Proceedings of the 2009 IEEE
`International Conference on Robotics and Biomimetics
`December 19 -23, 2009, Guilin, China
`
`Construction of Vision-based Manipulation System for 3D
`Industrial Objects
`
`Guangtao Zhao, Ying Jia and Zhicai Ou
`
`Abstract— Many applications call for robots to perform the
`tasks that the robot can accurately grasp and assemble many
`kinds of industrial parts. Using robots to perform these tasks
`can afford human safety as well as the high productivity. In
`this paper, an industrial robot system aiming to assemble the
`industrial parts with high precision is presented. The paper
`focuses on the design of the robot control system and the
`precise extraction of the position-and-orientation of the
`industrial parts. Finally, experiments are presented to show
`the efficiency of the system, in which sets of automotive parts,
`such as axletrees, bearings, pistons and pins are grasped from
`the conveyor and assembled together precisely.
`
`INTRODUCTION
`I.
`The application of vision-based industrial robotic assembly
`system is becoming increasingly widespread, especially in
`the capability of object grasping and assembly. In
`application of industrial assembly line, stability, effective
`manipulation and operation speed are very important to
`complete the whole task. This fact necessitates the high
`performance for the industrial assembly robotic system.
`And a typical industrial robot system should include the
`following functions: object identification, grasping, path
`planning and assembly. Aiming to the above, an industrial
`robot system with high precision is presented and this paper
`is focused on the design of the control system and the
`extraction of 3D position information for the target object.
`In the present industrial application of robotic assembly
`system, works in visual control can be classified into three
`main categories based on the control structure and the
`definition of the error signal. The first is the 3D visual
`control structure[1]. In this structure, the input of the error
`signal is calculated in three-dimension Cartesian coordinate.
`The features extracted from the image are utilized to
`estimate the pose of the target with respect to the camera.
`The second category is the 2D visual control structure [2] [3]
`& [4]. In this structure, the control error function is defined
`in the image space. The control input is defined neither in
`task space coordinates nor in joint coordinates since a
`closed loop scheme is performed directly in image plane.
`For this method, it is necessary to compute the image
`Jacobian matrix which establishes the relationships between
`the changes of
`the object and
`the changes of
`the
`corresponding image features. Different from the above two
`techniques, the third approach which is called 2-1/2-D
`
`Guangtao Zhao is with the Institute of Automation, Chinese Academy
`of Science.
`Ying Jia is with the College of Science, Minzu University of China.
`Zhicai Ou is with the Institute of Automation, Chinese Academy of
`Sciences.
`
`visual control structure[5] is defined by the control error
`function, in part in the image space and in part in the
`Cartesian space.
`Compared with the control structure mentioned above, a
`simplified image-based visual control system will be
`presented in this paper. With the assumption of a fixed
`height between the camera and the object, this visual
`control system provides a fast computation to meet the
`real-time operation requirement. What’s more, the shift
`position of the object in the image frame is proportional to
`that in the world frame.
`On the other hand, the grasping manipulation is one of
`the crucial steps for the assembly system. Considering the
`subsequent assembly operation with high precision, the
`industrial parts need to be grasped with a proper position
`and a proper orientation and this requires the accurate
`localization information for the target object. Therefore, it
`is necessary to precisely figure out the position and
`orientation of the parts to be grasped. Regarding to this,
`much work has been presented. Kefalea [6] presented an
`approach to localize the objects by combining the evidence
`provided by different visual cues, edges and colors. A
`method was presented by Speth et al. [7], who used the
`images of the target from different points of view to recover
`critical 3D information such as size, location and pose.
`Besides using the visual information, Haidacher and
`Hirzinger [8] proposed an approach for object localization
`by utilizing the contact measurements. Despite this, there
`are still problems and challenges for reliability and
`accuracy of the parts localization.
`In this paper, a robotic system for assembling different
`kinds of industrial parts is proposed. The target of this
`system is to grasp the industrial parts, and further to put
`them together with high precision (0.1mm). Considering the
`constraints and requirements of the actual assembly line, a
`simplified and efficient visual control system and a 3D
`grasping model are adopted to achieve fast, stable and
`efficient grasp and assemblage.
`This paper is organized as follows. Section II gives a
`brief introduction to the robotic assembly system. In
`Section III, a simplified and efficient visual control system
`is analyzed and in Section IV the efficient visual-based
`grasping model is presented. Section V gives the other
`important compositions of the system. Experimental results
`are showed in Section VI and conclusions are drawn in the
`last Section.
`
`II. SYSTEM FUNCTIONAL DESCRIPTION
`The practical target of the system is to make the system
`
`978-1-4244-4775-6/09/$25.00 © 2009 IEEE.
`1051
`Authorized licensed use limited to: Sterne Kessler Goldstein Fox. Downloaded on December 28,2023 at 00:59:35 UTC from IEEE Xplore. Restrictions apply.
`
`RoboticVisionTech EX2014
`ABB v. RoboticVisionTech
`IPR2023-01426
`
`

`

`able to search, identify, grasp and assemble the different
`industrial parts by itself. The robotic system consists of a
`six d.o.f. FANUC
`robot
`arm
`endowed with a
`camera-in-hand, a two-fingered hand, a conveyor and some
`designed fixtures. The hardware framework of the system is
`depicted by Fig. 1.
`
`choose the related grasping strategy according to the pose
`information of the object; (4) To control the robot to grasp
`the target parts with proper strategy; (5) To perform the
`corresponding assembly task aiming at different parts. In
`this paper, we focus on the construction of the system and
`the localization of the 3D target parts.
`
`Camera
`
`Robot Hand
`
`Industrial
`Workpieces
`
`(a)
`
`Fixtures
`
`Robot Gripper
`
`(b)
`Fig. 1. The Hardware Framework of the Assembly System. (a) shows
`the robot, the workpieces placed on the conveyor and the camera. (b) the
`fixtures for helping the assembly operation
`In Fig. 2, the whole software framework of the robotic
`assembly system is described. The manipulation process
`includes the following steps: (1) To search above the
`conveyor for the industrial parts and classify them; (2) To
`obtain the position and orientation of the target parts; (3) To
`
`Fig. 2. The framework of the software in the system
`
`III.
`
`
`
` ,( yxI
`
`)
`
`dxdy
`
`
`
` )(tR
`
` (5)
`
`IMAGE MOMENTS AS VISUAL FEATURES FOR THE VISION
`CONTROL STRUCTURE
`In the industrial assembly, the important performance is
`high speed and robustness which make the robot system
`more complicated. In the industrial system, however, there
`are some types of properties which help to make the
`construction of the robot system simple and efficient: (1)
`There are often only a few kinds of industrial parts to be
`identified. (2) The accurate geometric parameters of
`industrial parts are usually known. (3) The industrial parts
`often have distinctive features (holes and corners).
`Image moments were widely utilized for pattern
`recognition in the past. In our assembly system, the image
`moments are used because they can give a generic and
`geometric representation of the target part whatever the
`shape of the part is simple or complex. The key point to use
`the image moments is to try to avoid the image singularities
`that occur when redundant image features are utilized.
`The analytical form of the image Jacobian matrix about
`the image moments was determined in [9]. Here, the visual
`features using image moments are selected to manipulate
`the piston and peg, the bearing and axletree.
`A.
`Image Moments and Their Interaction Matrix
`)(tR
` be the
`For a binary or segmented image, let
`observed part image at time t. Then, the image moment of
`the target part is defined by
`³³
` )(tm
`
`
`ij
` )
` yxI ,(
`
`i yx
`.j
`where
`xg
`10 / mm
` and
`If the centroid of the object are
`00
`yg
`01 / mm
`, the centered moments with respect to the
`00
`centroid of the target object is
`³³
`
`
` )(tP
`
`y
`y
`dxdy
`x
`x
`)
`(
`()
`i
`ij
`g
`g
` )(tR
`
`Here, we take into account the relationship between the
`time variation of the moments and the relative kinematic
`V
` ),( Zv
`
`v
` , ,vvv
`
`[
`
`]
`, where
` and
`screw
`x
`y
`z
`x ZZZZ
`[
`,
`]
`,
` represent
`and
`translational
`the
`y
`z
`rotational velocity, respectively [10]
`PP 
`ij 
`m
`vL
`VL ij
`,
`ijm
`ij
` is the interaction matrix related to the origin
`
`
`
`j
`
` (6)
`
` (7)
`
`ijmL
`where
`ijLP
` the interaction matrix with respect to the
`and
`centered moments.
`
`1052
`Authorized licensed use limited to: Sterne Kessler Goldstein Fox. Downloaded on December 28,2023 at 00:59:35 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`IV. VISION-BASED OBJECT MODELING METHOD
`Position and orientation information of the industrial
`parts are often needed to decide how to grasp them and
`operate them. In addition, the precision of the obtained
`information has a great impact on the grasping speed and
`accuracy of the robot system. Therefore, how to obtain the
`information efficiently and quickly is one of the key issues
`in the system.
`Because of the central perspective projection, it is not
`accurate to directly compute the orientation of the target
`parts based on the image points. In our system, the camera
`lens is always parallel to the conveyer so the points in the
`conveyor plane and its corresponding image points meet the
`relation of simple-ratio invariability. Since that, an arbitrary
`angle constructed by the points in the conveyer plane is
`equal to the corresponding angle constructed by their image
`points. In addition, the accurate geometric parameters of
`industrial parts are known and the camera lens is calibrated
`ahead of time so the distance between the image plane and
`:
`the conveyer is known too. Let
` denote the set of all
`Z the set of
`3-D vertices of the industrial parts,
`cS
`pS
` and
` the
`corresponding image points in the plane,
`plane of the conveyor. The following procedure is adopted:
`:
`}jy
`}ix and {
`1. To classify the point set
` into {
` in
`order to make
`¦
`y S¦
`x
`{ }i
`
`{ }j
`
`. (1)
`
`C
`
` and
`
`CS
`
`i
`j
`Here, the position of image points of
`to change.
`{ }jy
`, apply the following
`2. Aiming at every point in
`3D grasping model to rectify its position.
`Based on the above procedure, an example is analyzed.
`wA and
`wB
`Two vertices
` of a part are like in Fig.3 and
`wA in the
`the robot’s task is to grasp it from the point
`,cT
`cA ,
`vertical direction. As depicted in Fig. 5, the points
`cS , and the points
`,pT
`pA ,
`wB
`cO
` and
` are in the plane
`pS . Additionally,
`pB
`pO
`pO
` is
` and
` are in the plane
`the principal point of the image,
` is the focus,
` is
`LO
`CO
`the intersection point of the optical axis and the plane
`.CS
`
`{ }ix
`
` is not necessary
`
`PT
`
`Py
`PS
`
`PA
`E
`
`PB
`
`PO
`
`LO
`
`Px
`
`The grasping
`direction
`WA
`
`WB
`
`B. Controller Design
`
`Assembly robot
`
`WZ
`
`Target object
`
`WT
`
`WY
`
`WX
`Fig. 3. The camera is not just over the workpiece but should obtain the
`position and orientation of the workpiece at a glance.
`
`The manipulation of the industrial parts is achieved by
`the use of a three-degree-of freedom stage with two
`translation and one rotation motion. To provide visual
`information,
`three moments-based visual features are
`gx
`gy
`selected. They are
` and
`, which are the center of
`gravity of an industrial part in the image, and the object
`D
`. These three visual features are given as
`orientation
`follows:
`
`(cid:3)(8)
`
`¹·
`
`©§
`
`arctan
`
`21
`
`
`D
`
`01
`
`,
`
`mm
`
`
`
`y
`
`g
`
`,
`
`mm
`
`10
`
`
`
`x
`
`g
`
`P
`2
`¸¸
`¨¨
`11
`PP
`
`20
`02
`00
`00
`Through the development of the image moments, we
`obtain the following relationship between the velocity of
`the stage motion and the feature change obtained in image
`coordinates
`
`»»» ¼º
`zyx
`
`vv
`««« ¬ª
`»»» ¼º
`
`g
`
`
`
`g
`
`
`
`««« ¬ª
`
`
`
`»»» ¼º
`
`
`
`gg
`yx
`««« ¬ª
`
` (9)
`
`Z
`
`Zf
`y
`/
`0
`g
`
`Zf
`x
`0
`/
`g
`
`
`D
`0
`0
`1
`gz
` is the depth of the gravity center. Since the
`where
`interaction matrix is upper triangular, it presents partially
`decoupling properties. Therefore, the image singularities
`problem is solved since the jacobian matrix is full rank all
`the time.
`In addition, due to image processing error and model
`computation error, the open-loop visual control is not
`accurate and robust enough
`to meet
`the
`industrial
`requirement. Thus, visual servo which is a close-loop
`control is adopted in our system (showed in Fig. 4 ).
`Desired
`Feature
`
`Transitional Error
`' '
`
`,x
`y
`(
`)
`
`Rotational error
`T'
`
`Work-Piece
`Pose
`
`Visual
`Control
`
`error
`
`+_
`
`Actual Feature
`
`Feature
`Extraction
`
`Captured
`image
`
`Fig. 4. Visual control configuration
`
`CS
`CT
`CA
` Fig. 5. Illustration of obtaining the orientation of the industrial parts
`
`CO
`
`1053
`Authorized licensed use limited to: Sterne Kessler Goldstein Fox. Downloaded on December 28,2023 at 00:59:35 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`
`
`arctan(
`
`Bp
`
`)
`
` (3)
`
`Based on Fig. 5, it is obvious that the image of the line
`pA B . Generally, we have p
`w wA B is
`segment
`the
`
`pA B as the orientation of p
`w wA B . However,
`orientation of
`
`in the assembly system, the correct grasping orientation is
`pT Bp
`identical with the orientation
` and the error angle is
`E.
`the angle
`pT
` is not known. Therefore,
`For this problem, the point
`the key point of this problem is to accurately find out the
`pT
`point
` in the image.
`Based on the geometric knowledge, we find out that
`A O
`O T
`v (2)
`
`
`r
`P
`P
`P P
`i
`A O
`O T
`C C
`C C
`ir
` is related to the distance of camera lens to the
`Here
`conveyer and it can be computed by camera calibration.
`Thus, the grasping direction can be computed as follows.
`
`y
`y
`D
`
`Tp
`
`x
`x
`Tp
`Bp
`Why does the robot system result in the error angle E?
`Through analyzing, the following reason is found out. Due
`to the central perspective projection, the image of the
`pT Bp
`pA B . If the camera focal p
`w wA B is not the
` but the
`
`
`f and the distance between the image plane and
`length is
`E is:
`d
`the conveyor plane is
`, the error angel
`
`

`˜
`
`TB
`TAd
`cw
`cw
`
`

`
`
`2
`TAd
`TA
`(
`cw
`cw
`
` E
`
`arccos(
`
`(
`
`TB
`cw
`
`2
`
`)
`
`˜
`
`)
`
`2
`
`˜
`
`2
`
`)
`
`TO
`(
`c
`c
`
`is known as the hand-eye calibration.
`Based on [11], the hand-eye calibration can be resolved
`by moving the robot more times and observing the
`corresponding motion of the camera and the robot at the
`same time.
`B. Searching for and Matching Objects with the Hausdorff
`Distance
`Before the robot grasps the objective part, the important
`tasks are to recognize them fast and efficiently.
`In our robotic assembly system, the matching method
`based on the Hausdorff distance has been widely applied in
`recent years. This method is quite tolerant of small position
`errors such as those that occur with edge detectors and other
`feature extraction methods. Given two finite point sets
` "
` "
`A
`a
`a
`B
`b
`b
`,
`}p
`, }q
`1{ ,
`1{ ,
` and
`,
`the Hausdorff
`Distance is defined as
`
`
`

`
`
`
`(H A B,
`
`,h A B h B A, ,
`
`
`) max
`
`

`
`
`a b
`,h A B
`
`where
`max min
`.
`
`
`b B
`a A
` In our assembly robotic system, the modified Hausdorff
`distance [12] is used as the matching model to recognize the
`different industrial parts. It is defined as
`1
`
`
`
` BAh( ,
`ba
`
`)
`in
`N
`Bb
`Aa
`A
` is the number of the sets
`
`

`
`

`
`
`
` (10)
`
` (11)
`
`A .
`
`m
`¦
`
`Where
`
`AN
`
`C. The Assembly of the Industrial Parts
`In the system, two kinds of devices for the parts assembly
`are designed to achieve the assembly task. Regarding to the
`assembly of the axletree and bearing, it is enough to only
`apply the mechanical devices. However, it is difficult to
`assemble the piston and peg because of interference fit. In
`this case, the principle of attractive region is used to meet
`the requirement with high precision as 0.1mm.
`
`VI. EXPERIMENTAL RESULTS
`Many experiments about the assembly of automotive
`parts are tested in our designed robotic system. The sets of
`the automotive parts include the bearing-axletree set and the
`piston-peg set. In our system, a Nikon color CCD camera is
`mounted on the robot end-effector. The acquired image is
`704u
`576
` pixels, and contains a lot of distortion and blur.
`According to the specific task, the whole manipulation is
`divided into two stages: the off-line stage and the on-line
`stage.
`z At the off-line stage, the following important steps
`should be completed. They are: (1) the hand-eye
`calibration which finds out the transformation matrix
`between the robot gripper and the camera in order to
`accurately control the hand; (2) to specify the base
`position of the automotive parts to be grasped and
`establish their templates of the objective parts.
`z At the on-line stage, the motion of the robot is
`
`)
`(4)
`Based on the equation (5), the influencing factors are:
`w cA T between the point
`z The vertical distance
`CS .
`and the plane
`cT
`cO
`.
` and
`z The distance of the point
`z The distance between the lens and the conveyor.
`w cA T and
`the
`Through
`the analysis,
`the distance
`cO Tc
`distance
` are proportional to the error angel. However,
`
`the distance between the lens and the conveyor is the
`contrary to the above two points. But, the image of the part
`to be grasped becomes small when the distance become
`longer. Therefore, a proper distance is chosen to meet our
`requirement.
`
`wA
`
`V. THE OTHER IMPORTANT COMPOSITIONS OF THE SYSTEM
`In the assembly system, there are still some important
`compositions except for the whole system configuration.
`They are the following three sections.
`A. Hand-Eye Calibration
`To accurately control the displacements of the robot, the
`relative position and orientation (transformation matrix)
`between the camera and the robot hand must be computed
`before the application of the assembly system. This problem
`
`1054
`Authorized licensed use limited to: Sterne Kessler Goldstein Fox. Downloaded on December 28,2023 at 00:59:35 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`controlled by the closed-loop visual system. The
`manipulation for the robotic system to grasp the
`target parts includes the following four steps: (1) To
`search for the parts to be operated; (2) To recognize
`different kinds of the parts; (3) To obtain the position
`and orientation information of the parts; (4) To
`operate the parts based on the vision feedback and
`pre-defined strategies.
`In this section, the experiments of bearing-axletree and
`piston-peg assembly are taken as the examples so as to
`illustrate the whole assembly system and validate the
`presented vision-based modeling method.
`A. Off-line Preparations before Experiment
`Considering the actual requirement, the CCD camera is
`mounted about 300mm from the conveyor. Respectively,
`we take the base images of the automotive parts: bearings,
`axletrees, pistons and pegs.
`Using the method proposed by [13], we obtain the
`camera
`distortion
`parameters
`which
`are
`
`
` 
`
`k
`e
`k
`e
`and
`. In order
`1.280739
`007
`2.366564
`012
`1
`2
`to finish hand-eye calibration, we move the robot twice and
`observe the changes about the position of the camera and
`the changes of the robot end-effector. So as to obtain a
`unique solution, the two robot end-effector movements
`must have different axes of rotation and their angles of
`S. Let
`1A and
`1B
` be the first
`rotation must not be 0 or
`2A
`movement of the camera and robot end-effector, and let
`2B
`and
` be the second movement of the camera and robot
`1A ,
`, 1B
`2A ,
`end-effector. The numerical values of the
`2B
`are:
`
`X
`
`
`
`ª
`«
`«
`«
`«
`¬
`
`
`0.2167
`
`0.9736
`0.0520
`0
`
`
`
`0.9717
`0.2105
`0.1071
`0
`
`
`0.1405
`0.0840
`0.9865
`0
`
`
`19.3037
`179.3834
`
`11.6820
`1
`
`º
`»
`»
`»
`»
`¼
`
`B. On-line Operation of the Robot System
`Based on the assembly procedure showed in Fig. 2, two
`kinds of assembly experiments are taken as the examples to
`validate the efficiency of the whole assembly system. The
`following is the details.
`1) Searching for and Classifying Industrial Parts
`As mentioned in section IV, the parts recognition method
`based on the modified Hausdorff Distance is applied in the
`searching process. While the robot moves the camera to
`scan the conveyor for finding target parts, the system
`performs template matching to classify the parts.
`2) Obtaining the Orientation by Vision-based Model
`Based on the new 3D grasping model proposed in this
`paper, the orientation of the axletree is obtained. In Fig. 6,
`two axletrees are placed randomly on the conveyor and the
`white lines stands for their orientation accurately through
`the model.
`Through many experiments, the relation between the
`error angle E and the impact factors is proved.
`
`(b)
`(a)
`Fig. 6. Orientation of the axletree by applying the vision-based model.
`The light lines in the two figures indicate the orientations of the axletrees.
`
`3) Obtaining the Position with Image Moments
`In this robotic system, it is important to obtain the
`corresponding position relationship between the present
`position and the base position in their images. Fig. 7 shows
`the centers of the axletree, bearing, and piston in the images.
`In experiments, the error is about 5mm. But it doesn’t
`influence the grasp and assembly of the part.
`
` and
`
`(b)
`(c)
`(a)
`Fig. 7. The center of the automotive parts by applying image moment. (a)
`the axletree. (b) the bearing. (c) the piston
`
`º
`»
`»
`»
`
`»¼
`
`º»
`
`º»»»»¼
`
`
`
`
`
`
`A
`2
`
`
`
`
`
`B
`2
`
`1055
`Authorized licensed use limited to: Sterne Kessler Goldstein Fox. Downloaded on December 28,2023 at 00:59:35 UTC from IEEE Xplore. Restrictions apply.
`
`ª«
`
`««¬
`
`«
`
`ª
`«
`«
`«
`
`«¬
`
`
`
`A
`1
`
`B
`1
`
`ª«
`
`48.4978
`0.0271
`0.9653 0.2598
`
`1.1852
`0.0341
`0.2606
`0.9648
`
`7.8113
`0.9991
`0.0172 0.0400
`1
`0
`0
`0
`0.9659 0.2588 0.0000 0
`0.2588
`0.9659 0.0000 0
`0.0000
`0.0000 1.0000 0
`1
`0
`0
`0
`
`20.8673
`0.9993
`0.0139 0.0333
`«
`
`0.0187
`0.9889 0.1477
`13.2461
`»
`«
`»
`
`0.0308
`23.4041
`0.1482 0.9885
`«
`»
`0
`0
`0
`1
`¬
`¼
`
`ª
`º
`0.9848 0.0000
`0.1736
`20
`»
`«
`0.0000 1.0000
`0.0000
`20
`«
`»
`»
`«
`0.1736 0.0000
`0.9848
`0
`«
`»
`0
`0
`0
`1
`¬
`¼
`XB
`1A X
`two
`Thus, with
`the
`equations
`1
`XB
`A X
`[14], the hand-eye calibration result is:
`2
`2
`
`

`

`success rate reaches to 90 percent when using the feedback
`information once.
`
`VII. CONCLUSION
`In this paper, an industrial robotic assembly system is
`presented to complete the grasp and assembly operation
`with high precision as 0.1mm. In order to achieve the above
`task, a simplified and efficient visual control structure is
`proposed to achieve the fast and robust control. In addition,
`a novel 3D grasping model is designed to accurately obtain
`the position and orientation of the parts. What’s more, the
`presented method is general for different kinds of the 3D
`industrial parts. In the end, the experiments about grasping
`and assembling of the bearing and axletree are presented to
`prove the efficiency of the designed robotic assembly
`system.
`
`[7]
`
`REFERENCES
`[1] W.J. Wilson, C.C. Williams Hulls and G.S. Bell, “Relative
`End-effector Control Using Cartesian Position Based Visual
`Servoing”, IEEE Transactions on Robotics and Automation, vol. 12,
`no. 5, pp. 684-696. 1996
`[2] N.J. Cowan and D. E. Chang, “Geometric Visual Servoing”, IEEE
`Transactions on Robotics, vol. 21, no. 6, pp. 1128-1138. 2005
`[3] P.I. Corke and S.A. Hutchinson, “A New Partitioned Approach to
`Image-Based Visual Servo Control”, IEEE Transactions on Robotics
`and Automation, vol. 17, no. 4. pp. 507-515. 2001,
`[4] R. Kelly, R. Carelli, O. Nasisi, B. Kuchen and F. Reyes, “Stable
`Visual Servoing of Camera-in-Hand Robotic Systems”, IEEE/ASME
`Transactions on Mechatronics, vol. 5. no. 1, pp. 39-48. 2000
`[5] E. Malis, F. Chaumette and S. Boudet, “2½D Visual Servoing”, IEEE
`Transactions on Robotics and Automation, vol. 15, no.2 pp. 238-250.
`1999
`[6] E. Kefalea, “Object Localization and Recognition for a Grasping
`Robot”, IECON '98, vol. 4, pp. 2057 – 2062, 1998.
`J. Speth, A.Morales, P.J.Sanz, “Vision-Based Grasp Planning of 3D
`Objects by Extending 2D Contour Based Algorithms”, IROS, pp.
`2240 – 2245, 2008.
`[8] S. Haidacher, G. Hirzinger, “Estimating Finger Contact Location and
`Object Pose from Contact Measurements in 3D Grasping”, ICRA, vol.
`2, pp. 1805 - 1810, 2003.
`[9] F. Chaumette, “Image moments: A general and useful set of features
`for visual servoing,” IEEE Trans. Robot., vol.20, no.4, pp.713-723,
`Aug. 2004
`[10] Espiau B, Chaumette F, Rives P, “A new approach to visual servoing
`in robotics”, IEEE Transactions on Robotics and Automation.
`8(3):313-326, 1992
`[11] Y.C. Shiu and S. Ahmad, “Calibration of Wrist-Mounted Robotic
`Sensors by Solving Homogeneous Transform Equations of the Form
`AX=XB”. IEEE Trans. Robotics and Automation. vol.5 no.1.
`pp.16-29. 1989.
`[12] Kwon Oh-Kyu, Sim Dong-Gyu, Park Rae-Hong, “Robust Hausdorff
`distance matching algorithms using Pyramidal structures”, Pattern
`Recognition, 34(7): 2005-2013, 2001
`[13] Guangtao Zhao, Hong Qiao, Zhicai Ou, “A method for calibrating
`camera lens distortion with cross-ratio invariability in welding seam
`system”, IEEE Conference of Intelligent Robotics and Applications
`[14] Shiu Y.C. and Ahmad S. Calibration of Wrist-Mounted Robotic
`Sensors by Solving Homogeneous Transform Equations of the Form
`AX=XB. IEEE Trans. Robotics and Automation 1989; 5(1): 16-29.
`
`Piston and peg
`
`crank
`
`Place the bearing
`
`Insert the peg
`
`Place the axle
`
`(a)
`(b)
`Fig. 8. The axletree-bearing grasping experiments
`
`Fig. 9. The axletree-bearing assembly experiments
`4) Grasping and Assemble the Industrial Parts
`The whole assembly process is illustrated in Fig. 8. The
`sets of bearing-axletree and piston-peg are assembled with
`all the techniques mentioned in this paper: searching for the
`industrial parts, classifying the parts, localizing the axletree
`and the bearing, grasping the parts and assembling them.
`Fig. 8 shows the process of the parts grasping. Fig. 9 shows
`the assembly of the different parts.
`C. Brief Summery of the Assembly System
`In the assembly system with high precision, the assembly
`experiments of the axletree-bearing and the piston-peg are
`performed successfully by the presented methods and
`procedure.
`Further experiments show that (1) The success rate is 75
`percent without visual feedback information; (2) The
`
`1056
`Authorized licensed use limited to: Sterne Kessler Goldstein Fox. Downloaded on December 28,2023 at 00:59:35 UTC from IEEE Xplore. Restrictions apply.
`
`