding
wheel machine.
Figure 6. Electric grinding equipment. (a) Angled polishing machine. (b) Table-top grinding wheel
machine.

In 1875, Brown and Sheeper designed a sawing machine and universal grinder, and
the grinding method combined with mechanical equipment began to appear. This grind-
ing method, combining manual and equipment-based grinding, continues to this day.
Micromachines 2022, 13, 1577 6 of 25

In 1875, Brown and Sheeper designed a sawing machine and universal grinder, and
the grinding method combined with mechanical equipment began to appear. This grinding
method, combining manual and equipment-based grinding, continues to this day. Special
grinding machines have been designed for special parts according to the principle of
the grinding machine [14]. In 1952, Massachusetts Institute of Technology successfully
developed the world’s first Computer Numerical Control (CNC) milling machine [15]. The
emergence of the CNC milling machine resulted in new mechanical grinding equipment
and a new process for post-processing grinding of castings. When a CNC milling machine
is used for casting post-processing, the workpiece to be polished is fixed to the standardized
clamping device in the workspace of the milling machine, and the grinding tool is controlled
by the CNC program for grinding processing [16]. Although the CNC milling machine
can be used in the post-processing grinding of castings, its working space is small and the
machine has poor flexibility. As a substitute for special machine tools, industrial robots are
increasingly being used in the field of grinding. In 1986, Tate, A. R. of MIT used robots to
realize automatic grinding of welds, and controlled the maximum normal force at 40 N
and the maximum frequency of reference force at 2.3 Hz [17]. Later, another researcher,
Peng J et al. [18], designed a passive grinding device, and studied the characteristics of
the grinding process and the influence of the deflection angle in the passive grinding
process. To meet the requirements of grinding complex parts, Guo Wanjin et al. from
Harbin Institute of Technology designed and developed a compound 5-Dof working robot
having large dexterity in its working space and dexterity in its attitude adjustment [19].
Table 1 scores the performance of different grinding methods. The average value
(Ep) of the table is evaluated according to the favorable grinding degree (0–5), where 0
represents very poor and 5 represents very good. Weight W: Stiffness W1 = 0.6, flexibility
W2 = 0.5, workspace W3 = 0.6, generality W4 = 0.7, cost W5 = 0.5.

Table 1. Comparison of conventional grinding machines with robotic grinding systems.

Universal Special Numerical Robot

Grinding Grinding Control Grinding Grinding
Machine Machine Machine System
Stiffness 4 4 4 1
Flexibility 2 0 3 5
Workspace 1 1 2 5
Versatility 2 0 3 5
Cost 3 2 2 4
Synthesis
6.3 4.1 8.2 11.6
Evaluation value

The fuzzy mathematics weighted average evaluation value (Epw ) and comprehensive
evaluation value (Ez ) can be derived as follows:

E pw = E p × W (1)

5
EZ = ∑ (Epi × Wi ) (2)
i =1

The universal grinding machine has good performance in terms of rigidity and cost,
but poor flexibility and a small working space, which are not suitable for the diversified
needs of intelligent manufacturing. The CNC grinder has excellent performance in terms
of stiffness and precision; however, for the processing of large complex surface parts, the
cost of a high-precision CNC grinder is too high, and it is thus not suitable for procurement
by small and medium-sized enterprises. In addition to its low stiffness, the robot grinding
system has outstanding performance in terms of flexibility, workspace, versatility, and cost.
The casting post-processing grinding process of an integrated robot grinding system has
great development potential; however, the main structure design method used by grinding
Micromachines 2022, 13, 1577 7 of 25

robots not only has a long design cycle, but also has poor grinding stability and stiffness,
which is not suitable for product promotion. Therefore, the end-effector with compliant
control theory combined with the grinding mode of industrial robots has been extensively
studied by researchers.

2.3. Polishing Mode Based on Modern Control Theory

In the process of grinding, the change in the force is the key to improving the grinding
performance to accurately control the force and displacement. It is difficult to control the
contact force between the end-effector and the workpiece. Researchers have conducted a
large number of studies on the control of grinding robot manpower, and found that there is
a large interaction force between the end-effector and the workpiece during grinding, and
that the control accuracy of this force will directly affect the machining accuracy [20–23]
and the position control of the robot. Therefore, the simultaneous control of position and
force is a challenge for grinding operations [24]. Researchers in the field of polishing
robots use active compliance control, passive compliance control, and active and passive
compliance control. The common means of active compliance control are power position
hybrid control [25–27] and impedance control [28–31].
In 1993, Perdereau, V proposed a hybrid robot position control scheme. Subsequently,
Zhou et al. proposed a hybrid control strategy of a grinding robot based on adaptive
impedance control [32]. Tian, Y et al. [33] recently designed a fuzzy force controller
that imitates human behavior in the process of rust removal. Subsequently, Zhao et al.
developed a force/bit disturbance rejection control strategy based on fuzzy Proportion
Integration Differentiation (PID). The proposed control strategy can achieve a force control
accuracy of 13.4% for the expected 15 N contact force, and a material removal depth of
0.0362 mm can reach a precision of 1.2 µm [34]. Zhu et al. proposed a dynamic control
method based on a one-dimensional force sensor PID controller. The roughness of the
polished surface Ra < 0.4 µm, the material removal depth is more stable, the deviation
remains at 0.003 mm, and the mean square deviation at 40 N is 0.37 N [32].
Grinding the workpiece to ensure accurate force control will reduce machining effi-
ciency. The control algorithm is much more complex than the passive compliant constant-
force mechanism. In addition, when the end-effector contacts the workpiece or the surface
is irregular, force overshooting will be inevitable and relatively large. Therefore, researchers
have used passive compliant mechanisms to develop end-effectors and conducted a large
number of studies. Mohammad et al. proposed a forced-end-effector design, which was
applied to the robot grinding system to make the grinding tool compliant and reduce the
impact of vibration [35]. Y.M. Li proposed a constant-force mechanism (CFM) based on
the combination of positive and negative stiffness mechanisms. The positive and nega-
tive stiffness of folded beam and bistable beam mechanisms was used to offset the zero
stiffness to produce constant force, as shown in Figure 7. The proposed CFM can produce
a stroke range of 2 mm in a constant-force mode [36], with a maximum of 12.63 N and a
minimum of 12.43 N, and a flatness of 98.41%. Qingsong Xu and his team at the University
of Macau proposed the design of an end-effector based on a constant-force mechanism
for robot grinding. The designed industrial robot drives the end-effector to polish, and
the end-effector passively adjusts the contact force. The precision of the force is ±0.3 N,
resulting in high consistency of the workpiece’s surface quality [37].
From the development of grinding robots mentioned above, it can be clearly seen that
the grinding robot is moving towards standardization. The control force and displacement
accuracy are the main research directions of end-effector design. The use of constant-force
grinding and constant-force clamping to precisely control the force greatly improve the
grinding accuracy and clamping stability. However, due to the limitation of material
properties and the size of the constant-force mechanism, the end-effector has insufficient
load, an overly complex structure, and insufficient plane stiffness when the end-tool moves.
 stiffness to produce constant force, as shown in Figure 7. The proposed CFM can produce
a stroke range of 2 mm in a constant-force mode [36], with a maximum of 12.63 N and a
minimum of 12.43 N, and a flatness of 98.41%. Qingsong Xu and his team at the University
of Macau proposed the design of an end-effector based on a constant-force mechanism for
robot grinding. The designed industrial robot drives the end-effector to polish, and the
Micromachines 2022, 13, 1577 8 of 25
end-effector passively adjusts the contact force. The precision of the force is ±0.3 N, result-
ing in high consistency of the workpiece’s surface quality [37].

Autonomoussanding
Figure7.7.Autonomous
Figure sandingscenario
scenariousing
usingCFM
CFMend-effectors.
end-effectors.

3. Demand
From theAnalysis of Grinding
development Technology
of grinding robotsfor Modern Casting
mentioned above, itProduction
can be clearly seen
Asgrinding
that the a result robot
of theis upgrading of industrialization
moving towards standardization.and
Thesocial progress,
control force andthe market
displace-
demand has increased for a casting grinding process having a large volume, low cost,
ment accuracy are the main research directions of end-effector design. The use of constant- fast
response, high degree of precision, and thin brittleness.
force grinding and constant-force clamping to precisely control the force greatly improve
the grinding accuracy and clamping stability. However, due to the limitation of material
3.1. Casting Scale Increase
properties and the size of the constant-force mechanism, the end-effector has insufficient
load, China’s total
an overly foundrystructure,
complex production reached
and 51.95 million
insufficient tons in 2020
plane stiffness when[38],
theanend-tool
increase
of 6.6%
moves. year-on-year, despite the impact of COVID-19. The casting output and growth rate
in China in recent years and their changes are shown in Figure 8.
3. Demand Analysis of Grinding Technology for Modern Casting Production
3.2. The Cost of Grinding Process Should Be Reduced
As a result of the upgrading of industrialization and social progress, the market de-
In the market, it is necessary to reduce the cost of post-processing of large quantities
mand has increased for a casting grinding process having a large volume, low cost, fast
of castings [39]. The cost can be roughly divided into four parts, as shown in Figure 9:
response, high degree of precision, and thin brittleness.
development and design, production preparation and processing, procurement of raw
materials and purchased parts, and management and sales [40]. The labor cost is not
negligible in the process of grinding. Since the Industrial Revolution, the demand of
enterprises for labor has been a topic of social concern, and technology also has an important
impact on people and enterprises. As a result of the adjustment in the global industrial
structure, especially in the situation of high-end competition in the manufacturing industry
all over the world, boosting the technology dividend instead of the demographic dividend
has become one of the main ways to reduce costs.
 Micromachines 2022, 13, x FOR PEER REVIEW 9 of 25

3.1. Casting Scale Increase

China’s total foundry production reached 51.95 million tons in 2020 [38], an increase
Micromachines 2022, 13, 1577 9 of 25
of 6.6% year-on-year, despite the impact of COVID-19. The casting output and growth
rate in China in recent years and their changes are shown in Figure 8.

Micromachines 2022, 13, x FOR PEER REVIEW 10 of 25

Figure8.8.China’s
Figure China’s casting
casting production
production and
andgrowth
growthrate
rateininrecent
recentyears.
years.

3.2. The Cost of Grinding Process Should Be Reduced

In the market, it is necessary to reduce the cost of post-processing of large quantities
of castings [39]. The cost can be roughly divided into four parts, as shown in Figure 9:
development and design, production preparation and processing, procurement of raw
materials and purchased parts, and management and sales [40]. The labor cost is not neg-
ligible in the process of grinding. Since the Industrial Revolution, the demand of enter-
prises for labor has been a topic of social concern, and technology also has an important
impact on people and enterprises. As a result of the adjustment in the global industrial
structure, especially in the situation of high-end competition in the manufacturing indus-
try all over the world, boosting the technology dividend instead of the demographic div-
idend has become one of the main ways to reduce costs.

Figure9.9.Cost
Figure Costimpact.
impact.

Upgrading the traditional grinding industry with modern equipment is the source of
power and the necessary means of achieving a low cost. This is of great significance for
promoting the progress of grinding technology, improving the quality of workers, im-
Micromachines 2022, 13, 1577 10 of 25

Upgrading the traditional grinding industry with modern equipment is the source of
power and the necessary means of achieving a low cost. This is of great significance for
promoting the progress of grinding technology, improving the quality of workers, improv-
ing the efficiency of casting enterprises, optimizing the industrial structure adjustment,
and promoting the development of the manufacturing industry. At the same time, the
market for a process having a fast response, high precision, and thin brittleness also has
great expectations.

3.3. Technical Challenges of Complex Workpiece Grinding: Fast Response, Thin Brittleness, and
High Precision
Fast response, high precision, and thin brittleness are further demands of the market
on the basis of high volume and low cost, and they are also the aspects that researchers in
the laboratory are tackling at present.
The market requires low cost and mass post-processing of casting parts, and a fast
manufacturing speed can greatly shorten product processing time in batch casting produc-
tion and processing [37]. To achieve a high volume and low cost, a rapid pace of work
is required. In turn, this requires a fast response. Ensuring a fast response is one of the
problems that must be currently solved to improve the working speed. Without a fast
response, it is very difficult to achieve mass production of polished products or low-cost
production. Parts of the surface of the casting do not require high precision and can be
sacrificed to improve speed. Where the casting is assembled with other parts (Figure 10),
high precision is required. At the same time, parts generally processed by casting have a
thin surface, which is a characteristic of casting parts with thin brittleness. However, due
to the use of brittle materials and a complex surface, thin brittleness is more serious. At
present, most grinding technology is aimed at the heavier parts, so thin brittle workpieces
have resulted in new requirements for grinding technology. An accurate and fast response
in force control technology, and a more accurate perception11and
Micromachines 2022, 13, x FOR PEER REVIEW of 25 planning strategy for
grinding workpieces, are required.

Brittle part
Installation fit

Figure
Figure 10. 10.
Complex casting.casting.
Complex
InIn
traditional high-precision
traditional grinding, market
high-precision demandmarket
grinding, has not been met duehas
demand to low
not been met due to low
efficiency, low accuracy, and damage to workers’ health. The market needs modern high-
efficiency, low accuracy, and damage to workers’ health.
precision grinding technology. Modern grinding technology uses visual processing toThe market needs modern high-
precision
achieve grinding
high-precision technology.
grinding; Modern
however, most grinding
of the grinding technology
technology uses visual processing to
on the market
uses 2D vision.
achieve In grinding system,
high-precision it usually
grinding; use a 2D most
however, visual method
of the for grinding technology
grinding the on the market
workpiece, simple data relating to the workpiece can be easily acquired and processed.
uses 2D vision. In grinding system, it usually use a 2D visual
However, if the shape is complex, the acquired data will not be complete, and blocking,
method for grinding the
workpiece, simple data relating to the workpiece can be easily
low precision, and path planning of the equipment will produce interference. The devel- acquired and processed.
opment
However, of 3D if
vision
the has occurred
shape very rapidly,
is complex, theand some researchers
acquired data willhavenot
started inves-
be complete, and blocking, low
tigating the use of 3D vision in casting post-processing. However, problems of slow speed
precision, and path planning of the equipment will produce interference. The development
and low accuracy still exist because only high-precision algorithm processing can be used
ofobtain
to 3D vision has occurred
high-precision data; these very rapidly,
data can andinsome
then be used researchers
the polishing have started investigating
and planning
the
of use of 3Dgrinding
high-precision visionofinworkpieces.
casting post-processing. However,
High-precision calibration problems of slow speed and
and a registration
algorithm are the necessary conditions for high-precision grinding. In recent years, in or-
low accuracy still exist because only high-precision algorithm processing can be used to
der to make up for the deficiency of traditional grinding, researchers have undertaken a
obtain
large high-precision
amount of exploration of data; these
intelligent datamethods.
grinding can then be used in the polishing and planning
of high-precision grinding of workpieces. High-precision calibration and a registration
4. Classification of Intelligent Grinding Methods
algorithm are the necessary conditions for high-precision grinding. In recent years, in order
In order to make up for the shortcomings of traditional grinding methods, and to im-
to make up for the deficiency of traditional grinding, researchers have undertaken a large
prove the grinding efficiency and accuracy of castings, researchers have combined intelli-
amount
gent sensingofdevices
exploration of intelligent
and artificial grinding
intelligence algorithms in methods.
terms of judgment and pre-
diction. The main achievements include grinding methods based on image vision, laser
sensing, data-driven grinding prediction, 2.5D local feature information, and comparison
between a design model and a 3D point cloud.

4.1. Judgment of Grinding Position Based on Image Vision

 Micromachines 2022, 13, 1577 11 of 25

4. Classification of Intelligent Grinding Methods

In order to make up for the shortcomings of traditional grinding methods, and to
im-prove the grinding efficiency and accuracy of castings, researchers have combined
intelligent sensing devices and artificial intelligence algorithms in terms of judgment and
prediction. The main achievements include grinding methods based on image vision, laser
sensing, data-driven grinding prediction, 2.5D local feature information, and comparison
between a design model and a 3D point cloud.

4.1. Judgment of Grinding Position Based on Image Vision

Manual grinding can be based on the actual situation, in which observation and
measurement need to be made after the completion of a process, and the subsequent
grinding strategy can be adjusted after real-time comparison of drawings and the size. With
the extensive use of machine vision, humanoid real-time feedback and intelligent planning
of the grinding strategy can be realized. With the support of artificial intelligence technology,
vision sensors have been applied in a wide range of intelligent applications, which have
played a positive role in inspiring the improvement in the grinding process [41]. Vision
sensors have the advantages of non-contact detection, high precision, strong repeatability,
fast speed, good stability, and low cost [42].
In 1981, Moravec proposed a corner detector for binocular vision image matching [43],
Micromachines 2022, 13, x FOR PEER REVIEW 12 of 25
and Harris proposed the Harris corner operator for image matching. In the early 21st
century, a large number of correlation methods emerged. In 2001, a digital phase-shifting
shadow technique was proposed, which takes only one image: the projection of the ref-
moving the virtual
erence raster line onreference
the surface grating
of theindeformed
its plane object.
[44]. InThe
2010, Mohammadi
phase proposed
shift is calculated bypro-
movingthe
jecting theMoire
virtual reference
grating grating
on the surfacein its plane
of the [44].the
object; In change
2010, Mohammadi
in the shapeproposed
of the object
projecting
surface the Moire
causes grating
the phase on the
change surface
in the gratingof the object;
fringe, andthe thechange in change
specific the shape of the
in the phase
object surface causes the phase change in the grating fringe, and
can be extracted to obtain the three-dimensional in-formation of the object surface [45]. the specific change in In
the phase can be extracted to obtain the three-dimensional in-formation
recent years, there have also been many applications in industrial robots and grinding- of the object sur-
face [45].
related In recent
fields. Fangyears, there havean
et al. proposed also beenvision
active many measurement
applications inframework
industrial robots and
for industrial
grinding-related fields. Fang et al. proposed an active vision measurement framework
robots, which is based on the motion planning technology of base sampling; the pictures
for industrial robots, which is based on the motion planning technology of base sampling;
taken by a camera at a certain frequency can be used for point cloud reconstruction of the
the pictures taken by a camera at a certain frequency can be used for point cloud recon-
measured workpiece [46]. Deng et al. developed an automated robotic repair system,
struction of the measured workpiece [46]. Deng et al. developed an automated robotic
which can
repair system, complete
which candefect detection
complete defectand polishing
detection andwithin
polishing 3 min, and
within realize
3 min, andthe surface
realize
burr
the surface burr removal process in the range of 60 cm × 60 cm and two elliptic rangesaxis
removal process in the range of 60 cm × 60 cm and two elliptic ranges (the long
is about
(the long12 cmis and
axis about the12small
cm and axis
theissmall
7 cm) [47].
axis is 7Tian, Y etTian,
cm) [47]. al. [33]
Y etproposed a method
al. [33] proposed a to
identify
method to rust as a key
identify rust technology in the process
as a key technology of rustofremoval.
in the process RustRust
rust removal. detection is per-
detection
formed automatically
is performed automaticallyby processing
by processinga series of camera
a series of cameraimages,
images,and the
and visual
the visualservo
servocon-
trol framework
control framework for for
thethe
process
processof of
rust removal
rust removalisisshownshownin inFigure
Figure 11.
11.

Camera

Fuzzy Force
Control Control

Grinder

Automated and
Robust Detection

Visual servo
Figure 11. Visual
Figure servosanding
sandingstructure.
structure.

In the process of grinding, the estimation of material removal and tracking of the
grinding track directly affect the grinding accuracy. At present, in order to obtain the ac-
tual material removal amount, the main method uses offline or online measurement to
establish a mathematical prediction model. In recent years, researchers have conducted a
Micromachines 2022, 13, 1577 12 of 25

In the process of grinding, the estimation of material removal and tracking of the
grinding track directly affect the grinding accuracy. At present, in order to obtain the actual
material removal amount, the main method uses offline or online measurement to establish
a mathematical prediction model. In recent years, researchers have conducted a large of
amount research on the estimation of material removal using image vision. Joshi et al. used
a machine vision method to acquire and recognize texture features of surface images on
polished surfaces, and used a regression model based on machine vision parameters to
evaluate surface roughness [48]. Wang et al. used a two-dimensional convolutional neural
network learning algorithm to monitor the material removal method, and extracted the
features of color, texture, and shape from visual signals. These features constitute a two-
dimensional feature matrix as the input parameter, and the material removal rate in the belt
grinding process as the output parameter. This method can be used to predict the material
removal rate of different sand belt specifications and different grinding parameters [49],
and is suitable for the regression prediction of the material removal rate under typical
working conditions.
Li et al. studied the mechanical spark generation of seven metals, namely, Q235 steel,
304 stainless steel, TC4 titanium alloy, 6061 aluminum alloy, H62 bronze alloy, AMAK3 zinc
alloy, and AZ31B magnesium alloy. The relationship between the physical and chemical
properties of friction sparks and their generation was evaluated. For 6061 aluminum alloy,
H62 bronze alloy, AMAK3 zinc alloy, and AZ31B magnesium alloy, no bright friction sparks
were observed at the maximum friction velocity of 12 m/s and the maximum surface
pressure of 3.75 N/mm2 , due to low hardness, high thermal conductivity, low melting
point, and lack of carbon [50].
The results show that the estimation of material removal and tracking of the grinding
track directly affect the grinding accuracy in the grinding process. Only the method based
on machine vision can produce obvious friction sparks for metals having higher hardness,
because soft metals cannot be predicted by the obvious external picture phenomenon;
hence, there are limitations. The data-driven approach of multi-information fusion provides
another approach for prediction with more complete information. The judgment grinding
based on visual sensing can be used for the grinding work of industrial structures, but its
robustness is affected to a certain extent due to the complex grinding environment and
insufficient environmental illumination.

4.2. Judgment of Grinding Position Based on Laser Sensing

In the process of casting grinding, uncertain and strong disturbances, such as tempera-
ture, noise, vibration, dust, and light, are inevitable, and limit the popularization and use
of vision sensors. Laser sensors can make up for some of the above deficiencies, especially
those caused by dust [51].
As early as in the 1970s, Nitzan et al. used the distance and intensity information of
the laser ranging system to describe indoor scenes, and the stability and reliability of laser
ranging were fully verified [52]. Laser scanning technology has been further developed in
the field of surveying and mapping. In the 1980s and 1990s, Kak proposed the installation
of a monocular laser vision sensor at the end of a robot to scan the surface of the object being
measured. Lindstrand studied the laser measurement method for measuring the diameters
of pipes and bars in the steel industry. Early applications were characterized by low
accuracy, slow processing speed, and vulnerability to environmental interference. In recent
years, researchers have studied the industrial applications of laser sensors. Xu Xiaohu
from Huazhong University of Science and Technology used laser sensors to optimize the
traditional hand–eye calibration algorithm and all its aspects [53–57], and established a
hand–eye calibration model based on the tool center coordinates to accurately obtain the
spatial pose relationship between the robot and the laser scanner, as shown in Figure 12.
The fitting error was calculated as F = 0.060 mm. More importantly, the entire automatic
calibration process lasts only 20 s when a proper path is planned, which greatly saves
calibration time. A. Seidel tried to use a cooperative laser profilometer to obtain geometric
 Xiaohu from Huazhong University of Science and Technology used laser sensors to opti-
mize the traditional hand–eye calibration algorithm and all its aspects [53–57], and estab-
lished a hand–eye calibration model based on the tool center coordinates to accurately
obtain the spatial pose relationship between the robot and the laser scanner, as shown in
Figure 12. The fitting error was calculated as F = 0.060 mm. More importantly, the entire
Micromachines 2022, 13, 1577 13 of 25
automatic calibration process lasts only 20 s when a proper path is planned, which greatly
saves calibration time. A. Seidel tried to use a cooperative laser profilometer to obtain
geometric shapes at the clamping position, and used an adaptive milling path planning
method to the
shapes at automatically offset changes
clamping position, in the
and used anposition
adaptiveand shapepath
milling of parts caused
planning by acci-
method to
dents [58].
automatically offset changes in the position and shape of parts caused by accidents [58].

Hand-eye Calibration

Laser Sensor
Industrial Robot
System

Posture Robotic Milling Scan and Analysis

Optimization System

Fixture Turntable Planning Workpiece

and Milling

Fixed
Figure
Figure 12.
12. Robotic
Roboticmilling
millingprocess
process system.
system.

Lasersensors
Laser sensorshave
havealsoalso been
been widely
widely usedused in field
in the the field of grinding.
of grinding. Gao, Y Gao, Y et al.
et al. carried
carried
out out research
research on robotongrinding
robot grinding
technologytechnology for welding
for welding pretreatment
pretreatment technology
technology of large of
large [59],
parts partsand
[59],adopted
and adopted
a lasera profilometer
laser profilometer for on-site
for on-site measurement,
measurement, planning
planning and pro-and
processing.
cessing.
Micromachines 2022, 13, x FOR PEER REVIEW At AUTOMATICA
At AUTOMATICA 20182018 in Munich,
in Munich, Fraunhofer
Fraunhofer IPA IPA
and and
PILZ,PILZ, a member
a member of
14ofofthe
25
the Ros Industrial Consortium, presented an on-site measure–plan–process
Ros Industrial Consortium, presented an on-site measure–plan–process grinding robot, as grinding robot,
as shown
shown in Figure
in Figure 13.13.
OnOn thethe basis
basis ofofprevious
previousresearch,
research,Ge Geetetal.
al.further
furtherproposed
proposed and and
constructedaa robot
constructed robot welding-line
welding-line grinding
grinding system
system based
based onon laser
laser sensors
sensors [60].
[60]. A
A sensing
sensing
device
device and
and aa self-made grinding tool
self-made grinding toolwere
wereintegrated
integratedatatthe
theend
endofofthe
therobot
robotfor
for the
the grind-
grinding
ing operation. After rough grinding, the weld height was kept at about 0.1
operation. After rough grinding, the weld height was kept at about 0.1 mm, and the average mm, and the
average surface roughness after fine grinding
surface roughness after fine grinding was 0.351 µm. was 0.351µm.

Industrial Robot
End-effector

Workpiece

Laser Sensor System

Figure 13. Robot laser scanning and grinding method.
Figure 13. Robot laser scanning and grinding method.
The polishing robot based on a laser sensor has a unique advantage in the face of the
harsh grinding
The environment
polishing robot basedinon
a factory, i.e., it ishas
a laser sensor nota affected by the environmental
unique advantage in the face oflight,
the
and the
harsh precision
grinding can also meet
environment in athe requirements
factory, for affected
i.e., it is not the post-processing of casting light,
by the environmental parts.
The the
and grinding accuracy
precision canmeet
can also be predicted online byfor
the requirements a data-driven prediction
the post-processing ofmethod.
casting parts.
The grinding accuracy can be predicted online by a data-driven prediction method.
4.3. Abrasive Quantity Prediction Based on Data-Driven
Grinding
4.3. Abrasive surfacePrediction
Quantity roughnessBased
is considered to be one of the key indicators of machining
on Data-Driven
quality [61]. However, due to the randomness and complexity of abrasive particle distribu-
Grinding surface roughness is considered to be one of the key indicators of machin-
tion, it is difficult to predict. In order to accurately estimate the roughness of the polished
ing quality [61]. However, due to the randomness and complexity of abrasive particle dis-
surface, it is necessary to obtain the amount of material removed by grinding in many ways.
tribution, it is difficult to predict. In order to accurately estimate the roughness of the pol-
The amount of material removed by grinding is related to multiple parameters [62], such as
ished surface, it is necessary to obtain the amount of material removed by grinding in
feed speed, rotational speed, contact stress, grinding time, workpiece material, geometric
many ways. The amount of material removed by grinding is related to multiple parame-
characteristics of the workpiece, and condition of the grinding head. In order to obtain
ters [62], such as feed speed, rotational speed, contact stress, grinding time, workpiece
the desired amount of removed material, the above parameters should be optimized and
material, geometric characteristics of the workpiece, and condition of the grinding head.
combined [63].
In order to obtain the desired amount of removed material, the above parameters should
be optimized and combined [63].

4.3.1. Model-Based Material Prediction

The most widely used model-based method is the Preston equation, according to
which the material removal rate during grinding is proportional to the pressure and the
Micromachines 2022, 13, 1577 14 of 25

4.3.1. Model-Based Material Prediction

The most widely used model-based method is the Preston equation, according to
which the material removal rate during grinding is proportional to the pressure and the
relative velocity between the medium and the workpiece. However, due to the lack of a
method to measure relative speed and pressure, this empirical law has not been applied to
finishing [64]. Preston’s law is expressed as:

dV 1
= F∗v (3)
dt k
where dV/dt is the surface indent distance (mm/s), t is time, V is indicates volume. 1/k is
the unit of Preston coefficient k, F is the contact force, v is the relative speed between the
workpiece and the tool. [65].
Equation (3) was studied by researchers, and Castillo-Mejia proposed a correlation
expression of the local material removal rate in grinding based on the Preston equation.
Lee et al. [66] established a model of the grinding material removal rate based on the
Preston equation for independently changing pressure and speed in real time. In the
above study, the Preston coefficient, grinding force, and rotational speed were regarded
as constant values. During the grinding process, the grinding head grinds the workpiece,
and the material properties of the workpiece surface change with the accumulation of heat.
Therefore, material removal cannot be accurately predicted using a constant polishing
coefficient, K. Pan R et al. [67] proposed extraordinary k-constructed correction functions
based on interface friction coefficients and verified their validity by experiments. In actual
grinding, the tool and workpiece do not have a single point of contact, but have regional
contact, which is inconsistent with Preston’s hypothesis. Calculations using the classical
Preston hypothesis will lead to inaccurate estimates of material removal [68]. Therefore,
based on Hertz contact theory and a local area grinding model, Wang et al. further predicted
the cutting depth of robot belt grinding [69] and found that when the cutting depth was
about 0.3 mm, the prediction error was less than 3.1%. Compared with the simplified
Hertz theory model, the root mean square value and mean absolute percentage error of
the material removal model proposed by Zhu et al. were reduced from 2.401 to 1.725 and
18.426 to 14.942%, respectively, considering the elastic deformation of contact. As a result of
the widespread use of artificial intelligence, many methods based on data-driven material
removal prediction have been proposed to solve the material removal prediction problem.

4.3.2. Data-Driven Prediction of Material Removal

Due to the complexity of the grinding process, some of its parameters cannot be
accurately detected in real time, which limits the application of model-based methods in
engineering implementation. An increasing number of researchers are using data-driven
methods to predict material removal.
As early as 2005, the data-driven method was applied to predict material removal.
Panda D predicted the material removal rate using an artificial feedforward neural net-
work [70]. Mathew, J. et al. used an artificial neural network to analyze the material removal
amount and established a parameter optimization model [71]. Wang et al. proposed a
material removal prediction algorithm using neural networks and genetic algorithms [72].
In order to ensure the whole process of grinding work detection, it is necessary to build
a real-time monitoring welding seam clearance prediction system. Thus, David Jin Hong
studied a deep learning vision system to automatically detect the grinding end-points of
welds in the grinding process and monitor the geometric changes in the grinding welds.
The deep learning method, which conducts end-to-end processing on a large number of
experimental grinding data, can obtain good material removal prediction results [73]. In or-
der to improve the accuracy of the polishing robot, Zhang et al. combined acoustic sensing
and the XGBoost algorithm to predict the material removal in sand belt polishing [74], with
an average absolute percentage error of 4.373%.
 Micromachines 2022, 13, 1577 15 of 25

When predicting material removal in grinding, the model-based method mainly

focuses on the basic parameters in the grinding process; thus, the model-based method
yields high-precision material removal results. However, many factors affect the prediction
of grinding materials. The data-driven method represented by neural networks provides a
solution for processing and analysis.

4.4. Grinding Method Based on 2.5D Local Feature Information

Two-dimensional image information is concentrated on the plane, and the depth
information provided by robot grinding is not accurate. Therefore, more accurate 2.5D
information has the advantage of representing 3D objects, thus improving the chance of
reliable recognition [75].
In 2008, Verma, A proposed a 2.5D processing feature recognition system. It was used
to screen out 2.5D part features to determine the machining direction [76]. In 2009, Siebert
et al. extended the 2D SIFT algorithm to 2.5D for application [77]; the proposed algorithm
can be directly matched using local features of 3D rotation invariance. Zhang Yuwei et al.
proposed a method to reconstruct the underlying 3D shape from a 2.5D bas-relief, and
optimized the face shape through normal transfer and Poisson surface reconstruction [78].
Micromachines 2022, 13, x FOR PEER REVIEW 16 of 25
Zhang et al. constructed a 2.5D height field for portrait relief to enhance the appearance
of the portrait [79], and the 2.5D technology was applied to the processing of the portrait
relief. The technology can also be applied to the surface grinding of a portrait relief.
Figure
Figure 14a14a
shows
showsthethe
datadata
template of a workpiece
template to be polished
of a workpiece obtained
to be polished by a sens-
obtained by a
ing device;device;
sensing FigureFigure
14b shows the local
14b shows template
the local features
template of the
features workpiece
of the workpieceto be polished;
to be polished;
Figure
Figure 14c
14cshows
shows thethe
approximate
approximate location
locationof of
BB inin
a rough
a rough registration; and
registration; andFigure
Figure 14d
14d
shows
shows thethe precise
precise location
location of BofinBAinthrough
A through fine registration.
fine registration. After registration,
After registration, it is
it is con-
convenient
venient for grinding
for grinding tools totools
planto theplan the machining
machining route of route of the workpiece,
the workpiece, which canwhich
greatly can
greatly improve
improve the machining
the machining accuracy.accuracy. High-precision
High-precision matching
matching is essential
is essential forfor automatic
automatic
grinding.
grinding.

(a) (b)

(c) (d)
Figure 14.14.
Figure Local feature
Local featurebased
basedmatchingMatching
matchingMatchingof of
local features.
local (a)(a)
features. Workpiece data
Workpiece template
data template (b)
(b)Local
Localfeature
featuredata
datatemplate
template(c)
(c)Coarse
Coarsealignment
alignmentpositioning
positioning(d)
(d)Fine
Finealignment
alignmentpositioning.
positioning

The
The above
above grinding
grinding method
method based
based onon 2.5D
2.5D local
local feature
feature information
information has
has lowlow accuracy
accuracy
in the depth direction,
in the depth direction, and and can be used for machining parts with low accuracy
used for machining parts with low accuracy require- requirements.
The disadvantage
ments. The disadvantageof using local information
of using polishing
local information is that is
polishing additional steps are
that additional required
steps are
to derive
required to the surface
derive information,
the surface and, like
information, and,thelike
2Dthe
method, this requires
2D method, representation
this requires repre-
from a separate
sentation specificspecific
from a separate viewpoint. There There
viewpoint. are errors in other
are errors parameters
in other calculated
parameters calcu-by
the by
lated geometric reasoning
the geometric method,
reasoning and the
method, feature
and recognition
the feature systemsystem
recognition based based
on a prompt
on a
requires prompt features in the workpiece. Therefore, 3D vision technology
prompt requires prompt features in the workpiece. Therefore, 3D vision technology must must be used
for deep and accurate machining.
be used for deep and accurate machining.

4.5. Grinding Method Based on Comparison between Design Model and 3D Point Cloud
In the grinding method based on the comparison between the design model and the
3D point cloud, the 3D design model of the workpiece to be polished is subtracted from
the real-time model of the detected workpiece [80], and the difference between the real-
Micromachines 2022, 13, 1577 16 of 25

4.5. Grinding Method Based on Comparison between Design Model and 3D Point Cloud
In the grinding method based on the comparison between the design model and the
3D point cloud, the 3D design model of the workpiece to be polished is subtracted from the
real-time model of the detected workpiece [80], and the difference between the real-time
model and the original 3D digital model represents the part to be polished [81], as shown
in Figure 15. Two sets of data are matched to facilitate the comparison of workpiece defect
errors. Kuss, A proposed a method to detect the workpiece shape deviation to adapt
to the robot grinding process. The method uses the model of product design to design
dimensional tolerance specifications to predict possible changes in the workpiece geometric
model, using the Iterative Closest Point (ICP) method to match each point cloud with the
measurement point cloud from the workpiece [80]. In order to further improve efficiency
and accuracy, Wei proposed a method to automatically evaluate the machining allowance
of casting parts. The scanned point cloud data were aligned with the design model through
Micromachines 2022, 13, x FOR PEER two
REVIEW 17 and
stages of “initial alignment” and “best registration” to find the best registration of 25
evaluate the machining allowance based on the registration results [82].

Figure15.
Figure 15. Comparing
Comparing the
the matching
matching graph.
graph.

In
In terms
terms of workpiece
workpiece grinding,
grinding,Hu Huetetal.
al.developed
developeda robot
a robot deburring
deburring and and cham-
chamfer-
fering system
ing system [83],
[83], ininwhich
whichthe thehuman
humanoperator
operatorcan canselect
select any
any feature on the the Computer
Computer
Aided
AidedDesign
Design (CAD)
(CAD) model
model and and export
export the
the selected
selected feature
feature for
for the
the tool
tool path
path ofof trajectory
trajectory
generation. However, artificial feature selection was inefficient. Zhang
generation. However, artificial feature selection was inefficient. Zhang et al. proposed et al. proposed an
an adaptive
adaptive grinding
grinding method
method forforprecision
precisioncasting
castingofofblades
bladeswith
with geometric
geometric deviation
deviation [5]. [5].
The
Themeasured
measureddata dataofofblades
bladeswere
werematched
matchedwithwiththe
the design
design model,
model, andand the
the corresponding
corresponding
matching
matchingmatrix
matrixwas
wassolved
solvedtotodetermine
determine the position
the position of of
thethe
castcast
blades.
blades.In order
In orderto further
to fur-
improve the accuracy of automatic laser deburring of ceramic cores, Huang
ther improve the accuracy of automatic laser deburring of ceramic cores, Huang et al. et al. proposed
aproposed
point cloud registration
a point cloud method combining
registration method global and local
combining feature
global and information
local feature [84],infor-
and
the final overall error was less than 35 µm.
mation [84], and the final overall error was less than 35 µm.
Due
Due to the uneven
to the uneven deformation
deformationofofcastingcastingparts,
parts, thethe optimal
optimal machining
machining route route is
is un-
unknown. Therefore, the ideal grinding machining route can be
known. Therefore, the ideal grinding machining route can be used as a benchmark toused as a benchmark to
quantitatively
quantitativelymeasure
measurethe theaccuracy
accuracyof ofdifferent
differentgrinding
grindingpaths,
paths,sosoas
astoto determine
determine whichwhich
registration method can grind paths with high precision. The key steps of grinding path
registration method can grind paths with high precision. The key steps of grinding path
generation are shown in Figure 16.
generation are shown in Figure 16.
The method based on the comparison of a design model and 3D point cloud data
has become an effective detection method for many digital design processes. Point cloud
matching is divided into two stages: rough matching and fine matching. Rough matching
algorithms include principal component analysis, four-point congruency, three-dimensional
normal distribution transformation, and local feature description, such as Fast Point Fea-
ture Histogram Feature. The precision of rough matching cannot reach the precision
specified in the manufacturing industry, so the precision must be further improved through
fine matching.
 feature information [84], and the final overall error was less than 35 µm.
Due to the uneven deformation of casting parts, the optimal machining route is
unknown. Therefore, the ideal grinding machining route can be used as a benchmark to
quantitatively measure the accuracy of different grinding paths, so as to determine which
Micromachines 2022, 13, 1577 registration method can grind paths with high precision. The key steps of grinding 17 path
of 25
generation are shown in Figure 16.

Figure16.
Figure 16.Key
Keysteps
stepsin
ingrinding
grindingpath
pathgeneration.
generation.

The traditional fine matching algorithm is represented by the ICP algorithm. The
principle of the algorithm is that the rotation matrix R and translation vector T are obtained
by solving Formula (4) of the ICP algorithm for point set Pi and Xi .

N
2
∑ = ∑ || Xi − ( R• Pi + T )||2 = min (4)
i =1

The algorithm has a good registration effect for point clouds having a high overlap
rate and close initial position, but has shortcomings in terms of the amount of computation
required and the iterative convergence speed. With rough matching, the ICP algorithm can
solve the problems of a low degree of overlap and large difference in initial position. As a
result of the development of artificial intelligence algorithms, researchers have also carried
out a large amount of research on the intelligent registration of the point cloud based on
deep learning.
(A) Traditional point cloud registration method
Based on the traditional registration algorithm, the speed and accuracy of the algo-
rithm are further improved through analysis; the performance principle of the traditional
algorithm is shown in Figure 17 [85].
As early as 1992 Besl, P. J used a general ICP algorithm for accurate and efficient
registration of 3D shapes. An important application scenario of this algorithm is to register
the data repeatedly measured by the sensing device of rigid objects with the ideal geometric
model before shape inspection, so as to improve the accuracy of the sensing data [86]. The
main disadvantages of ICP are slow convergence [87], sensitivity to outliers, missing data,
and partial overlap. In 2013, Pauly M proposed Sparse ICP, which achieved robustness
at the cost of computational speed through Sparse optimization [88]. In 2021, Zhang
proposed a fast convergence robust registration method. It was proved that the classical
point-to-point ICP can be regarded as a domination–minimization (MM) algorithm, and an
Anderson acceleration method was proposed to accelerate its convergence [89]. In addition,
the Welsch function-based robust error measurement method [90] was also introduced to
 The algorithm has a good registration effect for point clouds having a high overlap
rate and close initial position, but has shortcomings in terms of the amount of computation
required and the iterative convergence speed. With rough matching, the ICP algorithm
can solve the problems of a low degree of overlap and large difference in initial position.
Micromachines 2022, 13, 1577 As a result of the development of artificial intelligence algorithms, researchers have
18 also
of 25
carried out a large amount of research on the intelligent registration of the point cloud
based on deep learning.
(A) Traditional
achieve accuracypoint
that cloud registration
was similar method
to or better than that of Sparse ICP, while improving
speedBased
by anonorder of magnitude.
the traditional The fast
registration convergent
algorithm, the robust ICP accuracy
speed and registration method
of the algo-
provides
rithm are theoretical and method
further improved throughsupport forthe
analysis; theperformance
acquisition principle
of perception
of thedata of the
traditional
grinding
algorithmrobot, and in
is shown canFigure
further
17improve
[85]. the complexity of the grinding workpiece.

Point cloud Optimization

Correspondences

Transformation

Micromachines 2022, 13, x FOR PEER REVIEW 19 of 25

introduced to achieve accuracy that was similar to or better than that of Sparse ICP, while
improving speed by an order of magnitude. The fast convergent robust ICP registration
method
Figure 17. provides theoreticalregistration
Optimization-based and method support for the acquisition of perception data of
method.
the grinding robot, and can further improve the complexity of the grinding workpiece.
(B) Intelligent
(B) As early aspoint
Intelligent 1992 cloud
point registration
Besl, P.
cloud method
J used a general
registration methodICP algorithm for accurate and efficient
registration
The of 3D shapes. An importantlearning
application scenario of solve
this algorithm is to register
Themethod
methodbased
basedononend-to-end
end-to-end learningcan canbebeused
usedtoto solvethe
theregistration
registrationprob-
prob-
the
lem data repeatedly measured by the sensing device of rigid objects with the ideal geomet-
lemusing
usingan anend-to-end
end-to-endneural
neuralnetwork,
network,asasshown
shownininFigure
Figure18.
18.The
Theend-to-end
end-to-endlearning
learning
ric model
method before shaperegistration
inspection, problem
so as to improve the accuracy of the sensing data [86].
methodtransforms
transformsthethe registration problem into a regression
into problem,
a regression and
problem, the the
and transforma-
transfor-
The
tion main disadvantages
estimation is embedded of ICP
in are
the slow
neural convergence
network. [87], sensitivity to outliers, missing
mation estimation is embedded in the neural network.
data, and partial overlap. In 2013, Pauly M proposed Sparse ICP, which achieved robust-
ness at the cost of computational speed through Sparse optimization [88]. In 2021, Zhang
Point cloud
proposed End-to-end
a fast convergence robust framework
registration method. It was proved that the classical
point-to-point ICP can be regarded as a domination–minimization (MM) algorithm, and
an Anderson acceleration method was proposed to accelerate its convergence [89]. In ad-
dition, the Welsch function-based robust error measurement method [90] was also

Neural Transformation T
network

.
Figure 18.Learning-based
Figure18. Learning-basedpoint
pointcloud
cloudregistration
registrationalgorithm.
algorithm.

Due
Duetotothe
thelate
latedevelopment
developmentof ofintelligent
intelligentpoint
pointcloud
cloudregistration
registrationmethods,
methods,QiQipro-
pro-
posed
posed a neural network named PointNet in 2017, which provides a unified frameworkfor
a neural network named PointNet in 2017, which provides a unified framework for
applications
applicationssuch
suchasasclassification, segmentation,
classification, segmentation, and scene
and semantic
scene semantic analysis. According
analysis. to
According
this framework,
to this Yasuhiro
framework, AokiAoki
Yasuhiro proposed expanding
proposed the PointNet
expanding and LKand
the PointNet algorithms into a
LK algorithms
single trainable recursive deep neural network in 2019, opening up a new exploration
into a single trainable recursive deep neural network in 2019, opening up a new explora- path
for the application of deep learning in point cloud registration [91]. In 2020, Yuewang
tion path for the application of deep learning in point cloud registration [91]. In 2020, Yue- He
proposed a registration method combining PointNet++ and ICP. PointNet++
wang He proposed a registration method combining PointNet++ and ICP. PointNet++ can can extract
multiple features that are used as a basis for registration, using ICP algorithms to calculate
extract multiple features that are used as a basis for registration, using ICP algorithms to
rotation and translation. Recently, some researchers proposed solutions for intelligent
calculate rotation and translation. Recently, some researchers proposed solutions for in-
registration. Liu et al. proposed a robust point cloud registration method based on deep
telligent registration. Liu et al. proposed a robust point cloud registration method based
learning [92] called Point cloud Deep Cyclic Net, which uses the adjustment network based
on deep learning [92] called Point cloud Deep Cyclic Net, which uses the adjustment net-
on principal component analysis to quickly adjust the initial position between two pieces of
work based on principal component analysis to quickly adjust the initial position between
two pieces of the point cloud. Perez proposed a rigid point cloud registration method
called Point Cloud Registration Learning (PREL) [93], which allows the registration of
point clouds with high displacement or occlusion. The PREL algorithm does not need an
iterative process and estimates point distribution in a non-parametric way. In the highly
Micromachines 2022, 13, 1577 19 of 25

the point cloud. Perez proposed a rigid point cloud registration method called Point Cloud
Registration Learning (PREL) [93], which allows the registration of point clouds with high
displacement or occlusion. The PREL algorithm does not need an iterative process and
estimates point distribution in a non-parametric way. In the highly occluded point set, the
ICP method showed an average root mean square error (RMSE) of 98.8, followed by a Deep
Closest Point of 32.51 and PREL of 0.75.
The study of relevant literature in the field of grinding indicates that the efficiency
of the three-dimensional neural network algorithm is lower than that of the traditional
registration algorithm. To date, although the intelligent registration algorithm has good
performance in terms of accuracy, the calculation time is long and the cost is high, whereas
the traditional algorithm is fast and efficient. As a result of the continuous progress in
intelligent technology and the updating of algorithms, intelligent registration algorithms
have more potential than the traditional registration algorithm.
The design model is used to compare the workpiece to be polished. As a result, some
defects
Micromachines 2022, 13, x FOR PEER to be polished are detected and the situation of the polished workpiece
REVIEW is obtained
20 of 25

in real time by the feedback of the sensing device. Starting from the design source, the
errors generated in other processes in the casting post-processing and polishing process are
errors generated in other processes in the casting post-processing and polishing process
reduced.
are reduced.
5. Summary and Prediction
5. Summary and Prediction
TheThe
previous section can be summarized as shown in Figure 19. As demand changes
previous section can be summarized as shown in Figure 19. As demand changes
andand
modern technology
modern technology advances, grinding
advances, grinding methods
methods are being
are being developed.
developed.

Primitive man Stone

Demand Transformation of Casting
The earliest humans used Age
Grinding Development Process
stone tools for grinding tools
Bronz Qin Dynasty
e Age The earliest use of bronze tools,
Large noise
grinding with a file
1425-1475 Iron
The earliest use of iron tools, Age
the use of iron tools for
grinding, and the emergence of Siemens Non-structural
1866 castings
rotating tools The earliest use of motor grinding,
the first laboratory research and
1875 research out of the laboratory to
provide productivity for production Propensity to change
Universal grinder
overall shape over
CNC milling machine time
1952

Grinding robot
1986
Demand shift
Polishing based on image
1993 visual judgment
Judgment type based on laser
sensor polished
2002
Brittle Inexpensive
Predictive grinding based on
Polishing based on 2004 multi-information fusion
2.5-dimensional local feature Quick response
information
2008
High precision Multitudinous
Polishing based on the
comparison of design
2015 to model and 3D point cloud
present
Recent smart grinding methods

Figure 19. Demand conversion and grinding process development.

Figure 19. Demand conversion and grinding process development.
Different polishing methods have different advantages, but they also have disad-
vantages, as shown in Table 2.

Table 2. Comparison of grinding methods.

Grinding Method Core Technology Advantage Shortcoming Literature

Inefficient and damaging to the
Micromachines 2022, 13, 1577 20 of 25

Different polishing methods have different advantages, but they also have disadvan-
tages, as shown in Table 2.

Table 2. Comparison of grinding methods.

Grinding Method Core Technology Advantage Shortcoming Literature

Inefficient and damaging to
Manual grinding Human experience Flexible [8–12]
the body
Machine tool Poor flexibility and
Mechanical grinding High precision [14–17,19,94]
equipment small space
Grinding based on compliant Force and position High precision and
Poor rigidity [18,32,34–37]
control theory control large working space
Visual perception Limited two-dimensional
Judgmental grinding based Intelligence has a
devices, visual plane space judgment, [33,46,47]
on visual sensing judgment function
judgment algorithms affected by the environment
Laser perception
Judgmental grinding based The equipment is The device acquires
equipment, judgment [45,58–60]
on laser sensing robust and accurate data slowly
algorithm
Only harder materials have
Visual prediction Material removal
Predictive grinding based on better predictions and are
algorithms, advanced prediction, intelligent [48,49]
machine vision greatly disturbed by the
perception devices material removal
environment
Low prediction accuracy and
Predictive grinding based on Predict the model, Material removal
a large amount of training [64,65,95,96]
multi-information fusion training data prediction
data are required
The depth information is
Depth pre-estimation
Grinding based on 2.5D local Has depth inaccurate and the feature
method, feature [76,78,97]
feature information information recognition needs to be set
recognition algorithm
multiple times
The amount of information is
Based on the design model Accurate route
Laser sensors, large, and the amount of
and the 3D point cloud planning with 3D [5,84]
registration algorithms calculation of the intelligent
comparison grinding method information
algorithm is large

According to Table 2, the manual approach can be used to process almost any casting
grinding according to manual experience, which is flexible. However, with the increase
in casting output, manual grinding efficiency is low, and can seriously harm the lungs
and arms of grinding workers. In contrast to manual grinding, mechanical grinding does
not require workers to contact the workpiece, and the system has good rigidity and small
vibration, so the grinding accuracy is high. However, due to its poor flexibility and small
space, only specific castings can be polished. In view of the small workspace and poor
flexibility, industrial robot grinding based on the compliant control theory adopts force
and position control for grinding. In this approach, grinding precision is high, and the
workspace is large and flexible, but rigidity is poor. In recent years, many intelligent
grinding methods have been developed. Image vision-based grinding position judgment
uses visual perception equipment and visual judgment algorithms, so that the grinding
equipment performs the judgment function; however, this approach is seriously affected
by the environment and is limited to the two-dimensional plane. The laser sensing-based
polishing part judgment method uses a laser sensing device and a judgment algorithm;
this ensures the polishing device has the function of accurate judgment and provides good
robustness, but the data acquisition speed of the device is slow. The data-driven grinding
quantity prediction method uses advanced sensing equipment combined with a visual
prediction algorithm to predict the final grinding effect using the image and force data
in the grinding process; however, the obtained image data are seriously affected by the
Micromachines 2022, 13, 1577 21 of 25

material and environment. The grinding method based on 2.5D local feature information
adopts the combination of a feature recognition algorithm and the depth pre-estimation
method on the basis of the sensing device to achieve grinding to a depth based on part
of the depth information; however, the depth information is not accurate and the feature
recognition needs to be set several times. Based on the design model and the 3D point
cloud contrast grinding method, a laser sensor and registration algorithm can be used to
obtain 3D point cloud data with 3D information, which can provide accurate information
for path planning.
Table 1 shows that industrial robots have advantages in processing large parts. Com-
pared with CNC machine tools, the mechanical arm performs well in terms of cost, and
has large space and good flexibility, but the consistency of feeding is poor. It is difficult to
realize automation of robot processing of large or lightweight workpieces, and the series
structure has low stiffness, resulting in poor stability of the processing process. Therefore,
the parallel grinding robot has great development potential.
In the process of visual recognition, the grinding environment is complex, and sensing
equipment having good adaptability and high accuracy is the key breakthrough direction.
Sensing equipment needs to accurately perceive the position and shape of the workpiece
and other information. After sensing, a high-precision matching visual algorithm can be
used. There is urgent demand for improvement in registration.
The feedback control of the constant grinding contact force is required in high-speed
grinding systems. High-precision control of the grinding force is very important for the
consistency of the grinding surface of complex parts. The application of the constant-force
mechanism in the grinding field provides a new research idea for accurate control of the
grinding force.
In the process of grinding, the amount of material removed directly affects the grinding
accuracy. In order to obtain an accurate estimate of the amount of material removed,
offline measurement is needed to establish a prediction model. At present, the prediction
model has low accuracy and produces a serious environmental impact. A model that
can accurately predict material removal can obtain higher efficiency and precision in the
process of material removal by grinding.

Author Contributions: Conceptualization, H.L. and J.Q.; methodology, H.L.; Investigation, H.L.;
resources, J.Q.; writing—original draft preparation, H.L.; writing—review and editing, J.Q.; visu-
alization, H.L.; supervision, J.Q.; All authors have read and agreed to the published version of
the manuscript.
Funding: This research was funded by the Industry-University-Research Collaborative Innovation
Fund (Grant No. 2020-CXY46, 20200104), Key Research and Development Plan of Shandong Province
(Grant No. 2022CXGC010202, 2022CXGC010304), Key Projects of Natural Science Foundation of
Shandong Province (Grant No. ZR2020ME116), and Shandong Provincial Technology Innovation
Guidance Program (Grant No. 2020LYXZ020).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: This work was supported by the Industry-University-Research Collaborative
Innovation Fund (Grant No. 2020-CXY46, 20200104), Key Research and Development Plan of
Shandong Province (Grant No. 2022CXGC010202, 2022CXGC010304), Key Projects of Natural Science
Foundation of Shandong Province (Grant No. ZR2020ME116), and Shandong Provincial Technology
Innovation Guidance Program (Grant No. 2020LYXZ020).
Conflicts of Interest: The authors declare no conflict of interest.
Micromachines 2022, 13, 1577 22 of 25

References
1. Londershausen, T.; Vaupel, K.; Prziwara, P. Influence of Grinding Aids on the Dust Formation Tendency. Chem. Eng. Technik.
2019, 91, 1008–1013. [CrossRef]
2. Wei, N.Y.; Lin, H.S.; Chen, T. The hand-arm vibration syndrome associated with the grinding of handheld workpieces in a
subtropical environment. Int. Occup. Environ. Health Arch. 2021, 94, 773–781. [CrossRef]
3. Ding, H.; Yang, J.; Wang, W.; Guo, J.; Liu, Q. Characterization and formation mechanisms of rail chips from facing grinding by
abrasive wheel. J. Manuf. Process. 2022, 73, 544–554. [CrossRef]
4. Wang, J.; Sama, S.R.; Lynch, P.C. Design and Topology Optimization of 3D-Printed Wax Patterns for Rapid Investment Casting.
Procedia Manuf. 2019, 34, 683–694. [CrossRef]
5. Zhang, M.; Chen, T.; Tan, Y. An adaptive grinding method for precision-cast blades with geometric deviation. Int. J. Adv. Manuf.
Technol. 2020, 108, 2349–2365. [CrossRef]
6. Zeng, L.; Cui, D.; Fu, Y.; Li, L.; Liu, Z. Analysis on Curve Negotiating Ability of Rail Grinder in Grinding State. Sci. Rep. 2022, 12,
11668. [CrossRef]
7. Xie, H.; Li, W.-L.; Yin, Z.-P.; Ding, H. Variance-Minimization Iterative Matching Method for Free-Form Surfaces—Part I, Theory
and Method. IEEE Trans. Autom. Sci. Eng. 2019, 16, 1181–1191. [CrossRef]
8. Di Bella, M.; Italiano, F.; Martinelli, M.C. Archeometric characterization of prehistoric grindstones from Milazzo Bronze Age
settlement (Sicily, Italy). Archaeol. Anthropol. Sci. 2018, 10, 1571–1583. [CrossRef]
9. Sojak, M.; Wawrzczak, M. A study of the stones in the collection of the Mining Museum of Roznava. Res. News Inst. Archaeol.
Slovak Acad. Sci. 2017, 62, 7–36.
10. Li, X.J.; Martinón-Torres, M.; Meeks, N.D. Inscriptions, filing, grinding and polishing marks on the bronze weapons from the Qin
Terracotta Army in China. J. Archaeol. Sci. 2011, 38, 492–501. [CrossRef]
11. Gantsev, V. Rotary Millstones of the 9th–11th Centuries from the Excavations of the Mangup Princely Palace. Vestn. Volgogr. Gos.
Universiteta. Ser. 4. History. Regionovedenie. Intergov. Relat. 2021, 25, 131–141. [CrossRef]
12. Whitmore, L. A precision manual grinding tool for sample preparation. Ultramicroscopy 2022, 233, 113436. [CrossRef] [PubMed]
13. Dong, R.G.; Welcome, D.E.; Xu, X.S. Identification of effective engineering methods for controlling handheld workpiece vibration
in grinding processes. Int. J. Ind. Ergon. 2020, 77, 102946. [CrossRef]
14. Hashimoto, F.; Gallego, I.; Oliveira, J.F.G. Advances in centerless grinding technology. CIRP Ann. 2012, 61, 747–770. [CrossRef]
15. Yang, J.; Wang, L.; Fang, Y. A novel approach to wheel path generation for 4-axis CNC flank grinding of conical end-mills. Int. J.
Adv. Manuf. Technol. 2020, 109, 565–578. [CrossRef]
16. Gotlih, J.; Brezocnik, M.; Balic, J. Determination of accuracy contour and optimization of workpiece positioning for robot milling.
Adv. Prod. Eng. Manag. 2017, 12, 233–244. [CrossRef]
17. Tate, A.R. Closed Loop Force Control for a Robotic Grinding System. Ph.D. Thesis, Massachusetts Institute of Technology,
Cambridge, MA, USA, 1986.
18. Liu, P.; Zou, W.; Peng, J. Designed a Passive Grinding Test Machine to Simulate Passive Grinding Process. Processes 2021, 9, 1317.
[CrossRef]
19. Wanjin, G. Research on Robot for Grinding of Complex Shape Parts; Harbin Institute of Technology: Harbin, China, 2017.
20. Dong, J.; Shi, J.; Liu, C.; Yu, T. Research of Pneumatic Polishing Force Control System Based on High Speed on/off with PWM
Controlling. Robot. Comput. Integr. Manuf. 2021, 70, 102133. [CrossRef]
21. Ji, S.J.; Lei, L.G.; Zhao, J. An adaptive real-time NURBS curve interpolation for 4-axis polishing machine tool. Robot. Comput.
Integr. Manuf. 2021, 67, 102025. [CrossRef]
22. Xing, H.; Ding, L.; Gao, H. Adaptive tracking of axial position for valve-turning with a robot based on impedance control. Chin. J.
Mech. Eng. 2019, 15, 124–134.
23. Fan, C.; Xue, C.; Zhang, L.; Wang, K.; Wang, Q.; Gao, Y.; Lu, L. Design and control of the belt-polishing tool system for the blisk
finishing process. Mech. Sci. 2021, 12, 237–248. [CrossRef]
24. Chiwu, B. Control Study of a Wind Turbine Blade Grinding Robot; Harbin Engineering University: Harbin, China, 2012.
25. Yun, J.T.; Sun, Y.; Li, C.Q. Self-adjusting force/bit blending control based on quantitative factor-scale factor fuzzy-PID bit control.
Alex. Eng. Mag. 2022, 61, 4389–4397. [CrossRef]
26. Muller, D.; Veil, C.; Sawodny, O. Hybrid force/position control for quasi continuum manipulators. AT Autom. Tech. 2020, 68,
854–862. [CrossRef]
27. Xia, G.H.; Li, C.; Zhu, Q.D.; Xie, X. Hybrid Force/Position Control of Industrial Robotic Manipulator Based on Kalman Filter. In
Proceedings of the 2016 IEEE International Conference on Mechatronics and Automation, Harbin, China, 7–10 August 2016; pp.
2070–2075.
28. Ochoa, H.; Cortesao, R. Impedance Control Architecture for Robotic-Assisted Mold Polishing Based on Human Demonstration.
IEEE Trans. Ind. Electron. 2022, 69, 3822–3830. [CrossRef]
29. Kana, S.; Lakshminarayanan, S.; Mohan, D.M. Impedance controlled human-robot collaborative tooling for edge chamfering and
polishing applications. Robot. Comput. Integr. Manuf. 2021, 72, 102199. [CrossRef]
30. Li, J.; Guan, Y.S.; Chen, H.W.; Wang, B.; Zhang, T.; Hong, J.; Wang, D. Real-time normal contact force control for robotic surface
processing of workpieces without a priori geometric model. Int. J. Adv. Manuf. Technol. 2022, 119, 2537–2551. [CrossRef]
Micromachines 2022, 13, 1577 23 of 25

31. Dai, J.J.; Chen, C.Y.; Zhu, R.F. Suppress Vibration on Robotic Polishing with Impedance Matching. Actuators 2021, 10, 59.
[CrossRef]
32. Zhou, H.; Ma, S.; Wang, G. A hybrid control strategy for grinding and polishing robot based on adaptive impedance control. Adv.
Mech. Eng. 2021, 13, 2072279128. [CrossRef]
33. Tian, Y.; Zhang, G.; Morimoto, K. Automated Rust Removal, Rust Detection and Visual Servo Control. Autom. Constr. 2022, 134,
104043. [CrossRef]
34. Zhang, H.Y.; Li, L.; Zhao, J.B. The hybrid force/position anti-disturbance control strategy for robot abrasive belt grinding of
aviation blade base on fuzzy PID control. Int. J. Adv. Manuf. Technol. 2021, 114, 3645–3656. [CrossRef]
35. Mohammad, A.E.K.; Hong, J.; Wang, D. Design of a force-controlled end-effector with low-inertia effect for robotic polishing
using macro-mini robot approach. Robot. Comput. Integr. Manuf. 2018, 49, 54–65. [CrossRef]
36. Ding, B.X.; Zhao, J.Y.; Li, Y.M. Design of a spatial constant-force end-effector for polishing/deburring operations. Int. J. Adv.
Manuf. Technol. 2021, 116, 3507–3515. [CrossRef]
37. Wei, Y.; Xu, Q. Design of a new passive end-effector based on constant-force mechanism for robotic polishing. Robot. Comput.
Integr. Manuf. 2022, 74, 102278. [CrossRef]
38. 2020 China Foundry Industry Data Release. Cast. Technol. 2021, 42, 547–552. [CrossRef]
39. Haitao, N.; Sifu, M.; Zhenjie, Z. Design of a storage device for stamping dies under high-tempo and high-volume production
conditions. Mold Manuf. 2021, 21, 15–18.
40. Shaofei, J. Research on Cost Decision and Control Method and Technology of Mechanical Product Design Process. Ph.D. Thesis,
Zhejiang University, Hangzhou, China, 2003.
41. Horn, B. Robot Vision; The MIT Press: Cambridge, MA, USA, 1986.
42. Norouzi-Ghazbi, S.; Mehrkish, A.; Fallah, M.M.H.; Janabi-Sharifi, F. Constrained visual predictive control of tendon-driven
continuum robots. Robot. Auton. Syst. 2021, 145, 103856. [CrossRef]
43. Moravec, H.P. Rover Visual Obstacle Avoidance. IJCAI 1981, 81, 785–790.
44. Degrieck, J.; Van Paepegem, W.; Boone, P. Application of digital phase-shift shadow moiré to micro deformation measurements of
curved surfaces. Opt. Lasers Eng. 2001, 36, 29–40. [CrossRef]
45. Mohammadi, F.; Madanipour, K.; Rezaie, A.H. Application of digital phase shift moiré to reconstruction of human face. In Proceed-
ings of the 2010 Fourth UKSim European Symposium on Computer Modeling and Simulation, Pisa, Italy, 17–19 November 2010.
46. Fang, T.; Ding, Y. A sampling-based motion planning method for active visual measurement with an industrial robot. Robot.
Comput. Integr. Manuf. 2022, 76, 102322. [CrossRef]
47. Deng, D.; Polden, J.W.; Dong, J.; Tao, P.Y. Sensor Guided Robot Path Generation for Surface Repair Tasks on a Large-Scale
Buoyancy Module. IEEE/ASME Trans. Mechatron. 2018, 23, 636–645. [CrossRef]
48. Joshi, K.; Patil, B. Prediction of Surface Roughness by Machine Vision using Principal Components based Regression Analysis.
Procedia Comput. Sci. 2020, 167, 382–391. [CrossRef]
49. Wang, N.; Zhang, G.; Ren, L.; Li, Y.; Yang, Z. In-process material removal rate monitoring for abrasive belt grinding using
multisensor fusion and 2D CNN algorithm. Int. J. Adv. Manuf. Technol. 2022, 120, 599–613. [CrossRef]
50. Li, C.; Yuan, Y.; Amyotte, P.; Xiao, M.; Zheng, W.; Bu, Y.; Yuan, C.; Li, G. Friction spark generation and incendivity of several
metal alloys. J. Loss Prev. Process Ind. 2021, 70, 104406. [CrossRef]
51. Boillot, J.-P.; Cote, J.-L. Principle of Laser Vision Sensor and Its Application in Arc Welding and Laser Welding; Chinese Mechanical
Engineering Society: Beijing, China, 2002.
52. Nitzan, D.; Brain, A.E.; Duda, R.O. The measurement and use of registered reflectance and range data in scene analysis. Proc.
IEEE 1977, 65, 206–220. [CrossRef]
53. Xu, X.; Zhu, D.; Zhang, H.; Yan, S.; Ding, H. TCP-based calibration in robot-assisted belt grinding of aero-engine blades using
scanner measurements. Int. J. Adv. Manuf. Technol. 2017, 90, 635–647. [CrossRef]
54. Yan, S.; Xu, X.; Yang, Z.; Zhu, D.; Ding, H. An improved robotic abrasive belt grinding force model considering the effects of
cut-in and cut-off. J. Manuf. Process. 2019, 37, 496–508. [CrossRef]
55. Zhu, D.; Xu, X.; Yang, Z.; Zhuang, K.; Yan, S.; Ding, H. Analysis and assessment of robotic belt grinding mechanisms by force
modeling and force control experiments. Tribol. Int. 2018, 120, 93–98. [CrossRef]
56. Xu, X.; Zhu, D.; Zhang, H.; Yan, S.; Ding, H. Application of novel force control strategies to enhance robotic abrasive belt grinding
quality of aero-engine blades. Chin. J. Aeronaut. 2019, 32, 2368–2382. [CrossRef]
57. Xu, X.; Zhu, D.; Wang, J.; Yan, S.; Ding, H. Calibration and accuracy analysis of robotic belt grinding system using the ruby probe
and criteria sphere. Robot. Comput. Integr. Manuf. 2018, 51, 189–201. [CrossRef]
58. Seidel, A.; Gollee, C.; Schnellhardt, T.; Hammer, M.; Dassing, J.; Vogt, R.; Wiese, T.; Teicher, U.; Hellmich, A.; Ihlenfeldt, S.; et al.
Cyber-physical approach toward semiautonomous postprocessing of additive manufactured parts and components. J. Laser Appl.
2021, 33, 12033. [CrossRef]
59. Gao, Y.; Gao, H.; Bai, K.; Li, M.; Dong, W. A Robotic Milling System Based on 3D Point Cloud. Machines 2021, 9, 355. [CrossRef]
60. Ge, J.; Deng, Z.; Li, Z.; Li, W.; Lv, L.; Liu, T. Robot welding seam online grinding system based on laser vision guidance. Int. J.
Adv. Manuf. Technol. 2021, 116, 1737–1749. [CrossRef]
61. Ji, W.; Wang, L. Industrial robotic machining, a review. Int. J. Adv. Manuf. Technol. 2019, 103, 1239–1255. [CrossRef]
62. Rowe, W. Towards High Productivity in Precision Grinding. Inventions 2018, 3, 24. [CrossRef]
Micromachines 2022, 13, 1577 24 of 25

63. Domblesky, J.; Evans, R.; Cariapa, V. Material removal model for vibratory finishing. Int. J. Prod. Res. 2004, 42, 1029–1041.
[CrossRef]
64. Naruse, N.; Uneda, M.; Ishikawa, K. Study on oscillation controlled lapping by small diameter tool, Consideration based on
grinding characteristics and slurry flow considering overhang (machine elements, lubrication, design, production machining,
production system etc.). Nippon. Kikai Gakkai Rombunshu 2009, 75, 2581–2588.
65. Makiuchi, Y.; Hashimoto, F.; Beaucamp, A. Model of material removal in vibratory finishing, based on Preston’s law and discrete
element method. CIRP Ann. 2019, 68, 365–368. [CrossRef]
66. Lee, Y.; Wang, H.; Park, J.; Jeong, H. Experimental investigation of process parameters for roll-type linear chemical mechanical
polishing (Roll-CMP) system. Precis. Eng. 2014, 38, 928–934. [CrossRef]
67. Pan, R.; Zhong, B.; Chen, D.; Wang, Z.; Fan, J.; Zhang, C.; Wei, S. Modification of tool influence function of bonnet polishing
based on interfacial friction coefficient. Int. J. Mach. Tools Manuf. 2018, 124, 43–52. [CrossRef]
68. Dan, J.; Zhang, M. Study on edge-forming mechanism for aspheric surface compliant polishing. China Mech. Eng. 2013, 24, 1587.
69. Wang, W.; Liu, F.; Liu, Z.; Yun, C. Prediction of depth of cut for robotic belt grinding. Int. J. Adv. Manuf. Technol. 2017, 91, 699–708.
[CrossRef]
70. Panda, D.K.; Bhoi, R.K. Artificial Neural Network Prediction of Material Removal Rate in Electro Discharge Machining. Mater.
Manuf. Processes 2005, 20, 645–672. [CrossRef]
71. Somashekhar, K.P.; Ramachandran, N.; Mathew, J. Optimization of Material Removal Rate in Micro-EDM Using Artificial Neural
Network and Genetic Algorithms. Mater. Manuf. Process. 2010, 25, 467–475. [CrossRef]
72. Khalick Mohammad, A.E.; Hong, J.; Wang, D. Polishing of uneven surfaces using industrial robots based on neural network and
genetic algorithm. Int. J. Adv. Manuf. Technol. 2017, 93, 1463–1471. [CrossRef]
73. Nasir, V.; Sassani, F. A review on deep learning in machining and tool monitoring, Methods, opportunities, and challenges. Int. J.
Adv. Manuf. Technol. 2021, 115, 2683–2709. [CrossRef]
74. Gao, K.; Chen, H.; Zhang, X.; Ren, X.; Chen, J.; Chen, X. A novel material removal prediction method based on acoustic sensing
and ensemble XGBoost learning algorithm for robotic belt grinding of Inconel 718. Int. J. Adv. Manuf. Technol. 2019, 105, 217–232.
[CrossRef]
75. Fang, W.; Zheng, L.; He, B.; Wang, Q. Automatic 3D model acquisition for unknown objects based on hybrid vision technology.
Int. J. Precis. Eng. Manuf. 2017, 18, 275–284. [CrossRef]
76. Verma, A.K.; Rajotia, S. A hint-based machining feature recognition system for 2.5D parts. Int. J. Prod. Res. 2008, 46, 1515–1537.
[CrossRef]
77. Lo, T.R.; Siebert, J.P. Local feature extraction and matching on range images, 2.5D SIFT. Comput. Vis. Image Underst. 2009, 113,
1235–1250. [CrossRef]
78. Zhang, Y.W.; Wang, W.; Chen, Y.; Liu, H.; Jie, Z.; Zhang, C. From 2.5D Bas-relief to 3D Portrait Model. Comput. Graph. Forum 2020,
39, 258–268. [CrossRef]
79. Zhang, Y.; Qin, B.; Chen, Y.; Ji, Z.; Zhang, C. Portrait relief generation from 3D Object. Graph. Models 2019, 102, 10–18. [CrossRef]
80. Kuss, A.; Drust, M.; Verl, A. Detection of Workpiece Shape Deviations for Tool Path Adaptation in Robotic Deburring Systems.
Procedia CIRP 2016, 57, 545–550. [CrossRef]
81. Zhu, D.; Feng, X.; Xu, X.; Yang, Z.; Li, W.; Yan, S.; Ding, H. Robotic grinding of complex components, a step towards efficient and
intelligent machining—challenges, solutions, and applications. Robot. Comput. Integr. Manuf. 2020, 65, 101908. [CrossRef]
82. Gao, Y.; Du, Z.; Li, M.; Dong, W. An automated approach for machining allowance evaluation of casting parts. Int. J. Comput.
Integr. Manuf. 2019, 32, 1043–1052. [CrossRef]
83. Hu, J.; Kabir, A.M.; Hartford, S.M.; Gupta, S.K.; Pagilla, P.R. Robotic deburring and chamfering of complex geometries in
high-mix/low-volume production applications. In Proceedings of the 16th IEEE International Conference on Automation Science
and Engineering (CASE), Hong Kong, China, 20–21 August 2020.
84. Huang, W.; Mei, X.; Jiang, G.; Hou, D.; Bi, Y.; Wang, Y. An on-machine tool path generation method based on hybrid and local
point cloud registration for laser deburring of ceramic cores. J. Intell. Manuf. 2021. [CrossRef]
85. Huang, X.; Mei, G.; Zhang, J.; Abbas, R. A comprehensive survey on point cloud registration. arXiv 2021, arXiv:2103.02690.
86. Besl, P.J.; Mckay, H.D. A method for registration of 3-D shapes. IEEE Trans. Pattern Anal. Mach. Intell. 1992, 14, 3. [CrossRef]
87. Rusinkiewicz, S.; Levoy, M. Efficient variants of the ICP algorithm. In Proceedings of the Third International Conference on 3-D
Digital Imaging and Modeling, Quebec City, QC, Canada, 28 May–1 June 2001.
88. Bouaziz, S.; Tagliasacchi, A.; Pauly, M. Sparse Iterative Closest Point. Comput. Graph. Forum 2013, 32, 113–123. [CrossRef]
89. Zhang, J.; Yao, Y.; Deng, B. Fast and Robust Iterative Closest Point. IEEE Trans. Pattern Anal. Mach. Intell. 2021, 44, 3450–3466.
[CrossRef]
90. Paul, W.H.; Roy, E.W. Robust regression using iteratively reweighted least-squares. Commun. Stat. Theory Methods 1977, 6,
1973–2014.
91. Aoki, Y.; Goforth, H.; Srivatsan, R.A.; Lucey, S. PointNetLK, Robust & Efficient Point Cloud Registration using PointNet. In
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 15–20 June 2019;
pp. 7163–7172.
92. Liu, D.; Zhang, Y.; Luo, L.; Li, J.; Gao, X. PDC-Net, Robust point cloud registration using deep cyclic neural network combined
with PCA. Appl. Opt. 2021, 60, 2990. [CrossRef] [PubMed]
Micromachines 2022, 13, 1577 25 of 25

93. Perez-Gonzalez, J.; Luna-Madrigal, F.; Pina-Ramirez, O. Deep Learning Point Cloud Registration based on Distance Features.
IEEE Lat. Am. Trans. 2019, 17, 2053–2060. [CrossRef]
94. Janez, G.; Timi, K.; Karl, G.; Miran, B. Accuracy improvement of robotic machining based on robot’s structural properties. Int. J.
Adv. Manuf. Technol. 2020, 108, 1309–1329. [CrossRef]
95. Guo, W.C.; Wu, C.J.; Ding, Z.S.; Zhou, Q. Prediction of surface roughness based on a hybrid feature selection method and long
short-term memory network in grinding. Int. J. Adv. Manuf. Technol. 2021, 112, 2853–2871. [CrossRef]
96. Yu, T.; Li, Z.; Wu, D. Predictive modeling of material removal rate in chemical mechanical planarization with physics-informed
machine learning. Wear 2019, 426, 1430–1438. [CrossRef]
97. Kim, B.; Choi, M.; Son, S.-W.; Yun, D.; Yoon, S. Vision-force guided precise robotic assembly for 2.5D components in a semistruc-
tured environment. Assem. Autom. 2021, 41, 200–207. [CrossRef]