You are on page 1of 10

DESIGN AND FIELD EVALUATION OF A

ROBOTIC APPLE HARVESTING SYSTEM WITH A


3D-PRINTED SOFT-ROBOTIC END-EFFECTOR
C. J. Hohimer, H. Wang, S. Bhusal, J. Miller, C. Mo, M. Karkee

ABSTRACT. Fresh market apple harvesting is a difficult task that relies entirely on manual labor. Much research has been
done on the development of mechanical harvesting techniques. Several selective harvesting robots have been developed for
research studies, but there are no commercially available robotic systems. This article discusses the design and development
of a novel pneumatic 3D-printed soft-robotic end-effector to facilitate apple separation. The end-effector was integrated
into a robotic system with five degrees of freedom that was designed to simplify the picking sequence and reduce costs
compared to previous versions. Apples were successfully harvested using the low-cost robotic system in a commercial or-
chard during the fall 2017 harvest. A detachment success rate on attempted apples of 67% was achieved, with an average
time of 7.3 s per fruit from separation to storage bin. By conducting this study in an orchard where problematic apples were
not removed to increase the detachment success rate, current pruning and thinning practices were assessed to help lay the
foundation for future studies and develop strategies for successfully harvesting apples that are difficult to detach.
Keywords. Apple catching, Apples, Automated harvesting, Field experimentation, Harvesting robot, Soft-robotic gripper.

F
or the past several decades, researchers have been and design choice with performance factors as well as eco-
trying to develop systems for autonomous fresh nomic metrics. For a system to be commercially viable, it
market fruit harvesting, with the goal of reducing must add value to the current method for harvesting a spe-
labor cost and the overall dependence on labor. The cific crop. In the apple industry, this threshold seems to be
success and failure of such systems have been discussed in relatively low, as harvesting labor amounts to almost 30% of
detail in several review articles in the last several years (Bac the variable annual cost, and this seasonal labor force is in
et al., 2014; Peterson, 2005; Zhang et al., 2016a). Of specific increasingly short supply (Gallardo et al., 2010).
interest to this work are harvesting systems for fresh market One of the major hurdles faced by researchers in this field
apples. Despite several projects dating back as far as 1987, is the complex, unstructured environment in which a robot
researchers have been unable to bridge the gap from R&D to must operate. Unlike typical industrial robots that perform
commercialization or even to large-scale studies (Zhang et tasks repeatedly, harvesting robots must operate in a space
al., 2016a). Several projects are underway to create commer- with poorly defined positions and without a clear path from
cial systems, but even the most advanced of these projects is point A to point B within the workspace. Often, a robot is
still in the early startup stages (e.g., FFRobotics, Israel; required to operate without complete knowledge of the ob-
Abundant Robotics Inc., California). jects in the workspace that need to be avoided.
This gap between R&D and commercialization stems Over the last several decades, fruit growers have made
from several complex problems with autonomous harvest- significant progress in orchard design to maximize crop
ing, and any proposed system must balance each component yield and reduce the labor required to harvest and maintain
fruit trees. With this effort, highly structured, high-yield or-
chards are being developed and planted across the industry.
Several studies have been conducted to assess and develop
Submitted for review in June 2018 as manuscript number MS 12986;
approved for publication by the Machinery Systems Community of ASABE
tree structures to improve reachability in fruit trees. Bloch et
in December 2018. al. (2018) used environmental modeling to propose optimal
The authors are Cameron J. Hohimer, Graduate Student, School of robot configurations and designs to improve performance in
Mechanical and Materials Engineering, and Heng Wang, Graduate reachability based on the tree structure. Similarly, reachabil-
Student, School of Engineering and Applied Sciences, Washington State
University, Richland, Washington; Santosh Bhusal, Graduate Student, ity simulations in high-density trellised pear trees were con-
Center for Precision and Automated Agricultural Systems (CPAAS), ducted to assess pickable fruit using only linear motion; the
Washington State University, Prosser, Washington; John Miller, Professor, authors concluded that up to 93% of the fruit could be
School of Engineering and Applied Sciences, and Changki Mo, Professor,
School of Mechanical and Materials Engineering, Washington State
reached by linear motion only (Arikapudi and Vougioukas,
University, Richland, Washington; Manoj Karkee, Professor, Center for 2017; Vougioukas et al., 2016). These results provide insight
Precision and Automated Agricultural Systems (CPAAS), Washington into optimal tree architectures, but little work has been done
State University, Prosser, Washington. Corresponding author: Cameron to understand individual fruit locations in relation to other
Hohimer, 2710 Crimson Way, Richland, WA 99352; phone: 509-372-7356;
e-mail: cameron.hohimer@wsu.edu. fruit, tree branches, and trellis wires and how these factors

Transactions of the ASABE


Vol. 62(2): 405-414 © 2019 American Society of Agricultural and Biological Engineers ISSN 2151-0032 https://doi.org/10.13031/trans.12986 405
relate to robotic harvesting success rates. BACKGROUND
It has been stated that growers are willing to prune their Over the past several years, Washington State University
trees to facilitate robotic harvesting, but thus far no study has has been developing an autonomous robotic system for har-
been conducted to determine how apples might need to be vesting fresh market apples. The first prototype was tested
thinned to improve a robot’s harvesting success rate. Addi- in the fall of 2015 using a robot with seven degrees of free-
tionally, the lack of detailed reporting within robotic harvest- dom (DOF) and a tendon-driven three-fingered end-effector
ing research makes it difficult to assess the possible modifi- with little to no lateral reach, i.e., to the robot’s left or right
cations to the field environment, such as thinning, that have (Davidson and Mo, 2015; Silwal et al., 2017). This initial
been done to improve performance (Bac et al., 2014). In field trial was largely successful, but it had two areas for pos-
sweet pepper harvesting, the robotic harvesting success rate sible improvement: (1) expand the workspace to increase the
(harvested fruit vs. total fruit in canopy) can be increased by number of pickable apples, and (2) implement a fruit con-
more than ten times by removing clusters and occluding veyance system (previously, picked apples were simply
leaves before harvesting (Bac et al., 2017). dropped to the ground).
The research reported in this article addresses this issue In 2016, the robot workspace was expanded by adding an
by studying a grasping type end-effector and how harvesta- additional DOF (8-DOF total), so the robot had wider lateral
bility is affected when apples have not been thinned before movement. Attempts were made to address fruit conveyance
harvesting. This study is expected to provide essential infor- by adding a secondary robot to catch and store apples after
mation for optimizing both crop yields and robotic harvest- the harvesting arm had separated them from the tree. This
ing success rates. Several conditions that are not conducive work is described in detail by Davidson et al. (2016a, 2017).
to robotic harvesting are expected with this approach. Previ- In the 2016 field trial, a major shortcoming was a failure to
ously, apples that were too close to branches or trellis wires, detach fruit from the tree, as well as a high tendency for the
in clusters of two or more, or on long pendulum-like catching robot to collide with low-hanging branches. Be-
branches were removed before robotic harvesting. The re- cause of these difficulties, the performance of the catching
sults of this thinning yielded, on average, only a possible robot was not assessed.
3.5 fruits per cycle and fewer than 2.5 fruits successfully Excessive movement prevented the robot arm from con-
picked per cycle (Silwal et al., 2017). Without information sistently separating apples from the tree. The first field trial
on the number of fruits thinned or removed before harvest- in 2015 was largely successful because the robot was con-
ing, it is impossible to assess the true efficiency of a robot strained to ensure that it pulled apples away from the tree
because the tree was altered to improve performance. For along a vector normal to the end-effector palm, something
this field trial, the objectives were to: that was not true for the 2016 kinematics. These observations
 Develop a pneumatic soft-robotic end-effector to re- showed a need to further refine many of the design changes
duce the risk of damage to both the tree and robot due from the 2015 and 2016 field trials. These changes and the
to collisions. results of the 2017 field trials are discussed in this article. A
 Assess the performance of the harvesting system in an goal of this study was to assess the performance of the ro-
orchard that has not been thinned to remove problem- botic harvester in an environment that had not been artifi-
atic apples and thereby artificially improve harvesting cially thinned to increase the robot’s efficiency. Figure 1
efficiency and performance. shows the harvesting robot staged to pick ‘Jazz’ apples dur-
ing the fall 2017 harvest.

Figure 1. Robotic harvester preparing to harvest ‘Jazz’ apples in 2017.

406 TRANSACTIONS OF THE ASABE


ROBOTIC APPLE HARVESTER DESIGN to experience what a human picker would see at the start of
SOFT-ROBOTIC PNEUMATIC END-EFFECTOR the harvest, i.e., some apples in contact with trellis wires and
Fruit grasping with robotic end-effectors spans a wide branches as well as the trunks of the trees, making collisions
range of applications, and many different types have been inevitable.
implemented with various degrees of success (Blanes et al., To accomplish this without risking damage to the tree or
2011). Many of these end-effectors were designed for use in robot, a soft-robotic end-effector was developed. While sim-
a more controlled setting, such as a processing facility, and ilar end-effector prototypes have been previously developed
not in an unpredictable field environment (Ceccarelli et al., for different applications, few have moved beyond labora-
2000; Davis et al., 2008; Morales et al., 2014). Some were tory testing (Amend et al., 2012; Rateni et al., 2015; Udupa
designed for field or greenhouse applications but have been et al., 2014; Wang et al., 2017; Yap et al., 2016). The end-
limited to laboratory testing (Dimeas et al., 2015). Most of effector consists of three highly compliant, pneumatic actu-
these end-effectors rely on compliant underactuated designs ators mounted around a soft, flexible palm designed to sup-
using a combination of rigid components linked together by port an apple against the actuators, as shown in figure 2a.
flexible hinges that allow them to adapt to slight changes in Figure 2b shows the dimensions of the 3D-printed actuators.
object shape and size. These systems work well in controlled Figure 3 shows the functional prototype gripping an apple
settings, but a more compliant approach is required when po- before harvesting.
sition error and obstacle interference are common. Each actuator is symmetrically positioned about a circle,
A goal of this study was to assess the performance of the 120° apart. The upper two actuators are shorter (95.25 mm)
robotic harvester in an orchard that had not been artificially and are designed to stabilize the apple within the grasp of the
cleared of problematic apples (e.g., removal of clusters, pen- end-effector. The lower actuator is longer (152.4 mm) and is
dulum fruit, and fruit close to or behind trellis wires or designed to wrap almost fully around the apple. This config-
branches) or otherwise altered to enhance the robot’s perfor- uration reduces potential interference between the actuators
mance. While normal pruning and thinning occur throughout when grasping a particularly small apple.
the year to improve crop yield, based on traditional growing The actuators were fully 3D-printed without any post-
practices, these operations are not done with robotic harvest- processing required. Each actuator was fabricated in a single
ing in mind. In this study, the intent was to allow the robot step and then put into service. For this study, a custom three-

Figure 2. (a) CAD model of soft-robotic end-effector and (b) diagram of actuator with dimensions in millimeters.

Figure 3. Soft-robotic end-effector grasping an ‘Envy’ apple: (a) unactuated side view, (b) actuated side view, and (c) actuated top view.

62(2): 405-414 407


axis gantry system was built and fitted with a Titan Extruder al. (2016b) focused on straight-line pulls similar to the sep-
with a V6 HotEnd (E3D-Online, Chalgrove, U.K.) suitable aration technique used in the present study. Those previous
for printing flexible filaments to fabricate the actuators. Sim- results suggested that the soft actuators used in this study
ilar to other actuators fabricated with fused deposition mod- have sufficient grasping strength for fruit separation. For
eling (FDM), the actuators used in this study are comprised field testing, 103.4 kPa (15 psi) was set on the air supply
of bellows that are inflated, causing the actuator to bend regulator, and the compressor was allowed to recover fully
(Yap et al., 2016). The speed, motion, and direction are con- between harvesting sequences.
trolled by a combination of the actuator geometry and the
input air pressure (Mosadegh et al., 2014; Sun et al., 2013). HARVESTING ARM KINEMATICS
For this study, the actuators were printed with a larger cross- The harvesting arm is a 5-DOF system with two prismatic
sectional area and without the constraint of a specific print- (P) joints and three revolute (R) joints. The system first cap-
ing orientation, resulting in a significant increase in the pull- tures an image to identify apples and determine their loca-
ing force for a given input pressure. tions to generate individual fruit coordinates, as discussed
Previously, such devices have been formed out of soft sil- later. Branches and other possible obstacles are not identi-
icone using a mold and casting process that involves multi- fied, and obstacle avoidance was not included in this study.
ple steps (Ilievski et al., 2011). Fully 3D-printed actuators The apple coordinates are processed one at a time by the con-
have the advantage of being fabricated out of a much more trol software and inverse kinematics solver to calculate the
durable thermoplastic polyurethane (TPU) with a Shore joint positions and sequence for the picking process. Apples
hardness of 85 A, compared to silicone with a Shore hard- are prioritized from right to left starting with the smallest
ness of 00 to 30 A. This durability is necessary in an orchard y-coordinate, i.e., the apple farthest to the right in the image
environment where collisions between the tree and end-ef- frame. This minimizes movement of the y-axis translation
fector are common. (q2) and eliminates backtracking. Further optimization could
To determine their maximum pulling force, the actuators be done to maximize the speed or minimize the travel mo-
were clamped into a custom tensile loading machine and ac- tion. As tested, this picking process was a design simplifica-
tuated at various pressures while pulling against a ring at- tion chosen to reduce backtracking and translation.
tached to a load cell (MLP 750, Transducer Techniques, Figure 4a shows the kinematic model of the robot in its
Temecula, Cal.). Table 1 summarizes the pulling force at PPRRR configuration. Within this PPRRR configuration,
various internal pressures. The actuators tested in this study the prismatic joints are used to translate the robot in the x-y
required a minimum of 103 kPa (15 psi) for complete actua- plane. The prismatic joint along the y-axis (q2) is used to po-
tion. At this pressure, the force exerted by the actuator was sition the arm in line with an apple’s y-coordinate. The pris-
determined to be 49.53 6.3 N. The time required for actua- matic joint along the x-axis (q1) is used to approach the fruit.
tion is also given in table 1. A simple benchtop timing device The arm is first staged along the y-axis, 15 cm in front of the
was fabricated to record the delay between the pressure ap- fruit using the three revolute joints and then translated in the
plied and a pushbutton pressed by the actuator. Each tested x-direction using the prismatic joint. Figure 4b is a graphical
pressure was repeated ten times and averaged with standard representation of this staging and translation in the x-z plane.
deviations, also reported in table 1. The actuation time rep- Fruit detachment is the reverse of this movement. Table 2
resents a significant improvement over tendon-driven fin- lists the Denavit-Hartenberg parameters of the system.
gers, which required 1.5 s to close and 1.0 s to release (Silwal Simulations were conducted to compare the reachability
et al., 2017). of the 5-DOF robot to that of the 7-DOF robot used in the
Because of the many factors (e.g., stem location and 2015 field tests and the 8-DOF robot used in 2016 (Davidson
length, fruiting branch length, fruit maturity, fruit diameter) et al., 2017; Silwal et al., 2017). Fruit coordinates were ex-
involved in the force required to separate fruit from a tree tracted from the collected images and used to test the perfor-
(Erdoǧan et al., 2003; He et al., 2017), it is difficult to di- mance through simulation. The coordinates used in the sim-
rectly correlate an actuator’s pulling force to the force re- ulation were for apples located on the third and fourth trellis
quired to separate a given apple from a tree. Several hand- wires, which was the target workspace for all three robots.
picking studies conducted on different apple varieties rec- The results of the simulated reachability are shown in
orded the pulling force normal to the grasping fingers. Da- table 3. From this table, it is clear that the 5-DOF robot lim-
vidson et al. (2016b) found peak forces of 11 to 30 N de- ited the number of reachable fruit. Further analysis of the
pending on apple variety, while Li et al. (2016) reported simulation results showed that many of the missed fruit were
grasping forces of 5 to 15 N. Both studies looked at a variety located too close to the harvesting arm. There were two rea-
of grasping techniques using a three-finger grasp. Li et al. sons for this. First, the arm is limited in its motion on the
(2016) focused on a more traditional picking technique by prismatic joint in the x-direction. If a solution cannot be
applying a bending moment to the stem, while Davidson et found inside the limits of the revolute joints, the arm does
not adjust along the prismatic joint.
Table 1. Pulling force and activation time of soft-robotic actuator from The second cause for the missed fruit is that the distance
benchtop tests. Force and time values are means of ten tests.
between the transport vehicle (and therefore the robotic arm)
Pressure (kPa) Pulling Force (N) Actuation Time (ms)
103.4 49.54 6.37 82.5 1.95
and the fruit wall is not fixed. As the vehicle (driven by a
206.8 73.28 3.67 74.5 2.27 human operator) moves along an orchard row, this gap be-
275.8 96.12 1.70 65.1 0.74 comes wider or narrower. This is an unavoidable conse-
344.8 100.28 4.73 60.2 1.14 quence of the field environment, and the robot does not cur-

408 TRANSACTIONS OF THE ASABE


a)

Figure 4. (a) Kinemetric model of the 5-DOF arm and (b) 2D visual description of picking sequence.

Table 2. Denavit-Hartenberg parameters of 5-DOF harvesting robot.


r 2  l32  l42
Joint a (m)  (°) d (m)  (°)   cos1 (5)
1 0 -90 d1 0 2  l3  r
2 0 90 d2 -90
3 0.27 90 0.19 3 where Xt and Zt are the coordinates of the apple, l3 and l4 are
4 0.25 90 0 4 the lengths from the Denavit-Hartenberg parameters for
5 0.2 90 0 5
joints 3 and 4, and  is the desired end-effector pitch, which
Table 3. Simulation results of three robots tested from 2015 to 2017. was fixed at 0° in this study. Using an analytical solution
Robot Reachable Apples greatly reduces the computational time of the solver. Simu-
5-DOF 234 of 335 (69.9%) lated results showed that the 5-DOF solver is 10,000 times
7-DOF 260 of 335 (77.6%) faster than the numerical 8-DOF solver. While this does not
8-DOF 274 of 335 (81.8%)
amount to noticeable saving in time per apple picked, it is an
important step toward commercialization, when the control
rently account for this in any way. Ideally, the robot work- software will be implemented with a microcontroller.
space would completely overlap the fruit space at every stop Control was carried about using a Linux OS system with
of the arm. An area of active research is to find ways to allow custom C++ code. Dynamixel servomotors (Robotis Inc., Ir-
the robot to adapt to this dynamic environment (Wang et al., vine, Cal.) were used for the three revolute joints, and
2018). NEMA 23 stepper motors and a belt drive system were used
While the reduction in the number of DOF comes at a cost for the two prismatic joints. These actuators offered a low-
(table 3), it also has some benefits, the most significant of cost but sufficiently stable and responsive system. The Dy-
which is a great reduction in overall system cost. One of the namixel servomotors have internal feedback control and po-
main objectives of this project is to optimize the cost and sition sensors. The stepper motors used an open-loop soft-
speed of a robotic harvester to improve its commercial via- ware step-count to maintain positional accuracy.
bility. By removing three motors from the kinematic chain, For the 2017 harvest, the system was constrained such
the total cost of the hardware was reduced by almost 50% that the pitch of the end-effector was fixed at 0°. In previous
from previous versions. This is a significant cost saving con- harvest trials with this robot, the pitch was predefined based
sidering that one of the easiest ways to make robots faster is on the height of the apple. Handpicking data showed that
to employ several of them. there is little benefit to an inclined pull versus a horizontal
Another major advantage is that the inverse kinematics pull (a pitch equal to 0°) for separating fruit (Davidson et al.,
algorithm has an analytical solution. The three revolute joint 2016b). In this study, the pitch was fixed at 0° for all apples,
angles are determined using the following equations: regardless of height. By fixing the pitch of the end-effector,
3  atan2  Z t , X t    (1) the approach angle and grasping were held constant for all
attempted apples. This allowed better assessment of the
4     (2) grasping efficiency of the end-effector, particularly when the
stem location and surrounding branch locations were un-
5    1  2 (3) known. Fixing the pitch in such a way limited the robot
workspace and may have prevented some fruit from being
l32  l42  r 2 reachable (assumed to be a very few). However, it was not
  cos 1 (4) possible to allow an arbitrary pitch, as this would not con-
2  l3  l4
sistently separate fruit from the tree and would cause a

62(2): 405-414 409


Figure 5. Apples identified by vision system for automated harvesting.

greater number of unnecessary collisions. ples, while days three and four were conducted with ‘Envy’
apples. As previously stated, a goal of this study was to as-
APPLE DETECTION sess the performance of the robotic harvester in a commer-
In this study, a relatively inexpensive stereo vision sys- cial orchard without altering the environment; therefore, no
tem was used for apple detection. A BumbleBee XB3 indus- thinning and pruning was performed to improve the picking
trial stereo vision imaging sensor (FLIR Integrated Imaging efficiency. The only exception was the removal of apples di-
Solutions Inc., Richmond, BC, Canada) was used as a single rectly behind a trunk or large branch that were identified by
global camera system. The imaging sensor was mounted be- the vision system; attempts to pick these apples would cause
hind the robot’s prismatic gantry. The sensor was placed damage to the robot or tree. Given that these occluded apples
nearly 1.5 m away from the trees. Stereo image pairs were were rarely identified by the vision algorithm, only a handful
captured with five different exposure values to overcome is- of apples were removed. Unfortunately, the exact number
sues related to the variable lighting conditions (Silwal et al., was not recorded; however, a conservative estimate is that
2017). Each of the five images was fused to form a single approximately 15 apples were removed under this criterion.
image that minimized under- and overexposed regions in the Incorporating obstacle detection into the vision system, so
images. The intrinsic and extrinsic parameters for the camera that these occluded apples could be ignored by the robot and
were computed using checkerboard calibration. These pa- manually picked later, is an area of active research. In a com-
rameters were used in stereo rectification of the fused image, mercial orchard, pruning and thinning are common through-
making them epipolar. The camera calibration process for out the year. These management operations were performed
this study was similar to that of Silwal et al. (2017). The ste- outside the control of this study and were not done to im-
reo rectified image pairs were used to compute disparity prove robotic harvesting efficiency but were based on best
maps, which gave the depth information of each object in the practices of the growers to increase overall yield. Every ef-
camera’s field of view. fort was made to present the robot with a harvesting environ-
Apple detection was performed using a supervised classi- ment that a human picker would experience.
fication technique. Nearly 1400 images from the BumbleBee Simple assessment of the robot’s performance showed
XB3 sensor with different exposure levels were used for that the robot successfully harvested 116 apples, or ~67%,
training the detectors. Two apple detectors were formed us- of 172 attempted picks. Given the overall small number of
ing a boosted cascade of Haar-like features (Viola and Jones, harvesting robots reported in the literature, this is not outside
2001) and local binary patterns (Ahonen et al., 2006) using the range (42% to 77%) of success rates reported by Bac et
OpenCV-310. The two detectors were combined at the deci- al. (2014) for other autonomous apple harvesting studies.
sion level using the 50% overlap criteria defined by Evering- Several of these previous studies were performed in orchard
ham et al. (2010) to form the final apple detector. Combining environments that had been modified specifically to improve
the two detectors minimized the number of false detections the robot’s harvesting efficiency, e.g., removal of fruit clus-
and thus increased the precision of the system (up to 90%). ters, pendulum fruit, and fruit close to or behind trellis wires
Figure 5 shows a typical image after apple detection, with or branches (Silwal et al., 2017; Zhang et al., 2016b). This
eleven detected apples within the robot’s field of view. High reported success rate does not account for identified fruit that
precision was desired in the robotic harvesting of apples to were omitted because they were outside the robot’s work-
increase the speed of picking while minimizing the risk of space. This is significant when compared with table 3, where
damaging the tree. Finally, the 3D coordinates of all detected 69.9% of detected fruit were reachable, or inside the robot’s
apples within the robotic arm’s workspace were identified workspace. Thus, approximately 46.8% of the detected ap-
and transferred to the robotic arm’s controller. ples were harvested. Multiple harvesting arms could be used
to improve this success rate. Because of the space limitations
between orchard rows, it is difficult to design a robot capable
FIELD RESULTS of harvesting fruit over the entire vertical space of a tree. In
Field testing was performed in a commercial apple or- this study, many of the missed fruit were outside the robot’s
chard in Prosser, Washington, over four days (October 20 workspace because of the height limitation but could have
and 26, and November 1 and 3) in the fall of 2017. The first been picked by a robot positioned to pick fruit higher in the
two days of the experiment were conducted with ‘Jazz ‘ap- canopy. Active research is underway to optimize the robot
location to maximize the overall success of a single arm

410 TRANSACTIONS OF THE ASABE


(Wang et al., 2018). pseudo pendulum effect. In these misses, rather than apply-
Missed apples were classified into five categories: clus- ing force directly to the stem as the robot pulled on an apple,
ters, position error, pseudo pendulum, failure to detach, and the pulling effort caused the fruiting branch to swing away
obstructed grasp. Table 4 lists the counts for each category. from the tree, and the robot let go before separation occurred.
Apple clusters, apples obscured from view, and those that Finally, the remaining missed apples were caused by failure
were too close to obstacles and were removed prior to har- to detach (9%) or an obstructed grasp (5%). Apples that
vesting would pose no problems for a human picker. The av- failed to detach slipped out of the end-effector. Inconsistent
erage time per fruit was 7.3 0.78 s. This time accounts for airflow was a factor for these apples, as the supply pressure
the entire picking process, from apple identification to stor- from the portable air compressor decreased after several ap-
age bin. It does not include setup time and includes missed ples in a picking sequence. Obstructed apples occurred when
apples in the averages. The control software recorded the cy- one or more of the actuators closed around an adjacent
cle time, and the total cycle time was then divided by the branch or wire, resulting in incomplete grasp of the fruit.
number of apples attempted.
The most significant cause for missed apples was clus-
tered fruit, at 37% of the total missed apples. Most of these DISCUSSION
misses occurred because the end-effector collided with a sec- At present, there is no guidance on how commercial
ond apple in the cluster during the approach. The collision growers can revise their pruning and thinning strategies to
caused a displacement of the targeted apple, resulting in the maximize robotic harvesting efficiency and minimize the is-
end-effector’s inability to grasp the apple. Additionally, the sues currently faced by robotic harvesting systems. There-
collision occasionally caused the second apple to fall from fore, a robotic system that addresses the performance limita-
the tree, resulting in a failure to pick the second apple be- tions observed in this study due to fruit clusters, pseudo-pen-
cause it was no longer on the tree. This fallen apple was also dulums, and other issues would improve the commercial vi-
considered a clustered fruit miss. ability of such a harvesting system. Several observations
Another major source of missed apples (29%) was posi- were made during the harvesting tests and are discussed here
tion errors between the vision system and the robotic arm. as ways to improve robotic harvesting efficiency without al-
Most of the time, these errors resulted from collisions be- tering the orchard environment (e.g., by thinning clustered
tween the end-effector and the tree or trellis wire. While the apples).
end-effector was designed to be compliant and unaffected by Of the apples successfully picked, 36 (approx. 31%) were
collisions, the prismatic joints use an open-loop control classified as clustered. The robot’s success in picking a par-
scheme. Collisions tended to result in small position errors ticular cluster of apples was observed to be largely random,
that accumulated over multiple attempted apples. A feed- but observations were made as to why certain apples were
back control scheme should be implemented on the prismatic picked successfully. The biggest factor between success and
joints to avoid this type of failure. failure when picking a cluster was the apples’ relative posi-
The third largest source of missed apples (20%) was the tions within the cluster. Figure 6 shows sample images of
apple cluster identification. Apples in the same y-z plane
Table 4. Failure rates by category of missed apples.
Category Total Missed Total Attempted
(i.e., side by side) as the target apple (fig. 6a) were more
Clustered fruit 21 of 56 (37%) 21 of 172 (12%) likely to be pushed aside by the approaching end-effector, in
Position error 16 of 56 (29%) 16 of 172 (9%) contrast to apples with an offset from each other in the x-
Pseudo pendulum 11 of 56 (20%) 11 of 172 (6%) direction (i.e., depth) (fig. 6b).
Failure to detach 5 of 56 (9%) 5 of 172 (3%)
Obstructed grasp 3 of 56 (5%) 3 of 172 (2%)
In the cluster of three apples shown in figure 6a, the upper

Figure 6. Sample images of apple cluster identification: (a) a cluster of three apples that could be picked if properly sequenced, (b) occlusion of
an apple cluster misidentified as a single apple, and (c) a cluster of three apples misidentified as two apples.

62(2): 405-414 411


two apples were most difficult to pick. As the end-effector ever, in practice, this method was not as effective as ex-
tried to grasp one of these two apples, the actuators collided pected. Currently, the system does not identify apples that
with the other apple and fail to grasp either of them. The might be pendulums, nor does it acquire any information
lowest apple of the cluster was successfully picked during about the stem location or fruiting branch. Continuing re-
the harvesting test, but the upper two apples were left on the search on branch and obstacle detection could be helpful in
tree because the end-effector failed to grasp either of them addressing this problem. Other possible solutions to this
successfully. This occurred on several occasions with simi- problem are: (1) modifications to current pruning practices
lar fruit clusters. With higher-level logic for the picking se- to minimize the presence of long branches in tree canopies,
quence and end-effector approach vectors, all three apples and (2) incorporating sensors into the end-effector with feed-
could be picked, and the cluster picking performance could back control to help determine when an apple has been sep-
be improved to acceptable levels. With this proper sequenc- arated from the tree, to help the robot identify failures, and
ing, the lowest apple would be picked first, and the end-ef- to allow second attempts at missed apples. While the first
fector would then approach the other two apples from the solution seems the simplest to implement, it may increase
right and left, respectively. However, additional degrees of labor costs, as time and labor are already spent on training
freedom would be required to approach the fruit with a yaw and pruning the trees along trellis wires. In contrast, embed-
angle other than zero, which would negate the cost saving ded end-effector sensors with real-time feedback have po-
gained from the reducing the system to 5-DOF. tential to improve the robot’s efficiency beyond what could
Figure 6c shows another common difficulty in cluster be achieved with horticultural practices, including the ability
harvesting: an apple cluster has been misidentified, and three to sense picking failures, collisions, apple proximity, false
apples are identified as just two. In this case, the robot is un- positives, etc.
likely to pick any of the three apples successfully because
the misidentified apples are not well aligned to any of the
actual apples. Even with improvements to the motion plan- CONCLUSION
ning and picking sequence, these types of clusters will be Apples were successfully harvested using a low-cost ro-
very difficult to harvest with a grasping end-effector because botic system in a commercial orchard during the fall 2017
of the partial occlusion of the apples, which makes identifi- harvest. A detachment success rate on attempted apples of
cation difficult. Full occlusions are less of a problem because 67% was achieved, with an average time of 7.3 s per fruit
the system only sees the apple in the foreground, as in fig- from separation to storage bin. By conducting this study in
ure 6b. In this case, reimaging of the tree is necessary; oth- an environment in which problematic apples had not been
erwise, the robot will harvest only the identified apple, while removed to increase the detachment success rate, the current
the remaining occluded fruit are left on the tree. pruning and thinning practices were also assessed to help lay
Pendulum apples also posed a significant problem in suc- the foundation for future studies and develop strategies for
cessfully separating them from the tree. Figure 7 shows two successfully harvesting apples that are difficult to detach.
pendulum apples with their long fruiting branches outlined The soft-robotic end-effector offers several improve-
with boxes. An attempted solution to this problem was to ments over previous grasping end-effectors. The compliant
pitch the end-effector down from 0° to -90° before releasing actuators were unharmed by collisions with the tree canopy
the apple. By doing this, the direction of the applied load on and trellis wires and offered a significant improvement in
the stem was altered, which could detach the apple. How- grasping speed over previous tendon-driven fingers, closing
in several orders of magnitude less time (milliseconds versus
seconds) compared to the previous design. Clustered apples
caused some problems for both the vision system and the ro-
botic arm; however, future studies with high-level sequenc-
ing logic may improve the performance in picking apple
clusters.

ACKNOWLEDGEMENTS
This research was supported in part by USDA Hatch and
Multistate Project Funds (Accession Nos. 1005756 and
1001246), a USDA National Institute for Food and Agricul-
ture competitive grant (Accession No. 1000339), the Wash-
ington State University (WSU) Agricultural Research Cen-
ter (ARC), and the National Science Foundation (NSF)
Graduate Research Fellowship Program (NSF Grant No.
121477-005). Any opinions, findings, and conclusions ex-
pressed in this publication are those of the authors and do not
necessarily reflect the views of the USDA, NSF, or Wash-
ington State University. The authors would also like to
Figure 7. Apples with long fruiting branches classified as pendulums. acknowledge Allan Brothers, Inc., for their support.

412 TRANSACTIONS OF THE ASABE


challenge. Intl. J. Comput. Vision, 88(2), 303-338.
REFERENCES https://doi.org/10.1007/s11263-009-0275-4
Ahonen, T., Hadid, A., & Pietikainen, M. (2006). Face description
Gallardo, R. K., Taylor, M., & Hinman, H. (2010). 2009 Cost
with local binary patterns: Application to face recognition. IEEE
estimates of establishing and producing Gala apples in
Trans. Pattern Anal. Mach. Intel., 28(12), 2037-2041.
Washington. Factsheet FS005E. Pullman, WA: Washington
https://doi.org/10.1109/TPAMI.2006.244
State University Extension.
Amend, J. R., Brown, E., Rodenberg, N., Jaeger, H. M., & Lipson,
He, L., Fu, H., Karkee, M., & Zhang, Q. (2017). Effect of fruit
H. (2012). A positive-pressure universal gripper based on the
location on apple detachment with mechanical shaking. Biosyst.
jamming of granular material. IEEE Trans. Robotics, 28(2), 341-
Eng., 157, 63-71.
350. https://doi.org/10.1109/TRO.2011.2171093
https://doi.org/10.1016/j.biosystemseng.2017.02.009
Arikapudi, R., & Vougioukas, S. G. (2017). Linear reachability of
Ilievski, F., Mazzeo, A. D., Shepherd, R. F., Chen, X., &
fruits in orchard trees with actuator constraints. ASABE Paper
Whitesides, G. M. (2011). Soft robotics for chemists.
No. 1701421. St. Joseph, MI: ASABE.
Angewandte Chemie Intl. Ed., 50(8), 1890-1895.
https://doi.org/10.13031/aim.201701421
https://doi.org/10.1002/anie.201006464
Bac, C. W., Hemming, J., van Tuijl, B. A., Barth, R., Wais, E., &
Li, J., Karkee, M., Zhang, Q., Xiao, K., & Feng, T. (2016).
van Henten, E. J. (2017). Performance evaluation of a harvesting
Characterizing apple picking patterns for robotic harvesting.
robot for sweet pepper. J. Field Robotics, 34(6), 1123-1139.
Comput. Electron. Agric., 127, 633-640.
https://doi.org/10.1002/rob.21709
https://doi.org/10.1016/j.compag.2016.07.024
Bac, C. W., van Henten, E. J., Hemming, J., & Edan, Y. (2014).
Morales, R., Badesa, F. J., Garcia-Aracil, N., Sabater, J. M., &
Harvesting robots for high-value crops: State-of-the-art review
Zollo, L. (2014). Soft robotic manipulation of onions and
and challenges ahead. J. Field Robotics, 31(6), 888-911.
artichokes in the food industry. Adv. Mech. Eng., 6, article
https://doi.org/10.1002/rob.21525
345291. https://doi.org/10.1155/2014/345291
Blanes, C., Mellado, M., Ortiz, C., & Valera, A. (2011). Review:
Mosadegh, B., Polygerinos, P., Keplinger, C., Wennstedt, S.,
Technologies for robot grippers in pick and place operations for
Shepherd, R. F., Gupta, U., ... Whitesides, G. M. (2014).
fresh fruits and vegetables. Spanish J. Agric. Res., 9(4), 1130-
Pneumatic networks for soft robotics that actuate rapidly. Adv.
1141. https://doi.org/10.5424/sjar/20110904-501-10
Funct. Mater., 24(15), 2163-2170.
Bloch, V., Degani, A., & Bechar, A. (2018). A methodology of
https://doi.org/10.1002/adfm.201303288
orchard architecture design for an optimal harvesting robot.
Peterson, D. L. (2005). Harvest mechanization progress and
Biosyst. Eng., 166, 126-137.
prospects for fresh market quality deciduous tree fruits.
https://doi.org/10.1016/j.biosystemseng.2017.11.006
HortTech., 15(1), 72-75.
Brown, P., Rasch, M., & Dietrich, C. (2014). Vacuum apple
https://doi.org/10.21273/horttech.15.1.0072
harvester. Conklin, MI: Phil Brown Welding. Retrieved from
Rateni, G., Cianchetti, M., Ciuti, G., Menciassi, A., & Laschi, C.
http://www.philbrownwelding.com/index.php/new-products
(2015). Design and development of a soft robotic gripper for
Ceccarelli, M., Figliolini, G., Ottaviano, E., Mata, A. S., & Criado,
manipulation in minimally invasive surgery: A proof of concept.
E. J. (2000). Designing a robotic gripper for harvesting
Meccanica, 50(11), 2855-2863. https://doi.org/10.1007/s11012-
horticulture products. Robotica, 18(1), 105-111.
015-0261-6
https://doi.org/10.1017/S026357479900226X
Silwal, A., Davidson, J. R., Karkee, M., Mo, C., Zhang, Q., &
Davidson, J. R., & Mo, C. (2015). Mechanical design and initial
Lewis, K. (2017). Design, integration, and field evaluation of a
performance testing of an apple-picking end-effector. Proc.
robotic apple harvester. J. Field Robotics, 34(6), 1140-1159.
ASME Intl. Mechanical Engineering Congress and Exposition.
https://doi.org/10.1002/rob.21715
New York, NY: ASME. https://doi.org/10.1115/IMECE2015-
Sun, Y., Song, Y. S., & Paik, J. (2013). Characterization of silicone
50482
rubber based soft pneumatic actuators. Proc. IEEE/RSJ Intl.
Davidson, J. R., Hohimer, C. J., & Mo, C. (2016a). Preliminary
Conf. on Intel. Robots and Systems, (pp. 4446-4453).
design of a robotic system for catching and storing fresh market
Piscataway, NJ: IEEE.
apples. IFAC-PapersOnLine, 49(16), 149-154.
https://doi.org/10.1109/IROS.2013.6696995
https://doi.org/10.1016/j.ifacol.2016.10.028
Udupa, G., Sreedharan, P., Sai Dinesh, P., & Kim, D. (2014).
Davidson, J. R., Hohimer, C. J., Mo, C., & Karkee, M. (2017). Dual
Asymmetric bellow flexible pneumatic actuator for miniature
robot coordination for apple harvesting. ASABE Paper No.
robotic soft gripper. J. Robotics, 2014, article 902625.
1700567. St. Joseph, MI: ASABE.
https://doi.org/10.1155/2014/902625
https://doi.org/10.13031/aim.201700567
Viola, P., & Jones, M. (2001). Rapid object detection using a
Davidson, J., Silwal, A., Karkee, M., Mo, C., & Zhang, Q. (2016b).
boosted cascade of simple features. Proc. IEEE Computer
Hand-picking dynamic analysis for undersensed robotic apple
Society Conf. on Computer Vision and Pattern Recognition.
harvesting. Trans. ASABE, 59(4), 745-758.
Piscataway, NJ: IEEE.
https://doi.org/10.13031/trans.59.11669
https://doi.org/10.1109/CVPR.2001.990517
Davis, S., Gray, J. O., & Caldwell, D. G. (2008). An end effector
Vougioukas, S. G., Arikapudi, R., & Munic, J. (2016). A study of
based on the Bernoulli principle for handling sliced fruit and
fruit reachability in orchard trees by linear-only motion. IFAC-
vegetables. Robot. Comput. Integ, Manuf., 24(2), 249-257.
PapersOnLine, 49(16), 277-280.
https://doi.org/10.1016/j.rcim.2006.11.002
https://doi.org/10.1016/j.ifacol.2016.10.051
Dimeas, F., Sako, D. V., Moulianitis, V. C., & Aspragathos, N. A.
Wang, H., Hohimer, C. J., Bhusal, S., Karkee, M., Mo, C., &
(2015). Design and fuzzy control of a robotic gripper for
Miller, J. H. (2018). Simulation as a tool in designing and
efficient strawberry harvesting. Robotica, 33(5), 1085-1098.
evaluating a robotic apple harvesting system. IFAC-
https://doi.org/10.1017/S0263574714001155
PapersOnLine, 51(17), 135-140.
Erdoğan, D., Güner, M., Dursun, E., & Gezer, I. (2003). Mechanical
https://doi.org/10.1016/j.ifacol.2018.08.076
harvesting of apricots. Biosyst. Eng., 85(1), 19-28.
Wang, Z., Torigoe, Y., & Hirai, S. (2017). A prestressed soft
https://doi.org/10.1016/S1537-5110(03)00024-2
gripper: Design, modeling, fabrication, and tests for food
Everingham, M., Van Gool, L., Williams, C. K., Winn, J., &
handling. IEEE Robot. Automat. Letters, 2(4), 1909-1916.
Zisserman, A. (2010). The Pascal visual object classes (VOC)

62(2): 405-414 413


https://doi.org/10.1109/LRA.2017.2714141 technology: A review. Trans. ASABE, 59(5), 1165-1180.
Yap, H. K., Ng, H. Y., & Yeow, C.-H. (2016). High-force soft https://doi.org/10.13031/trans.59.11737
printable pneumatics for soft robotic applications. Soft Robot., Zhang, Z., Heinemann, P. H., Liu, J., Schupp, J. R., & Baugher, T.
3(3), 144-158. https://doi.org/10.1089/soro.2016.0030 A. (2016b). Design and field test of a low-cost apple harvest-
Zhang, Z., Heinemann, P. H., Liu, J., Baugher, T. A., & Schupp, J. assist unit. Trans. ASABE, 59(5), 1149-1156.
R. (2016a). The development of mechanical apple harvesting https://doi.org/10.13031/trans.59.11708

414 TRANSACTIONS OF THE ASABE

You might also like