1

A Fruit Detection System and an End Effector for Robotic Harvesting

of Fuji Apples
D. M. Bulanon 1, T. Kataoka 2
1

Department of Agricultural and Biological Engineering,

University of Florida, Gainesville, FL 32611
2
Crop Production Engineering, School of Agriculture, Hokkaido University, Sapporo, Japan
Corresponding authors e-mail: bulanon@[Link].

ABSTRACT
The challenges in developing a fruit harvesting robot are recognizing the fruit in the foliage and
detaching the fruit from the tree without damaging both the fruit and the tree. The objectives of
this study were to develop a real-time fruit detection system using machine vision and laser
ranging sensor and to develop an end effector capable of detaching the fruit similar to the human
picker. The Fuji apple variety was used in this study. In the detection of the fruit, machine vision
was combined with laser ranging sensor. The machine vision recognized the fruit and the laser
ranging sensor determined the distance of the fruit. The system detected single fruit with 100 %
accuracy in both front and back lighted scenes and the distance measurement accuracy was 3
mm. In detaching the fruit from the tree, an end effector was developed with a peduncle holder
and a wrist; the peduncle holder pinches the peduncle of the fruit and the wrist rotates the
peduncle holder to detach the fruit. Field test results of the end effector showed more than 90%
success rate in detaching the fruit with an average time of 7.1 seconds.
Keywords: Apple, end effector, image processing, machine vision, robotic harvesting, Japan
1. INTRODUCTION
The two main tasks of a fruit harvesting robot are to detect the fruit and to pick the fruit without
damaging it and the tree. These two tasks are the challenges faced by researchers today and this
is also the reason why robotic fruit harvesting is not yet commercially applied. The development
of a fruit harvesting robot is a viable solution to the decreasing number of farm workers and the
increasing cost of agricultural operations such as harvesting.
To detect the fruit, several authors e.g. (Jiminez et al., 2001; Bulanon et al., 2001; Hannan &
Burks, 2004; Ling et al., 2004; Monta et al., 1998) have reported the development of fruit
detection systems. Most of the works on fruit detection used machine vision wherein a CCD
(charge coupled device) camera was used to capture the scene and a PC (personal computer) to
do the image processing. The techniques in image processing could be divided into spectralbased or shape-based analysis. Spectral-based analysis was effective for fruits with reflectance
different from the background (Bulanon et al., 2002a) while shape-based analysis was used to
look for a specific shape of the fruit (Ling et al., 2004). Although promising results have been
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

2
obtained, problems were also encountered. One of these was uneven lighting condition (Bulanon
et al., 2002b). This condition could affect the reflectance of objects which could result in not
detecting the fruit or detecting a non-fruit object. This leads to the second problem, which is false
detection. In some shape-based approach, leaves were detected as fruit. The third problem is
occlusion where fruits are partially hidden by other fruits and leaves. Some researchers have
reported methods to detect occluded objects. One of the popular methods is the circular Hough
Transform, which is effective for circular objects such as oranges, apples and tomatoes (Plebe &
Grasso, 2001). However, results showed that this method was computationally intensive which
would pose a challenge for real-time application and they also reported that the contour of other
objects such as the leaves generated false detection. Another research reported the use of air
blowing device to avoid leaf occlusion (Dobrusin et al., 1992), however this may not be
applicable to the apple trees. Finally, the lack of distance or range information is a challenge for
researchers. The acquired image gives only two-dimensional information, however the distance
of the fruit is unknown. The use of stereo vision, ultrasonic sensor, and laser ranging sensor have
been used to supplement the distance information (Hannan & Burks, 2004). A robust fruit
detection system is required to work in a complex environment such as an orchard.
Picking of the fruit is the task wherein the robot makes contact with the fruit. It should be
pointed out that fruits for the fresh market should be free of damage. This is one of the
challenges of end effector development. Another challenge is the manner of removing the fruit
from the tree. Different fruits have different ways of harvesting. In case of the Fuji apple (Fig. 1),
the fruit should be lightly cradled between the palm and the finger, the thumb or the forefinger
against the base of the peduncle. The apple should be removed with a twisting and lifting motion.
Figure 1 shows that the center of rotation is the topmost portion of the peduncle. This topmost
portion is called the abscission layer, located between the peduncle and the fruit spur. This
procedure is strictly followed because it is important that the peduncle remain on the apple, as an
apple without a peduncle has a less storage life and a low market value especially in Japan.

Figure 1. Manual harvesting of Fuji apples; (a) initial position; (b) start of rotation;
(c) fruit rotation; (d) fruit detached.
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

3
The objectives of this paper were 1) To develop a fruit detection system that could detect single
fruit and measure its distance from the camera, 2) to develop an end effector prototype that
mimics the human harvesting method, and 3) to evaluate the performance of both the fruit
detection system and the end effector.
2. MATERIALS AND METHODS
2.1 Tested Fruit
The fruit tested in this study is the red Fuji apple. This is one of the popular apple varieties in
Japan. 50 % of the apple production in Japan is of this variety. Fuji apples are harvested in the
beginning of November.
2.2 Development of Fruit Detection System
2.1.1 Hardware Development
The fruit detection system is composed of a machine vision system to recognize the fruit and a
laser ranging sensor to determine the distance to the fruit. The machine vision system consists of
a compact color CCD video camera to capture images of the apples, a USB frame capture device
to digitize the acquired images, and a PC (Pentium 1 GHz) for image processing. The acquired
image was a 320 240 bitmap image.
The tested laser ranging sensor can measure distance from 30 cm to over 100 cm. The camera
was mounted on the laser ranging sensor (Fig.2). This position was configured to align the
optical axis of the camera with the laser. The goal here is once the desired object, in this case the
apple is positioned at the center of the image through visual servoing, the laser ranging sensor
could easily measure the distance to the fruit. The laser ranging sensor was connected to the
computer through the RS-232.

Figure 2. End-effector with fruit detection system.

[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

4
2.2.2 Software Development
The fruit detection algorithm (Fig.3) for the machine vision has six steps; acquisition,
segmentation, filtering, labeling, edge extraction, and feature extraction. Segmentation is the first
step of object recognition. Van Henten et al.(2003) reported segmentation as one of the sources
of failure during harvest. The segmentation method used was a color based method developed by
Bulanon et al.(2002a). In this approach the chromaticity coefficients r and g were used as the
feature space. Two decision functions, d1 and d2, that separated the fruit from the other classes in
the feature space were derived using decision theoretic approach (Gonzalez & Woods, 1992).
This method could be applied under different lighting conditions. The chromaticity coefficients
and the decision functions, d1 and d2, are expressed by the following equations:
r=

R
R+G+ B

(1)

g=

G
R+G+ B

(2)

d 1 0.09 0.11 r
d = 0.12 0.06 g

2

(3)

Where d1 and d2 are the decision functions that separate the fruit from the leaf and the branch
respectively. Although, the other parts of the background such as the ground and the sky were
not included in the derivation of the decision functions, results showed that the two functions
were sufficient to separate the fruit from the background.
Using the decision functions, a segmented image g(x, y) is defined as
1 if d1 > 0 and d 2 > 0
g ( x, y ) =
otherwise
0

(4)

Figure 3. Image processing algorithm for fruit detection.

After segmentation, a low pass filter was passed over the segmented image to remove noise.
Connected pixels were classified as one segment or blob (Davies, 1997). A Laplacian edge
detector was used to extract the edges of the segments. At this point, the segments were not yet
considered as fruit. The morphological properties such as area, major axis, length, width, aspect
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

5
ratio, segment center were determined in feature extraction. These features were used to classify
the segments as single fruit and occluded fruit. In this study, single fruit was considered as
harvestable while occluded fruit was not. A single fruit is a fruit that has 25% occlusion or less.
Occlusion here is defined as leaf/branch occlusion and fruit occlusion. Although apples dont
grow in clusters, an apple which is located behind a certain apple could be viewed as a single
multiple fruit cluster and it is considered here as fruit occlusion. However, they could also be
viewed as single fruits in another view. The goal here is to harvest the apples with high accuracy.
To determine if the segment was a single fruit or an occluded fruit, a shape area factor was
defined. The shape area factor is the ratio of the area of the segment to the area of the circle with
the diameter of the major axis of the segment. The shape area factor of a single fruit was found to
be more than 0.75.
For the distance measurement, the camera and the laser ranging sensor were mounted on a
cylindrical manipulator and the motion of the manipulator was controlled by visual servoing
(Bulanon et al., 2005). Visual servoing positions the target fruit (apple) in the center of the
image. Once the fruit center was aligned with the image center, the distance to the fruit was
measured automatically using the laser ranging sensor.
The program for the image processing, laser control, interface of machine vision and laser
ranging sensor was developed using the Visual C# programming language.
2.2.3 Performance Evaluation
To evaluate the performance of the fruit recognition system, it was tested in the field in the first
week of November 2004. One hundred images were acquired; fifty images were under front
lighting and fifty images were under back lighting condition.
2.3 Development of End Effector
2.3.1 End Effector Design
The design of the end effector was based on the manner the human picker removes the fruit from
the tree and the apple harvesting force analysis made by Kataoka et al. (1999). The end effector
has two components: the peduncle holder and the wrist (Fig.4). The peduncle holder is a DC
motor equipped with two fingers with an opening width of 15 mm and a gripping force of 11 N.
This is enough to hold the fruit by its peduncle, as the average weight of Fuji apples is estimated
to be less than 400 g. The wrist is a stepper motor and it rotates the peduncle holder after it
pinches the peduncle. It has a torque of 1.5 Nm, which, based on force analysis, is sufficient for
harvesting.

[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

Figure 4. End effector prototype.

The control of the prototype is shown in Fig.5. The peduncle holder receives open/close signal
from the PC through the digital I/O and it also sends back its status to the PC. The wrist is
controlled by a pulse signal at 500 Hz produced by a microcontroller which is connected to the
PC also through the digital I/O. The control interface in the PC was also developed using Visual
C#.
PC

Digital
Input/Output

Microcontroller

DC motor
control

Stepper motor
control

DC Motor

Stepper Motor

Figure 5. Control of End effector.

[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

7
2.3.2 Performance of End Effector
To evaluate the performance of the end effector, it was also tested in the field during the
harvesting season of the Fuji apples in Hokkaido University. The harvested fruits were from
three Fuji trees. Twenty two apples were harvested. In the end effector performance test, the
fruit was positioned with the peduncle inside the fingers of the peduncle holder. This was the
initial position with the fingers parallel to the ground. Performance evaluation started with the
closing of the fingers and then the fingers were rotated 120 degrees.

2.4 Field Test of the Apple Harvesting Robot Prototype

The apple harvesting robot prototype is composed of the developed machine vision (eyes), the
fruit picker (hand) and the developed cylindrical manipulator (arm). The apple harvesting robot
was mounted on a vehicle lift so that it could easily move in the orchard and it also increases the
work area of the robot. Harvesting test of the robot was made in the first week of November
2005, which was the harvesting period for the Fuji apples. The tests were done both in the Yoichi
experimental orchard and the apple orchard inside Hokkaido University in Sapporo. Seventy five
(75) apples were tested.

Figure 6. Field Test of Robotic Harvesting Prototype

[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

8
3. Results and Discussion
3.1 Fruit recognition system
The fruit recognition system was tested in the field. Figure 7 shows the result of the image
processing algorithm. Two sets are shown here; Fig.7(a) is when the fruits are subjected to
uneven lighting and Fig.7(b) is when there is an occlusion.
Although Fig.7(a-1) was taken under front lighting condition, the presence of other objects in the
scene such as a branch could cause uneven lighting. The two fruits had different illumination; the
larger fruit is brighter as compared to the other fruit, which is shaded. Although the fruits had
different illumination, the segmented image shows that both fruits were recognized in the image.
Fruits in both frontlighted and backlighted images were successfully detected. This shows that
the segmentation method was able to adjust to the different lighting conditions of the fruit. The r
and g feature space transformation decouples intensity from the original RGB image and this
facilitated the segmentation to adjust to different lighting condition. An effective segmentation
method should be utilized because it affects the subsequent processes.
Figure 7(b-1) shows the performance of the algorithm when a fruit is occluded (extreme left fruit
is occluded by leaves). The final image shows that only the two single fruits were detected. The
algorithm identified the occluded fruit and so it was not considered as a harvestable fruit. The
biggest challenge of occlusion is false detection, which would affect the accuracy of the
harvesting system. Therefore in this study, the main focus is to determine if the segmented
portion is a single fruit or an occluded fruit and treat the single fruit as the fruit to be harvested.
In this way, false detection could be avoided and the probability of harvesting the fruit
successfully is increased. Furthermore, in the case of Fuji apples, flower thinning operation
lessens the probability of fruits occluded by other fruits in the image.

[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

Figure 7. Image processing results; (a) sample of uneven lighting; (b) sample of occlusion;
(1) acquired image;(2) segmentation; (3) labeling; (4) edge extraction; (5) feature extraction.

Table 1 shows the result of the performance of fruit recognition system. There were a total of
190 single fruits and 73 occluded fruits. All the single fruits were detected while there were false
detections in the detection of occluded fruits. Some parts of the image such as the branch were
segmented as fruit and they were detected as an occluded fruit. Although they were detected as
occluded fruits, they were not considered as harvestable fruit. This shows that the shape area
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

10
factor was effective in dealing with non-fruit objects, which were falsely segmented. It is also
noted here that this image processing step could be considered a spectral based approach plus a
shape based approach because of the addition of the shape area factor to classify the objects in
the image. Table 2 shows the execution time for each image processing step and its relative
percentage. The total time is less than 500 ms, which means that real-time application of this
algorithm is possible and this could be implemented using a maximum frame capture rate of 2
frames per second. Labeling took much of the total execution time. Improving the labeling
algorithm could decrease the total time and increase the frame capture rate.
Table 1. Performance of fruit detection algorithm
Total
Detected

False Positives

Single Fruit
Occluded Fruit

0
11

190
73

190
84

Table 2. Processing time for fruit detection

Processing time, ms
Segmentation
15
Filtering
46
Labeling
187
Edge extraction
46
Feature extraction
15
Total
309

Selected
Harvesting
190
0

Relative percentage, %
4.8
14.9
60.6
14.9
4.8
100

To evaluate the performance of the laser and the machine vision system, the image processing
algorithm was implemented in real-time with frame capture rate of 1 frame per second. This was
used in the visual servoing of the manipulator. When a single fruit was detected, the camera was
moved to position the center of the detected single fruit to the center of the image. Once the fruit
center coincided with the image center, the camera was moved 50 mm upward and the laser
measured the fruit center. The distance of the detected single fruit was measured with a 3 mm
accuracy.
3.2 Performance of end effector
Figure 8 shows the robotic harvesting of Fuji apples. Figure 8(a) shows the starting position.
From here the operation was controlled by the PC. Then the fingers closed and held the peduncle
(Fig. 8(b)). Once the close signal was received by the PC, the wrist rotated the peduncle holder
from the initial position to 120 degrees (Fig.8(c)-(d)). Comparing this with Fig.1, the twisting
motion of the stem at the abscission layer in robotic harvesting is similar with manual harvesting.
Results (Table 3) showed that the end-effector had more than 90% success rate. In case of the
fruits that were not successfully removed, one had a short peduncle and when it was removed,
the fruit had no peduncle. The other fruit fell down when the holder pinched the peduncle. Table
4 shows the physical properties of the harvested apples. Although, the number of trials was not
sufficient to warrant the reliability of this prototype, it was enough to show that it could remove
the fruit similar to the way human picks the fruit. Removal of the fruit took an average of 7.1
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

11
seconds with a minimum of 3 seconds and a maximum of 14 seconds. Removal time depends on
the pulse signal to the wrist stepper motor. Decreasing this time would mean increasing the
frequency of the pulse signal.

Figure 8. Robotic harvesting of Fuji apples; (a) initial position; (b) fingers closed;
(c) fruit rotation; (d) fruit detached.
Table 3. Performance of end effector
Number of fruits
Successful removal
20
Unsuccessful removal
2
Total
22
Table 4. Physical properties of harvested apples
Diameter, mm
Height, mm
Maximum
89
82
Minimum
66
67
Average
77.8
78.7

Percentage, %
90.9
9.1
100

Peduncle length, mm
19.5
12.5
16.5

Weight, g
300
150
238.5

The main advantage of this end effector is its contact with the fruit. It makes contact with the
peduncle only and not the fruit. It is possible to hold the fruit with the peduncle only because of
the high tensile strength between the peduncle and the fruit. Other developed end effectors
(Monta et al. (1998); Cho et al. (2002)) had direct contact with the fruit while controlling the
gripping force. Although, the gripping force is controlled, there is still a high probability of
causing damage to the fruit. In designing an end effector for fruit harvesting, contact area should
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

12
be one of the considerations. Lesser contact area is better for the fruit but without sacrificing
grasping capability of the end effector.
The limitation of this end effector is the way it approaches the fruit. The end effector should
approach the fruit horizontally. When attaching this end effector to a manipulator, this is one of
the constraints that should be considered in the trajectory planning. In addition, the fruit
recognition system should also take into consideration that this end effector requires the position
of the peduncle with high accuracy. Bulanon et al. (2001) had reported an image processing
technique to determine the peduncle position. In this method, the fruit center and the fruit outline
are required. These two features are easily determined in the present fruit recognition system.
3.3 Field Test of Harvesting Robot Prototype
Results of the field test showed that the robot successfully harvested about 89% of the apples.
Eleven percent (11%) of the apples was not successfully picked. There were several factors that
were considered for the failure; (1) the position of the peduncle, (2) size of the peduncle and (3)
difficulty in the fruit recognition. In reason (1), some fruits did not have a position where the
position of the peduncle is straight because of blockage by the branch or leaf. Because of this,
the machine vision was not able to correctly calculate the position of the peduncle. In reason (2),
some fruits have very short peduncle and the picker held the branch instead of the peduncle and
when it picked the fruit, the branch was taken together with the fruit. In case of reason (3), there
were instances where the machine vision failed to recognize the fruit because of the background,
there were other fruits behind the target fruit which would make the fruit looked overlapped in
the image. So we had to move the robot to another position where it would be recognized as a
single fruit. Although, the success rate is considerably high, the response time of the robot was
still slow. The main reason for this one as stated above is the machine vision image processing
rate. If the image processing could be improved and its speed be increased, then the robot could
be used for apple harvesting. Another improvement that will be looked into is the design of its
fruit picker. It is thought that the current length of its peduncle holder is short and its width is
large. If the length would be increased and its width decreased, the chance of holding the fruit
with shorter peduncle is high. Further development in the manipulator area should also be looked
into specifically the determination of the configuration of the manipulator that is suitable for the
apple tree plant training system. The current manipulator was not developed based on the plant
training system of the apple. Although, there are still a lot of improvements needed for the
current apple harvesting robot, the results showed a bright future in this research area.

4. Conclusion
A fruit detection system and an end effector that can be attached to a commercial manipulator
were developed for robotic apple harvesting. The fruit detection system used machine vision to
recognize the fruit and laser ranging sensor for the fruits distance information. Results showed
that it could detect single apples with a 100% accuracy and no false detection of single fruit. It
could measure the fruits range with 3mm accuracy. Image processing took less than one
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

13
second and it can be concluded that real-time application is possible. The developed end effector
prototype was based on the way human picks the apple. It makes contact with the peduncle of the
fruit only. Performance test of the end effector showed that it has a success rate of over 90%. The
machine vision system and the end effector were attached to a cylindrical manipulator. Field
tests showed that the robotic harvesting prototype successfully picked 89% of the apples. Future
studies would involve improving frame rates of machine vision system, handling system of the
end effector, and development of a manipulator suitable for the apple trees.
Acknowledgement
The authors gratefully acknowledge the Japan Society for the Promotion of Science for the
fellowship (No.15-3253) and the Japan Ministry of Education, Culture, Sports, Science and
Technology for the research grant.
References
Bulanon, D., Kataoka, T., Ota, Y., Hiroma, T. 2001. A Machine Vision System for
the Apple Harvesting Robot. Agricultural Engineering International: the CIGR Journal of
Scientific Research and Development Manuscript PM 01 006. Vol. III.
Bulanon, D., Kataoka, T., Ota, Y., Hiroma, T. 2002a. A Color Model for Recognition of Apples
by a Robotic Harvesting System. Journal of the JapaneseSociety of Agricultural Machinery
64(5): 123-133.
Bulanon, D., Kataoka, T., Ota, Y., Hiroma, T. 2002b. A Segmentation Algorithm for the
Automatic Recognition of Fuji Apples During Harvest. Biosystems Engineering 83(4):
405-412.
Bulanon, D., Kataoka, T., Okamoto, H., Hata, S. 2005. Feedback Control of Manipulator Using
Machine Vision for Robotic Apple Harvesting. ASAE Paper 053114. [Link], MI.:
ASAE.
Cho, S. I., Chang, S. J., Kim Y., An K. J. 2002. Development of a Three-degrees-of-freedom
Robot for harvesting Lettuce using Machine Vision and Fuzzy logic Control. Biosystems
Engineering 82(2): 143-149.
Davies, E. R. 1997. Machine Vision: Theory, Algorithms, Practicalities. Academic Press.
Dobrusin, Y., Edan, Y., Grinshpun, J., Peiper, U., Hetzroni, A. 1992. Real-time image processing
for robotic melon harvesting. ASAE paper No. 92-3515. St. Joseph, MI.: ASAE.
Gonzalez, R., Woods, R. 1992. Digital Image Processing. Addison-Wiley Publishing Company
Hannan, M., Burks, T. 2004. Current Developments in Automated Citrus Harvesting. ASAE
paper 043087. St. Joseph, MI.: ASAE.
[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.

14
Jimenez, R., Ceres, R., Pons, J. L. 2001. A survey of Computer Vision for Locating Fruit on
Trees. Transactions of the ASAE 43(6): 1911-1920.
Kataoka, T., Ishikawa, Y., Hiroma, T., Ota, Y., Motobayashi, K., Yaji, Y. 1999. Hand
Mechanism for Apple Harvesting Robot. Journal of the Japanese Society of Agricultural
Machinery 61(1): 131-139.
Kondo, N., Nishitsuji, Y., Ling, P., Ting, K. 1996 Visual feedback guided robotic cherry tomato
harvesting. Transactions of the ASAE 39(6): 2331-2338.
Ling, P., Ehsani, R., Ting, K. C., Chi, Y., Ramalingam, N., Klingman, M., Draper, C. 2004.
Sensing and End-Effector for a Robotic Tomato Harvester. ASAE paper 043088. St.
Joseph, MI.: ASAE.
Monta, M., Kondo, N., Ting, K. C. 1998. End-Effectors for Tomato Harvesting Robot. Artificial
Intelligence Review. 12: 11-25.
Plebe, A., Grasso, G. 2001. Localization of spherical fruits for robotic harvesting. Machine
Vision and Applications. 13: 70-79.
Van Henten E. J., Van Tuijl B. A. J., Hemming, J., Kornet, J. G., Bontsema, J., Van Os E. A.
2003. Field Test of an Autonomous Cucumber Picking Robot. Biosystems Engineering
86(3): 305-313.

[Link], T. Kataoka. A Fruit Detection System and an End Effector for Robotic
Harvesting of Fuji Apples. Agricultural Engineering International: the CIGR Ejournal.
Manuscript 1285. Vol. XII. February, 2010.