Computers and Electronics in Agriculture 211 (2023) 108023

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Robotic Tree-fruit harvesting with arrays of Cartesian Arms: A study of fruit

pick cycle times
Rajkishan Arikapudi, Stavros G Vougioukas *
Department of Biological and Agricultural Engineering, 2030 Bainer Hall, University of California Davis, CA 95616, USA

A R T I C L E I N F O A B S T R A C T

Keywords: In some cases, it has been shown that fruits on trees with SNAP (Simple, Narrow, Accessible, and Productive)
Automation architectures are reachable by robot arms using three linear degrees of freedom; hence, high fruit-picking effi­
Computer Simulation ciencies can be achieved with simpler arms. This paper uses digitized fruit position data to compute the fruit pick
Harvest Throughput
cycle times (PCT) of robotic fruit harvesters with multiple arms arranged in grid configurations, i.e., operating in
Mechanization
Robotic Harvesting
disjoint rectangular work cells independently of each other. The effects of the robot joints’ maximum linear
Telescopic Arms acceleration and maximum linear velocity on PCTs were studied. As Vmax increased, the PCT followed a negative
exponent power law (diminishing return) for any given Amax. Similarly, for a constant Vmax, the improvement of
PCT as Amax increased slowed down at higher acceleration values.
Also, the PCTs were computed when each arm work cell was designed based on equal fruit load or equal size
criteria, using four workspace partitioning schemes (height split, length split, height split matrix, and length split
matrix). Equal fruit-load configurations resulted in load balancing and exhibited lower PCTs; the height-split
configuration was the best among the four different partitioning schemes, possibly because it compensated
better for the fruit distribution’s non-uniformity along the trees’ height.
Finally, the PCTs were computed while harvesting one side of an orchard row or both sides concurrently.
Harvesting sides separately resulted in lower PCTs (greater speed) due to non-uniform fruit distributions on the
different sides of trees. The insights gained in this paper can inform the design of harvesting robots utilizing
arrays of 3-dof linear arms.

narrow, almost two-dimensional canopies that are easier to harvest

1. Introduction because of increased fruit visibility and reachability. Researchers have
used digitized models of SNAP-style pear and cling-peach trees to show
Harvesting is among the most labor-intensive and costly operations that one manifestation of the general notion of “easier to harvest” was
in fresh-market fruit production. Additionally, growers depend on a that most of the fruits were reachable by linear-only motions along small
farm labor workforce that has been shrinking due to various socio- sets of approach directions (Vougioukas, Arikapudi and Munic, 2016;
economic reasons (Taylor, Charlton, Yúnez-Naude, 2012). Robotic Arikapudi & Vougioukas, 2017; Arikapudi & Vougioukas, 2021). Such
harvesting can offer a solution to the labor cost and scarcity challenges. motions can be achieved by Cartesian, spherical or cylindrical-type ro­
The technical feasibility of robotic fruit picking was demonstrated in bots with telescoping arms instead of multi-degree-of-freedom harvest­
principle more than thirty years ago (Grand D’Esnon, 1985; Harrel et al., ing arms with revolute joints (Hayashi and Ueda, 1991; Burks et al.,
1985). However, fruit-harvesting robots are still in the prototype, pre- 2005; Baeten et al., 2007; Mehta and Burks, 2014; Xiong et al., 2020).
commercial stage (Bac et al., 2014; Bogue, 2020). To a large extent, Because of their simple workspace, linear, telescoping arms are also less
this lack of progress can be attributed to robots’ low cost-effectiveness complicated, cheaper, and easier to deploy on a multi-arm machine than
(high harvest cost) and the inability to harvest the wide range of exist­ multi-dof revolute joint arms. So, an approach that is gaining popularity
ing tree architectures (Bac et al., 2014; Vougioukas, 2019). - for appropriate tree architectures - is multi-arm harvesters that
To simplify the harvesting task, growers have been increasingly comprise several 3-dof linear robot arms (Schueller, 2012; Tibbets,
adopting high-density SNAP (Simple, Narrow, Accessible, and Produc­ 2018; Bogue, 2020; Barnett et al., 2020).
tive) tree architectures (Karkee & Zhang, 2012). Such trees feature Properly designed and controlled harvesters with many linear arms

* Corresponding author.
E-mail address: [email protected] (S.G. Vougioukas).

https://doi.org/10.1016/j.compag.2023.108023
Received 15 February 2021; Received in revised form 2 June 2023; Accepted 25 June 2023
Available online 6 July 2023
0168-1699/© 2023 Elsevier B.V. All rights reserved.
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Nomenclature R Number of rows of arms

Ta Time to approach the fruit
1SHN One single-sided harvester carrying N arms Tap Time to approach and pick a fruit
1THN/2 One two-sided harvester with N/2 arms on each side Tapr Time to approach, pick and retract to the robotic arm’s
2D Two dimensional initial position
2SHN/2 Two independent single-sided harvesters carrying N/2 Tc Picking cycle time
arms Td Fruit detachment time
3D Three dimensional TTSΣij Time to harvest all fruits on the ith Tree and its jth Side
Σ A fruit-picking sequence using a single arm in picking sequence Σ
Amax Maximum acceleration TTSΣijq Time to harvest all fruits within the qth work-cell on the ith
C Number of columns of arms Tree and its jth Side using a single arm in picking sequence
Dof degrees of freedom Σ
D Distance (a subscript denotes axis) TTSMij Time to harvest all fruits on the ith Tree and its jth side with
Fij Total number of Fruits on the jth side of the ith tree (j = 1 or Multiple arms
2) TO Time to harvest an Orchard block
FPT Fruit picking throughput TSk Time needed to harvest one (the kth) Single fruit
HS Height-split Tv Time of a velocity profile11
HSM Height-split-Matrix Tx Time to move in the x-direction
LS Length-split Ty Time to move in the y-direction
LSM Length-split-Matrix Tz Time to move in the z-direction
N Number of arms on the harvester TQ Time to harvest each work-cell
PCT Picking Cycle Time of a single-arm harvesting robot THT Total harvest time
PCTM Pick Cycle Time of Multi-arm harvesting robot Vmax Maximum velocity
PPP Prismatic-Prismatic-Prismatic joints (x, y, z) Coordinates of an entity in the workspace frame.

Table 1
Detailed information on the orchards where the data was collected.
Pears Peaches

Cultivar Bartlett Late Ross

Training High-density trees trained Open vase trees trained to V-
system using a trellis system shaped architecture
Plantation New orchard New orchard
Row spacing 10′ 18′
Tree-to-tree 5′ 9′
spacing
Age 12 years 7 years
Number of 20 20
trees
digitized
Total fruits on 2922 6920
all trees
Height of the 2.7′ 3.4′
tree
Name of the Chris Ruddick Ranch, Ukiah Peter Martini Ranch, Escalon
orchard
Location 39.113323, − 123.175990 37.780423, − 121.024035
Row N-S E-W
orientation Fig. 1. Example of the actual and the corresponding digitized/simulated high-
Type of Sprinkler Sprinkler density trellised pear tree.
irrigation
Dates of 07.24.17, 07.25.17, 07.26.17, 08.28.17, 08.29.17, 08.30.17,
manipulators to harvest oranges; however, they didn’t report any har­
harvest 07.28.17, 07.31.17, 08.01.17, 08.31.17, 09.01.17, 09.02.17,
08.02.17, 08.03.17, 08.04.17. 09.03.17, 09.04.17, 09.05.17. vesting performance metrics. Kondo et al. (1993) evaluated four tomato-
picking robot configurations quantitatively based on a manipulability
measure, normalized workspace volume, and the so-called redundant
(Zion et al., 2014; Mann et al., 2016) can reduce the fruit pick cycle time space. Ceres et al. (1998) developed an aided fruit harvesting robot and
(PCT), which is a key performance metric that heavily influences tested it with artificial trees with reported cycle times of 2 s under
robotic-harvest cost (Harrel, 1987). However, the PCT of robotic har­ laboratory conditions. Muscato et al. (2005) used a manipulator with
vesters depends on the interrelationships among orchard layouts, tree prismatic joints to harvest citrus fruit and reported cycle times of 5.93 s.
canopy structures, and spatial fruit distributions with robot mechanics Song et al. (2007) modeled the working space of an eggplant robot based
and control (Vougioukas, 2019). on the growth characteristics and fruit distribution of the fruits. They
Few researchers have developed harvesting simulators in the litera­ optimized two-link length parameters based on the coverage of the
ture that can estimate the PCT of robot designs using real-world data. predefined workspace. Edan et al. (1991) measured fruit locations on
Harrel et al. (1990) used a spherical manipulator to harvest oranges and twenty small orange trees and used that data to compute the fruit-
reported cycle times of about 3 s to 7 s, with 75% fruit detected. Juste picking itinerary that would minimize the travel time between succes­
et al. (1992) achieved a 90% fruit detection rate using spherical sive fruits for a single robot arm. Sivaraman and Burks (2006, 2007)

2
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 2. A linear robot arm with its gripper at the initial position (x0, y0, z0) is in front of a tree and picks a single fruit at (x1, y1, z1). a) The “approach” part of the
motion is shown on the left; b) the “extension” part of the motion is shown on the right.

used canopy volume estimates to perform robot manipulator kinematic 84%, and PCTs (per fruit, not averaged) range from 3 to 14.3 s (Bac
design, with maximum fruit reachability as a design criterion. Han et al. et al., 2014). The values of these parameters are based on very limited
(2007) analyzed the fruit distribution and harvest expectation in an (often unreported) numbers of experiments and harvested fruits using
eggplant greenhouse and designed a 4-DOF manipulator with RRRR conventional trees, i.e., non-trellised trees with large canopies. Also,
configuration, i.e., four revolute (R) joints, which was reported to be they did not report how they modeled the trees and fruits (figures show a
optimal for eggplant picking. They optimized the length of each link and rendering of one tree with a few fruits), and they did not compute
workspace volume, but no details on how they were implemented were productivity metrics or discuss design issues. Davidson et al. (2020)
reported. Van Henten et al. (2009) used workspace and manipulator reviewed 39 robotic harvesting technologies between 1985 and 2018
specifications stemming from cucumber harvesting operations to opti­ and stated that about 41% didn’t report any results related to picking
mize a three-link robot arm’s kinematic structure and link parameters. time. The other 59% reported a wide range of picking times ranging
Edan and Miles (1994) evaluated the performance of a melon-harvesting from 1 s to 23 s due to significant variations in sample sizes and testing
robot (a 2D problem) by simulating and comparing different types of conditions.
robots, the number of arms, multiple arm configurations, and workspace This paper aims to inform the design of harvesting robots utilizing
designs. Li et al. (2008) computed the workspace of the gripper of a arrays of 3-dof Cartesian/Prismatic linear arms (PPP joints). More spe­
tomato harvesting robot arm by computing the forward kinematics of cifically, this work uses digitized fruit position data to compute the PCTs
the arm within its joint limits and verified visually that the workspace is of linear arms arranged in grid configurations, i.e., operating in disjoint
large enough to contain all the tomatoes on a plant. Nguyen et al. (2013) rectangular work cells independently from all other arms 1) under
used a robotics simulator to evaluate sensing, motion, and task-planning different cell-partitioning geometries (sub-section 2.3.3) and 2) in
algorithms for a selective apple-harvesting robot. Baca et al. (2016) alternative orchard row harvesting configurations (sub-section 2.4). By
evaluated the performance of two different algorithms via simulation to computing the PCT of arrays of linear arms – a metric of harvesting
check how different parameters would affect path planning. They speed - this work complements the work reported by Arikapudi et al.,
decided that reducing the overall en- effector dimension and increasing 2021, which computed the fruit reachability of linear arms using
the sweet pepper stem spacing would improve success in reaching the computational geometrical tools. Reachability is an equivalent metric –
goals from 63% to 84%. Bloch et al. (2018) developed simulation tools under linear motion - to fruit visibility, which is an important factor
to optimize the mechanical system and the environment simultaneously. shaping fruit picking efficiency.
They pointed out that the design of the robot should match a specific
task in a specific environment to be optimal. However, as they discussed, 2. Materials and methods
they replicated a single tree to represent the entire orchard which is not
a true representation since trees’ geometrical features vary greatly. This work uses digitized fruit position data to compute the PCTs of
Overall, reported fruit-picking efficiencies (FPE)s in literature for single- linear arms that operate within disjoint rectangular work cells inde­
arm robots harvesting apple or citrus trees range anywhere from 50% to pendently from all other arms. Subsection 2.1 briefly describes the
dataset; the full description of the digitization process is given in (Ari­
kapudi et al., 2015). To calculate the PCT of several arms ‘harvesting’
1
The code is available upon request. the digitized fruit dataset, picking a single fruit with a single linear arm

3
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 3. A) in minimum-time motion, a trapezoidal velocity profile occurs when the travel distance is long enough to reach maximum velocity; otherwise, b) a
triangular velocity profile is generated.

is modeled first, and the duration of the process is calculated (subsection harvest models are: 1) fruit locations remain fixed, 2) all fruits in the
2.2). Next, the process of picking all the fruits on one side of a tree using reachable dataset are harvested in one pass, 3) the time taken by the
only one arm is modeled, and its time duration is computed (subsection harvester to move from one tree to the next is not considered, and 4)
2.3.1). Subsections 2.3.2.1 and 2.3.2.2 present in detail how the work­ grasping and detaching times are considered constant.
space of the harvester is partitioned in Q disjoint work cells when Q arms
operate in parallel in a grid configuration, and the rest of subsection
2.3.2 presents the modeling of the operation of the arms inside the work- 2.1. Tree fruit data
cells and the calculation of the time required to harvest a tree side with Q
arms and the corresponding PCT. After modeling the harvesting of a A novel system was developed, which utilized electromagnetic fields
single tree side, subsection 2.4 presents the modeling of harvesting both for collecting tree-fruit location data in commercial orchards. Manual
sides of a tree and three scenarios of harvesting orchard rows. All models digitization was carried out to obtain tree geometries, fruit size, and
were implemented in Matlab. locations from various pear and peach trees in commercial orchards in
The main picking-time-related assumptions/simplifications of the California. The accuracy of the collected data is better than 1 cm. The
topological branching structure is also encoded in the digitization

4
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 5. Multi-arm harvester and the tree with fruits.

and let the fruit be located at (x1, y1, z1), and the robot arm gripper at an
initial position (x0, y0, z0). The initial position corresponds to a config­
uration where the arm is fully retracted and has zero initial velocity and
acceleration. It is further assumed that the harvester – and the base of
the arm - is located far enough from the tree, so that when the arm is
fully retracted, its gripper can move on the x = x0 plane without inter­
fering with branches; essentially, no branch extending into the row in­
tersects the x = x0 plane. Also, we assume that the harvester’s work area
– all points on x = x0 where a gripper can go - covers the entire height
and width of the tree side.
The first part of the motion is referred to as the “approach” of the
fruit (Fig. 2 a). The fully retracted arm moves along the Y-axis, from y0 to
y1 (distance Dy), and independently along the Z-axis, from z0 to z1
(distance Dz), and comes to a full stop. Once the gripper reaches the
point (x0, y1, z1) – and not earlier, to avoid possible branch collisions -
the arm enters the second part of the motion (Fig. 2b), in which it ex­
tends along the X-axis from x0 to x1 (distance Dx), and comes to a full
Fig. 4. Fruit harvesting flow chart.
stop. The traveled distances in each axis are:

procedure, along with the steps to reconstruct the digitized trees (Ari­ Dy = |y1 − y0 | (1)
kapudi et al., 2015). Details on the collected data are listed in Table 1.
The data repository is available online (Vougioukas and Arikapudi, Dz = |z1 − z0 | (2)
2020).
A reconstructed fruit tree model is shown in Fig. 1. The fruits were Dx = |x1 − x0 | (3)
modeled using spheres, the branches were modeled using polyhedra, It is assumed that fruit detachment requires a constant time, Td (if
and reachability calculations were performed using the Bullet physics detachment time varies, an upper bound can be used). When the fruit is
engine, as described in Arikapudi et al. (2020). The reachable fruits were detached, the arm retracts along the X-axis and returns to position (x0,
imported into the MATLAB workspace and used to estimate harvesters’ y1, z1). The time required to pick a single fruit is computed as the total
PCTs with several linear arms. time the gripper takes to approach the fruit, detach it, and retract to the
x = x0 plane.
2.2. Harvesting a single fruit with one linear arm The arm’s motion in each axis is modeled as position-control “min­
imum-time” motion, with zero initial velocity and acceleration (Lynch
The picking process of harvesting a single fruit by one linear arm is and Park, 2017). Let D be the distance to be traveled. If this distance is
modeled as follows. All coordinates are expressed in the “workspace long enough for the arm to reach its maximum speed (D ≥ Vmax 2
/Amax ),
frame,” which is located at the bottom left corner of the gantry structure the arm will move at a maximum acceleration Amax until it reaches its
on which the robotic arm is mounted, as shown in Fig. 2. Let the mini­ maximum speed Vmax, where it will switch to zero acceleration; at a
mum and maximum height of the workspace be zmin and zmax, respec­ proper position before its destination, the arm will decelerate, with
tively. Also, let the outmost left and right coordinates of the workspace -Amax (“bang-coast-bang” motion). This motion results in a trapezoidal
be ymin and ymax, respectively. Let the index of the fruit to be picked be i,

5
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

The time required to cover this distance using the trapezoidal ve­
locity profile can be shown to be (Lynch and Park, 2017):
D Vmax
Tv = + (6)
Vmax Amax
To compute the time a Cartesian robotic arm needs to reach a 3D
fruit location (x1, y1, z1) from its initial position (x0, y0, z0), one must
first calculate the times Tx, Ty, and Tz it takes to move along the X, Y and
Z axes independently, using equations 5 or 6. Since the movement
during the approach stage of the motion involves independent and
simultaneous motions along the Y and Z axes, the total time to approach
and extend to the fruit (i.e., reach it) is:
Tr = Tx + max(Ty , Tz ) (7)

Once the robotic arm gripper has reached the fruit, the next step is to
detach it. The total time to reach and detach is :
Trd = Tr + Td (8)
Once the fruit is detached, the gripper needs Tx to retract from the
fruit’s x-coordinate to the gripper’s initial x-coordinate. Therefore, the
total time the gripper takes to pick a fruit (approach, extend, detach, and
retract) is:
Tp = Trd + Tx (9)

The fruit must be transported to a bin once the gripper is retracted in

the YZ plane. The implementation of this stage depends on the overall
design of the conveyance system. For example, if a horizontal conveyor
belt is below the arm (Tibbets, 2018), the arm would need to travel a
vertical distance to release the fruit on the belt. However, if a vacuum
gripper is used (Tibbets, 2018), the robot arm transports the fruit to the
bin without any motion. Therefore, we do not include the fruit-transport
time in the pick cycle time (PCT) in this work. Also, the time required to
detect a fruit is not considered; the vision system is effectively decoupled
from the mechanical harvesting system. Therefore, the total time to pick
the kth single fruit, TSk is:
TSk = Tp (10)

2.3. Harvesting one side of a tree

This section presents a model for harvesting the entire side of one
tree in an orchard row, using one and multiple linear arms, and com­
putes the associated PCTs. The jth tree side (j = 1 or 2) is harvested while
the harvester is static in front of it. In this research, the time the
harvester takes to move from one tree to the next is not considered.

2.3.1. Harvesting one side of a tree with a single linear arm

Let the number of fruits on the jth side of the ith tree be Fij. To harvest
one side of a tree using a single arm, we need to set the initial position of
the gripper and generate the sequence Σ in which the arm will harvest
Fig. 6. Fruit density histograms for (a) pears and (b) peaches. (Arikapudi, the fruits. The total harvesting time can be computed using the single-
Vougioukas, Saracoglu, 2015). fruit harvesting procedure described in the previous section applied to
the successive fruits in Σ.
velocity profile (Fig. 3a). If the travel distance is too short, then the The gripper’s initial position in the y and z-axis (y0, z0) is obtained by
acceleration switches from Amax to - Amax (“bang-bang” motion) in the averaging all the fruits’ y coordinates and z coordinates on the respec­
middle of the distance (Fig. 3b). The relation between Vmax, Amax, and D tive side of the tree. The initial position in the x-axis (x0) of the robotic
for this profile was calculated as arm gripper is located in such a position with respect to the tree that
when the arm is fully retracted, the gripper moves on the YZ plane
(Amax D)
2
Vmax = (4) without interfering with branches in the x-axis.
2
Regarding the picking sequence, let Σ = {σ(1), σ(2), …, σ(Fij)} be the
The time, Tv , required to cover distance D using the triangular ve­ fruit picking sequence. For example, if σ(1) = 10, fruit number 10 is
locity profile is: picked first; if σ(2) = 75, fruit number 75 is picked second, etc. The
√̅̅̅̅̅̅̅̅̅ sequence resulting in near-minimum travel distance could be computed
D
Tv = 2* (5) using a solver for the Traveling Salesman Problem (TSP) (Edan et al.,
Amax
1991). However, the large number of fruits on each tree side would
result in excessive TSP solution times and long simulation times. In this

6
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 7. Examples of equal fruit-load work-cell partitioning. (a) Height-split; (b) Length-split; (c) Height-split-matrix, and (d) Length-split-matrix configurations.

Fig. 8. (a) Total number of fruits along the Z -axis. (b) Cumulative distribution function along the Z-axis.

work, at the arm’s current x-coordinate, the fruits are sorted by 2.3.2. Harvesting one side of a tree with multiple linear-arms
increasing the y-coordinate and are presented to the robotic arm gripper In this section, the process of harvesting one side of a tree with
in that sequence. A flowchart of the tasks is shown next (See Fig. 4): several linear arms is studied as a function of the number of arms and the
The PCT for a tree side using a single-arm harvester is computed as spatial configurations of the arms, i.e., the geometric arrangement of the
follows: the time TSk to harvest the σ(k)th fruit (k = 1, 2, …, Fij) is work cells inside which the arms move. To harvest a tree-side using
calculated using equation (10). The time TTSΣij required to harvest Fij several linear arms, the workspace for each arm is set by partitioning the
fruits on one side of the tree using the picking sequence Σ is: entire tree-side workspace into N disjoint work cells such that each arm
is allocated to one cell, as shown in Fig. 5.
Fij
∑
TTSΣij = TSσ(k) (11) It is known that the distribution of fruits in tree canopies is not
k=1 uniform (Arikapudi, Vougioukas, Saracoglu, 2015) (e.g., Fig. 6).
Therefore, partitioning the work area into work cells (the area an
The PCT for a single-arm harvester is calculated by dividing the total
individual arm operates in) is very important, as work cells may have
time required to harvest the fruits by the number of harvested fruit on
different numbers of fruits, and the amount of work per arm inside each
the tree side:
work cell will vary. Consequently, some arms may finish picking earlier
TTSΣij than others and remain idle, thus affecting the overall PCTM.
PCT = (12)
Fij The partitioning of the workspace is done using two methods. In the
first method, the workspace is partitioned into N work cells of equal fruit
load (number of fruits is 1/N). When the numbers of fruits in each cell
can’t be exactly equal, they are as close as possible to 1/N. In the second

7
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 9. Equal-sized work-cell partitioning configurations ((a) Height-split; (b) Length-split; (c) Height-split-matrix, and (d) Length-split-matrix configuration.

Fig. 10. Side view of the multi-arm harvester with two tree sides.

method, the workspace is partitioned so that the work cells have equal 2.3.2.1. Equal fruit load work-cell partitioning. Here, we explain how N
size. The equal-load method is meant to distribute the fruit load equally work cells are created – one for each robot arm - based on the equal fruit-
among arms (workload balancing). In contrast, the equal-size method load principle (method 1) for each of the four partitioning
provides a way to configure the arms on a harvester when fruit load configurations.
distributions are not known in advance.
Four different configurations for partitioning the workspace into a) Equal Load Height-split partitioning
work cells were tested (Fig. 7): Length-split (LS) configuration (1D array
of work-cells), Height-split (HS) configuration (1D array of work-cells), The arms are placed one above the other in a 1-D vertical configu­
Length-split-Matrix (LSM) configuration (2D array of work-cells), and ration. The total height of the workspace is split into N horizontal work
Height-split-Matrix (HSM) configuration (2D array of work-cells). cells (Fig. 7a), and each work cell contains as close as possible to 1/N
The coordinates of the bottom-left and top-right corners of the work fruits. To calculate the bottom and top heights [zq-1 zq ] of the qth work
cell define the geometry of each (rectangular) cell. In the next sub­ cell, the frequency histogram and the cumulative distribution function
sections, we explain in detail how these coordinates are calculated using (cdf(z)) of the fruits inside the workspace is calculated based on the
the above principles and partitioning schemes. fruits’ z-coordinates, i.e., their heights in the tree canopy (Fig. 8). The
coordinates of all fruits are available from the digitization process

8
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

and top coordinates of the qth zone correspond to the [zq-1 zq ] co­
ordinates of a work cell that contains 1/N of the fruits. For example, for
the histogram shown in Fig. 8, and four arms, the first (q = 1) y-axis zone
[0, 0.25) corresponds to the 1st (bottom) work-cell spanning the height
interval [zmin z1], with zmin = cdf-1(0) = 0 and z1 = cdf-1(0.25) = 100 cm.
So, the 1st work cell will start at 0 cm and extend vertically until 100 cm.
The zone [0.25, 0.5) corresponds to the work cell directly above the
bottom one (q = 2), which extends from z1 = cdf-1(0.25) = 100 cm to z2
= cdf-1(0.5) = 155 cm. The work-cell above it extends from a height of
z2 = cdf-1(0.5) = 155 cm to z3 = cdf-1(0.75) = 190 cm and the top work-
cell spans z3 = 190 to z4 = zmax = cdf-1(1.0) = 265 cm. In general, the
coordinates of the bottom-left and top-right corners of the qth work-cell
are [ymin, cdf-1((q-1)/N)] and [ymax, cdf-1(q/N)].

b) Equal Load Length-split partitioning

The arms are placed one next to the other in a 1-D horizontal
configuration. The total length of the workspace is split into N vertical
work cells (Fig. 7b), and each work cell contains as close as possible to 1/
N fruits. The outmost left and right y-coordinates [yq-1 yq ] of the qth
work-cell are calculated in the same way as the [zq-1 zq ] z-coordinates of
the qth work-cell in height-split, i.e., using the cumulative distribution
of the fruits, albeit along the y-axis. Therefore, the coordinates of the
bottom-left and top-right corners of the qth work-cell are [cdf-1((q-1)/
N), zmin,] and [cdf-1(q/N), ymax].

Fig. 11. Fruit harvesting using one single-sided harvester with N arms. c) Equal Load Height-split Matrix partitioning

An even number of arms is placed in a 2-D array/grid configuration.

(Section 2.5). By definition, cdf(zmin) = 0 and cfd(zmax) = 1.0, and the
First, the equal load height-split partitioning is calculated for N/2 arms
cumulative distribution’s horizontal axis corresponds to fruit z-co­
following the procedure described in subsection 2.3.2.1-a. The result is
ordinates, with zmin = cdf-1(0), zmax = cdf-1(1). The vertical axis of the
N/2 horizontal workspace zones that have an equal number of fruits.
cumulative distribution (Fig. 8b) has a range from 0 to 1 and can be split
Next, these zones are split into two equal-load work cells using the
into Q equal zones, with each zone “containing” 1/N th of the fruits; for
equal-load length-split partitioning method described in subsection
example, if we used four arms (i.e., N = 4), each zone would contain
2.3.2.1-b. Therefore, the coordinates of the bottom-left and top-right
25% of the fruits. Since each zone contains 1/N of the fruits, the bottom

Fig. 12. Fruit harvesting using one two-sided harvester with N/2 arms on each side.

9
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

height-split partitioning method described in subsection 2.3.2.1-a.

Therefore, the coordinates of the bottom-left and top-right corners of
the bottom work cell in the qth column (q = 1,.., N/2) are [ymin + cdf-1(2
(q-1)/N), zmin] and [ymin + cdf-1(2q/N), zmin+(zmax-zmin)/2]. The cor­
responding coordinates of the rightmost work-cell in the qth row are
[ymin + cdf-1(2(q-1)/N), zmin+(zmax-zmin)/2] and [ymin + cdf-1(2q/N),
zmax].

2.3.2.2. Equal size work-cell partitioning. Next, we explain how N work-

cells are set – one for each robot arm - based on the equal work-cell size
principle (method 2) for each of the four partitioning configurations
(Fig. 9).
In Height-split partitioning, the height of the entire workspace is split
into N segments, and each segment has a height equal to (zmax-zmin)/N
and spans the length of the workspace. So, the y-z coordinates of the
bottom-left corner of the qth work-cell are [ymin, zmin+(q-1)(zmax-zmin)/
N], and those of the top-right corner are [ymax, zmin + q(zmax-zmin)/N].
The same scheme is applied in Length-split partitioning, albeit along
the y-axis. So, the y-z coordinates of the bottom-left corner of the qth
work-cell are [ymin+(q-1)(ymax-ymin)/N, zmin], and those of the top-right
corner are [q(zq-zmin)/N, zmax ].
In Matrix partitioning, the workspace is partitioned into R rows and
C columns, where R*C = N. The workspace is split into R horizontal
segments using the Height-split scheme, and each segment is split into C
equal-sized work cells. So, the coordinates of the bottom-left corner of
the ijth work-cell are [ymin+(j-1)(ymax-ymin)/C, zmin+(i-1)(zmax-zmin)/R]
and those of the top-right corner [ymin + j(ymax-ymin)/C, zmin + i(zmax-
Fig. 13. Fruit harvesting using two independent single-sided harvesters with zmin)/R]. When there are more rows than columns (R > C), this scheme
N/2 arms. is referred to as Height-split matrix partitioning, whereas when there are
more columns, the scheme is referred to as Length-split matrix
corners of the leftmost work cell in the qth row (q = 1,..N/2) are [zmin + partitioning.
cdf-1(2(q-1)/N), ymin] and [zmin + cdf-1(2q/N), ymin+(ymax-ymin)/2]. The The next step is to set the initial position for each of the robotic arms
corresponding coordinates of the rightmost work-cell in the qth row are within their allocated work cell. We follow the single-arm method
[zmin + cdf-1(2(q-1)/N), ymin+(ymax-ymin)/2] and [zmin + cdf-1(2q/N), (Section 2.3.1) applied in each work cell. Namely, the × coordinate of
ymax]. the gripper for each arm is set to x0. We compute the mean values yq and
zq of all the work cell’s fruits’ y and z coordinates, respectively. The
d) Equal Load Length-split Matrix partitioning initial position of the gripper is set to (x0, yq, zq). The steps of sorting and
harvesting fruits in each work cell are the same as those for a single-arm
An even number of arms is placed in a 2-D array/grid configuration. harvester.
First, the equal load length-split partitioning is calculated for N/2 arms The time TTSΣijq to harvest the fruits within the qth work-cell of the
following the procedure described in subsection 2.3.2.1-b. The result is jth side of the ith tree is calculated using equation (11). The maximum
N/2 workspace zones that have an equal number of fruits. Next, these time to harvest inside the N work cells of a tree side is defined as the
zones are split into two equal-load work cells using the equal-load time, TTSMij , to harvest the entire tree side with multiple arms.

Fig. 14. PCT of the harvester to harvest pears as a function of Vmax for a given Amax.

10
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 15. PCT of the harvester as a function of Vmax for a given Amax to harvest peaches.

Fig. 16. PCTM as a function of the number of arms harvesting pears under different configurations with equal-sized work cells.

2.4. Harvesting Tree-rows and associated PCTM

Table 2
Results of the t-test for PCTM harvesting pears for each combination of the Harvesting a tree row involves harvesting both sides of each tree and
different configurations with equal-sized work cells. all trees in the orchard row. In the previous section, harvesting a tree-
Configuration p-value Decision side was explained. In this section, harvesting an entire tree, i.e., both
HS vs. HSM 0.62 Accept Null sides, is studied, followed by harvesting adjacent trees in the orchard
HS vs. LS 0.91 Accept Null row.
HS vs LSM 0.95 Accept Null To harvest a tree using a harvester with one or multi-arms, first, the
HSM vs. LS 0.67 Accept Null side of the tree that needs to be harvested is selected. Based on the
HSM v.s LSM 0.12 Accept Null
LS vs LSM 0.98 Accept Null
selected tree side, the initial position of the gripper for the robotic arms
and the work cells in which the robotic arms operate are calculated. The
next steps of the harvesting process follow the tree-side harvesting
( )
TTSMij = max TTSΣij1 , TTSΣij2 , TTSΣij3 , ⋯., TTSΣijN (13) detailed in the previous section.
For the side of the tree that has fruits with × coordinates greater than
The PCT of the multi-arm harvester (PCTM) for a tree side is calcu­ or equal to zero (to the right of the harvester), the harvester is stationed
lated by dividing the time to harvest the tree side by the number of at the YZ plane on the positive x-axis with the robotic arms facing the
harvested fruit (Fij) as follows. fruits in the current orchard row, as shown in Fig. 10. The initial position
TTSMij in the x-axis is (x0) for the gripper of the robotic arm positioned in the
PCTM = (14)
Fij current orchard row. For the side of the tree that has fruits with × co­
ordinates less than zero (left side of the harvester), the harvester is
The idle time of the harvester is estimated as follows: The picking stationed at the YZ plane on the negative x-axis with the robotic arm
time for each arm to harvest all the fruits that are reachable from inside facing the fruits in the next orchard row, as shown in Fig. 10. The initial
its work cell is computed using the single-arm harvesting method, and so position in the x-axis is (-x0) for the gripper of the robotic arm positioned
are the maximum and minimum times among the N arms to pick all their in the next orchard row. The robot arm gripper’s initial y and z co­
corresponding fruits. The difference between the maximum and mini­ ordinates are calculated using the same method detailed in the previous
mum times is defined as the ‘idle time’ of the harvester. section. The next steps of calculating the work cells based on the arms
and fruit harvest sequence are the same as those of single-side

11
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 17. PCTM as a function of the number of arms harvesting peaches under different configurations with equal-sized work cells.

the final PCTM for this harvesting scenario.

Table 3 In the second harvesting scenario (referred to as “1THN/2”), one
Results of the t-test for PCTM harvesting peaches for each combination of the harvester is “two-sided,” i.e., it carries N/2 arms on each side. Hence, it
different configurations with equal-sized work cells. picks two tree halves/sides (each having a different fruit distribution)
Configuration p-value Decision simultaneously as it traverses the orchard row. This scenario corre­
HS vs. HSM 0.01 Reject Null sponds to two possible machine configurations. In the first one, the
HS vs. LS 0.00 Reject Null machine straddles trees to pick both sides at once, as shown in Fig. 12
HS vs. LSM 0.01 Reject Null (left); in the second, the machine picks two opposite sides of the row at
HSM vs. LS 0.58 Accept Null once Fig. 12 (right). The time taken on the left side is calculated as that
HSM vs. LSM 0.62 Accept Null
LS vs. LSM 0.43 Accept Null
of a multi-arm harvester scenario for each tree side. In the same way, the
time taken on the right side is calculated. The maximum time of the two
sides is defined as the time to harvest that tree. The overall PCTM to
harvest all the fruits is obtained by adding the times for individual trees
Table 4 and dividing the sum by the number of harvested fruits from all the
Pear fruit distribution in each work cell under different configurations with trees.
equal fruit loads. In the third harvesting scenario (referred to as “2SHN/2”), two in­
Number of Configuration Cell Cell Cell Cell Cell Cell dependent single-sided harvesters with N/2 arms pick each of the two
arms 1 2 3 4 5 6 tree halves/sides, as shown in Fig. 13. Here, the PCTMs for the two
2 Height split 1121 1146 single-sided harvesters are calculated separately. The PCTM for each
2 Height split 1124 1143 harvester is calculated by adding the times for individual tree sides and
matrix
dividing by the total number of harvested fruits. The average of the two
4 Height split 553 568 563 583
4 Height split 553 568 562 584 PCTMs is calculated and is defined as the PCTM for this harvesting
matrix scenario.
6 Height split 362 376 383 376 368 402
6 Height split 360 378 368 391 373 397
matrix 2.5. Experimental design

All the simulations used fruit position data from twenty pear and
harvesting.
twenty peach trees (available at Vougioukas et al., 2020). The fruit
This completes the harvesting process of one tree. The same steps are
positions were manually digitized using a PowerTRAK 360™ digitizer
repeated to harvest the next tree and all the trees in the orchard row.
(Polhemus, Colchester, VT, USA) and established geometric modeling
However, the collected tree-fruit data (Arikapudi et al., 2015) indicated
techniques (Arikapudi et al., 2015). A wide range of actuation system
that the fruit load varies between the trees and within the sides of the
types is commercially available to produce linear motion, such as screw,
trees. Also, in commercial orchards, there might not be enough room for
rack-pinion, belt transmission, rod-in-cylinder, etc. These systems offer a
all the arms on the robot to pick fruit without staggering the arm loca­
wide range of maximum velocity (Vmax) and acceleration (Amax)
tions. Therefore, how the harvesting of the trees takes place will have an
specifications.
impact on overall PCTMs.
In this work, we first investigated the effects of Amax and Vmax on the
Three harvesting scenarios were considered to improve the PCTMs
PCTM of a multi-arm harvester for values of Amax between 10 and 150 m
while harvesting trees in an orchard row. In the first scenario – referred
s− 2 and Vmax between 1 and 10 m s− 1 (sub-section 3.1).
to as “1SHN” - one harvester carries N arms on one of its sides. Hence, it
Next, the PCTM was computed for an array of arms, where each arm
picks the fruits from one side of the trees in its traversing row. After
moved inside a work cell, and each work cell had: 1) equal size (sub-
turning to the neighboring row, it picks the other side of the trees, as
section 3.2) or 2) an equal number of fruits inside it (sub-section 3.3).
shown in Fig. 11. The times needed to harvest individual tree sides are
For each case, four configurations were used to partition the workspace
added. The sum is divided by the number of harvested fruits to obtain
into work cells: length-split, height-split, height-split-matrix, and

12
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 18. PCTM as a function of the number of arms harvesting pears under different configurations with equal fruit load cells.

3. Results
Table 5
Results of the t-test for PCTM harvesting pears for each combination of the This section presents the effect of actuator design specifications, such
different configurations with equal fruit load.
as maximum acceleration, maximum velocity, and fruit detachment
Configuration p-value Decision time on PCTs for a harvester equipped with one 3-DOF linear robotic
HS vs. HSM 0.00 Reject Null arm to harvest pears and peaches. Next, the PCTMs for two spatial
HS v.s LS 0.79 Accept Null configurations (cells with equal size vs. cells with equal fruit load) are
HS vs. LSM 0.00 Reject Null presented. Finally, the PCTM for three harvesting scenarios (harvesting
HSM vs. LS 0.00 Reject Null
HSM vs. LSM 0.19 Accept Null
one side with one harvester equipped with N arms vs. harvesting both
LS vs. LSM 0.00 Reject Null sides with one harvester equipped with N/2 arms on each side vs. har­
vesting each side with two harvesters, each equipped with N/2 arms) to
harvest both pears and peaches are presented.
length-split-matrix (see Figs. 7 and 9). The effect of each split configu­
ration on the PCTM was computed for pears and peaches using Vmax = 5
m s− 1 and Amax = 50 m s− 2 for the Y-axis and Z-axis, and Vmax = 2 m s− 1 3.1. Effects of maximum linear acceleration and maximum linear velocity
and Amax = 10 m s− 2 for the X-axis. The fruit detachment time was set to on PCTs
0.25 s.
Finally (in sub-section 3.4), the PCTM was estimated for the three The effect of maximum linear velocity for different maximum linear
different harvesting scenarios presented in sub-section 2.4, i.e., 1SHN, accelerations on PCTs is presented, and the results are shown in Fig. 14
1THN/2, and 2SHN/2. and Fig. 15.
Fig. 14 and Fig. 15 indicate that for any given combination of Vmax

Fig. 19. PCTM as a function of the number of arms harvesting peaches under different configurations with equal fruit load cells.

13
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Table 6 Height-split-Matrix (HSM)). The null hypothesis states no significant

Peach fruit distribution in each work cell under different configurations with difference in the PCTM between different configurations. The results of
equal fruit load. the t-test are given in Table 2.
Number of Configuration Cell Cell Cell Cell Cell Cell The p-value suggests no significant difference in the PCTM between
arms 1 2 3 4 5 6 different configurations. The PCTM harvesting peach fruits under
2 Height split 2736 2759 different configurations for equal-sized cells is shown in Fig. 17.
2 Height split 2734 2761 Results indicate that the height-split configuration yielded lower
matrix PCTM than the other configurations. The PCT for a 4-arm harvester had
4 Height split 1353 1383 1368 1391
yielded low PCTM for the length-split-matrix and height-split-matrix
4 Height split 1355 1381 1369 1390
matrix compared to the height-split and length-split schemes, as observed in
6 Height split 899 917 920 910 910 939 pears. The PCTM followed a power law as a function of the number of
6 Height split 897 919 905 925 915 934 arms per machine. A t-test was conducted to check if there is any sig­
matrix nificant difference in the PCTM between different configurations
meaning that the PCTM obtained for each of these configurations is
different statistically, and the probability of obtaining that difference by
Table 7 chance is very small. The null hypothesis states no significant difference
Results of the t-test for PCTM harvesting peaches for each combination of the in the PCTM between different configurations. The results of the t-test
different configurations with equal fruit load. are given in Table 3.
Configuration p-value Decision
The p-value suggests no significant difference in the PCTM between
height-split-matrix, length-split, and length-split-matrix configurations.
HS vs. HSM 0.14 Accept Null
However, there is significantly difference in PCTM between the height-
HS vs. LS 0.74 Accept Null
HS vs. LSM 0.08 Accept Null split and the rest of the configurations.
HSM vs. LS 0.00 Reject Null
HSM vs. LSM 0.49 Accept Null 3.3. PCTM for work cells with equal fruit loads
LS vs. LSM 0.61 Accept Null

The number of fruits in each cell is given in Table 4. The PCTM

and Amax, the PCT of the harvester for pears is lower than that of harvesting pear fruits under different configurations with cells having
peaches. The reason is that the average distance traveled to harvest the (almost) equal fruit load is shown in Fig. 18.
fruit is 2.4 m in pears, whereas, for peaches, it is 4.4 m. Also, it is A t-test was conducted to check for any significant difference in the
observed that PCT decreased rapidly at the start and evened out after PCTM between different configurations. The null hypothesis states no
reaching a certain threshold maximum velocity Vtmax for a given Amax. significant difference in the PCTM between different configurations. The
For example, for Amax = 10 m s− 2, the PCT decreased rapidly below 4 m results of the t-test are given in Table 5.
s− 1 and stayed constant afterward. The reason is that, for a constant The PCTM harvesting peach fruits under different configurations for
maximum acceleration as the maximum velocity increased, the time cells with equal fruit loads is shown in Fig. 19, and the number of
taken to cover the same distance was reduced based on equation (5). At peaches in each cell is shown in Table 6.
the same time, if the distance that needs to be traveled is short, the A t-test was conducted to check if there was any significant difference
actuator may not reach the maximum velocity, which results in a in the PCTM between different configurations. The null hypothesis states
triangular velocity profile. The regression models shown in the graph a difference in the PCTM between different configurations. The results of
show how the PCT changes as a function of Vmax for a constant Amax the t-test are given in Table 7.
equal to 150 ms− 2. The regression functions can predict PCT for a given For equal fruit load work cells, as seen from Table 4 and Table 6, the
maximum velocity. The R2 coefficient of determination, which gives the fruit count is not the same in each of the wwork celllocation of fruit to
statistical measure of how well the regression predictions approximate the work cells varies based on the resolution used to divide the fruit
the data points for this model, is greater than 0.9 for both peaches and among the work cells. Since the PCTM depends on the number of fruits
pears, indicating a strong relationship between the velocity and PCT. allocated in each work cell, no conclusions can be drawn from the t-test
Based on equation (4), at a constant distance D, the peak velocity results.
attained by the robotic arm is limited for a given Amax. Therefore, there
is no improvement in PCT after reaching the peak velocity (maximum 3.4. PCTM harvesting one side vs. Both tree sides
threshold velocity). For example, in pears, with an average distance D of
2.4 m at Amax = 10 m s− 2, the Vtmax obtained is 3.5 m s− 1. Therefore, Results for the 1st harvesting scenario referred to as one single-sided
selecting an actuator that can attain a Vmax of 4 m s− 1 for a given Amax of harvester with N arms (1SHN); the 2nd harvesting scenario referred to as
10 m s− 2 cannot improve PCTs. A similar effect is observed in peaches. one two-sided harvester with N/2 arms on each side (1THN/2); the 3rd
For example, in peaches, with an average distance D of 4.4 m at Amax = harvesting scenario, referred to as two independent single sided har­
25 m s− 2, Vtmax obtained is 7.4 m s− 1. This information could help size vesters with N/2 arms (2SHN/2) are presented in this section. Results
the robotic arm motors. indicate that PCTMs achieved for all the scenarios followed a decaying
power law as a function of the number of arms (See Figs. 20 and 21).
The PCTMs achieved for the 1THN/2 harvesting scenario are higher
3.2. PCTM for work cells with equal-sized cells than the PCTMs achieved by 1SHN and 2SHN/2 scenarios for the case of
equal fruit load allocation to arms. The reason is that since the fruit loads
The PCTM harvesting pear fruits under four different configurations on the two sides of a tree are generally not equal, the advancement of the
of the work area, with equal-sized work cells, are shown in Fig. 16. harvester in the orchard row is slowed down by the side having more
The PCT of a 4-arm harvester yielded low PCTM for the length-split- fruit. Therefore, the 1THN/2 scenario took more time to harvest the
matrix and height-split-matrix compared to the height-split and length- orchard row, thereby having higher PCTMs.
split schemes. The reason is that the fruit load is distributed better in a 4- The PCTMs achieved for the 2SHN/2 harvesting scenario are lower
arm matrix configuration. A t-test was conducted to check if there is any than the PCTMs achieved by 1SHN and 1THN/2 scenarios for the case of
significant difference in the PCTM between different configurations equal cell size allocation to arms. The reason for the 2SHN/2 scenario
(Length-split (LS), Height-split (HS), Length-split-Matrix (LSM), and having lower PCTMs compared to the 1THN/2 scenario is again that

14
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 20. PCTM as a function of the number of arms harvesting peaches a) work-cells based on equal fruit load; b) work-cells based on equal size for three har­
vesting scenarios.

since the fruit load on both sides of the tree is different, the advancement 4. Summary and conclusions
of the harvester in the orchard row is slowed down by the side having
more fruit. Also, the PCTMs obtained using the 2SHN/2 scenario are This paper presented simulation studies that estimated the PCTMs of
lower compared to the 1SHN scenario because the number of arms on a multi-arm harvesters comprising 3-dof linear arms arranged in a 1D or
harvester in the latter scenario is higher, and both scenarios followed a 2D grid configuration. Each arm moved independently of the others
decaying power law as a function of the number of arms. inside a work cell. Digitized fruit locations from high-density pear and
Since the distribution between PCTM and the number of arms was a peach trees were used in the simulations.
power law, the harvester with N arms (1SHN scenario) would have First, the effects of maximum motor acceleration, Amax, and Vmax
higher PCTM than 2 harvesters with N/2 arms (2SHN/2) harvesting the velocity on the PCTM were investigated. As Vmax increased, the PCT
same data set. This was true for pears and peaches in work cells with followed a negative exponent power law (diminishing return) for any
equal cell size and fruit load. given Amax: it decreased rapidly initially and plateaued as Vtmax
A t-test was conducted to check for any significant difference in the increased. Similarly, for a constant Vmax, the improvement of PCT as
PCTM between different harvest scenarios (1SHN, 1THN/2, 2SHN/2). Amax increased slowed down at higher acceleration values. Also, similar
The null hypothesis states no significant difference in the PCTM between PCTs could be achieved with combinations of low (higher) Vmax and
different harvest scenarios. The results of the t-test are given in Table 8. higher (lower) Amax.
The p-value suggests a significant difference in the PCTM between Next, the effect of arranging the same number of arms in different
the three harvesting scenarios for work cells of equal fruit load. In the grid configurations/partitions on fruit PCTM was presented. Results
case of equal-sized cells, there was a significant difference in PCTM indicated that the height-split configuration had the lowest PCTMs
between the 2SHN/2 harvesting scenario and the 1SHN and 1THN/2 among the four different partitioning schemes for pears and peaches,
harvesting scenarios. There was no significant difference in PCTM for regardless of whether the work cells had an equal size or fruit load.
equal-size work cells between 1SHN and 1THN/2 harvesting scenarios. Additionally, it was shown that partitioning cells to contain equal
numbers of fruit resulted in lower PCTMs than if the cells had equal
sizes, presumably, because of better load balancing. Finally, the results

15
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Fig. 21. PCTM as a function of the number of arms harvesting pears a) work-cells based on equal fruit load; b) work-cells based on equal size for three harvest­
ing scenarios.

augmented by economic and operations factors - can help make design

Table 8 decisions, such as motor sizing and arm placement, that increase the
Results of the t-test for each combination of the different harvesting scenarios for cost-effectiveness of multi-armed robotic harvesters.
harvesting pears and peaches under equal cell size and fruit load.
Pears equal cell size
1SHN vs 1THN/2 1SHN vs 2SHN/2 1THN/2 vs 2SHN/2 Declaration of Competing Interest
p-value 0.13 0.00 0.00
Decision Accept Null Reject Null Reject Null The authors declare that they have no known competing financial
Peaches equal cell size interests or personal relationships that could have appeared to influence
1SHN vs 1THN/2 1SHN vs 2SHN/2 1THN/2 vs 2SHN/2 the work reported in this paper.
p-value 0.09 0.00 0.00
Decision Accept Null Reject Null Reject Null
Pears equal fruit load Data availability
1SHN vs 1THN/2 1SHN vs 2SHN/2 1THN/2 vs 2SHN/2
p-value 0.02 0.01 0.00 Data will be made available on request.
Decision Reject Null Reject Null Reject Null
Peaches equal fruit load
1SHN vs 1THN/2 1SHN vs 2SHN/2 1THN/2 vs 2SHN/2 Acknowledgments
p-value 0.00 0.00 0.00
Decision Reject Null Reject Null Reject Null We thank numerous California growers for letting us gather data
during their busy harvesting season. This work was partly funded by
USDA-NIFA Grants 2016-67021-24532 and 2020-67021-30759.
showed that, on average, harvesting each side of the peach and pear
trees separately resulted in lower PCTMs than harvesting both sides
concurrently, regardless of whether the work-cells had equal size or References
equal fruit load.
Arikapudi, R., & Vougioukas, G S. (2017). Linear Reachability of Fruits in Orchard Trees
Overall, the information provided by the proposed methodology – if with Actuator Constraints. 2017 ASABE Annual International Meeting 1701421.

16
R. Arikapudi and S.G. Vougioukas Computers and Electronics in Agriculture 211 (2023) 108023

Arikapudi, R., & Vougioukas, G S. (2021). Robotic Tree-fruit Harvesting with Kondo, N., Monta, M., Shibano, Y., and K. Mohri. 1993. Basic mechanism of robot
Telescoping Arms: A study of Linear Fruit Reachability under Geometric Constraints. adapted to physical properties of the tomato plant. In: Proceedings of the
2021 IEEE Access. International Conference for Agricultural Machinery & Process Engineering, Seoul,
Arikapudi, R., Vougioukas, S., Saracoglu, T., 2015. Orchard tree digitization for Korea, pp. 840–849.
structural-geometrical modeling. In: Precision agriculture’15. Wageningen Li, Z., Liu, J., Li, P., and Li, W. (2008). Analysis of workspace and kinematics for a tomato
Academic Publishers, pp. 161–168. harvesting robot. In International Conference on Intelligent Computation
Bac, C.W., Henten, E.J., Hemming, J., Edan, Y., 2014. Harvesting Robots for High-value Technology and Automation (ICICTA), 2008 volume 1, 823–827. IEEE.
Crops: State-of-the-art Review and Challenges Ahead. J Field Rob 31 (6), 888–911. Lynch, K.M., Park, F.C., 2017. Modern Robotics. Cambridge University Press.
Bac, C.W., Roorda, T., Reshef, R., Berman, S., Hemming, J., Van Henten, E.J., 2016. Mann, M.P., Zion, B., Shmulevich, I., Rubinstein, D., Linker, R., 2016. Combinatorial
Analysis of a motion planning problem for fruit harvesting in a dense obstacle optimization and performance analysis of a multi-arm cartesian robotic fruit
environment. Biosyst Eng 146, 85–95. harvester—extensions of graph coloring. J Intell Rob Syst 82 (3–4), 399–411.
Johan Baeten, Kevin Donné, Sven Boedrij, Wim Beckers, Eric Claesen. Autonomous Fruit Mehta, S.S., Burks, T.F., 2014. Comput Electron Agric 102, 146–158.
Picking Machine: A Robotic Apple Harvester. 6th International Conference on Field Muscato, G., Prestifilippo, M., 2005. A prototype of an orange picking robot: Past history,
and Service Robotics FSR 2007, Jul 2007, Chamonix, France. inria-00194739. the new robot and experimental results. Industrial Robot: An International Journal
Barnett, J., Duke, M., Au, C.K., Lim, S.H., 2020. Work distribution of multiple Cartesian 32 (2), 128–138.
robot arms for kiwifruit harvesting. Comput Electron Agric 169, 105202. Nguyen, T. T., Kayacan, E., De Baerdemaeker, J., and Saeys, W. 2013. Task and Motion
Bloch, V., Degani, A., Bechar, A., 2018. A methodology of orchard architecture design for Planning for Apple Harvesting Robot. In Proceedings of the IFAC Agricontrol
an optimal harvesting robot. Biosyst Eng 166, 126–137. Conference, 4(1): 247-252.
Bogue, R., 2020. Fruit-picking robots: Has their time come? Ind Robot 47 (2), 141–145. Schueller, J.K., 2012. Another report from 25 years hence. ASABE Resources Magazine
Burks, T. , Villegas, F. , Hannan, M. , Flood, S. , Sivaraman, B. , Subramanian, V. , and 19 (4), 16.
Sikes, J. 2005. HortTechnology. January-March 2005 15(1). Sivaraman, B., Burks, T.F. 2007. Robot Manipulator for Citrus Harvesting: Configuration
Ceres, R., et al., 1998. Design and implementation of an aided fruit-harvesting robot Selection. Transactions of the ASABE.
(Agribot). Ind Robot 25 (5), 337–346. Sivaraman, B., Burks, T.F., 2006. Geometric performance indices for analysis and
Davidson, J., Bhusal, S., Mo, C., Karkee, M., Zhang, Q., 2020. Robotic Manipulation for synthesis of manipulators for robotic harvesting. Trans ASABE 49 (5), 1589–1597.
Specialty Crop Harvesting: A Review of Manipulator and End-Effector Technologies. Song, J., Sun, X., Zhang, T., Zhang, B., Xu, L., 2007. Design optimization and simulation
Global Journal of Agricultural and Allied Sciences 2 (1), 25–41. of structure parameters of an eggplant picking robot. N Z J Agric Res 50, 959–964.
Edan, Y., Flash, T., Peiper, U.M., Shmulevich, I., Sarig, Y., 1991. Near-minimum-time Taylor, J.E., Charlton, D., Yúnez-Naude, A., 2012. The end of farm labor abundance.
task planning for fruit-picking robots. IEEE Trans Rob Autom 7 (1), 48–56. Appl Econ Perspect Policy 34 (4), 587–598.
Edan, Y., Miles, G.E., 1994. Systems engineering of agricultural robot design. IEEE Trans Tibbets, J.H., 2018. Not too Far from the Tree. Mech Eng 140 (02), 28–33.
Syst Man Cybern 24 (8), 1259–1265. Van Henten, E.J., Van’t Slot, D.A., Hol, C.W.J., Van Willigenburg, L.G., 2009. Optimal
Grand D’Esnon, A., 1985. Robotic harvesting of apples. Proc. Agri-Mation 1, 210–214. manipulator design for a cucumber harvesting robot. Comput Electron Agric 65 (2),
Han, S., Xueyan, S., Tiezhong, Z., Bin, Z., Liming, X., 2007. Design optimization and 247–257.
simulation of structure parameters of an eggplant picking robot. NewZealand Vougioukas, S., and Arikapudi, R. “Digitized pear and peach trees with fruits, v2, UC
Journal of Agricultural Research 50 (5), 959–964. Davis,” Dataset, 2020. https://doi.org/10.25338/B8WW41.
Harrell, R., 1987. Economic analysis of robotic citrus harvesting in Florida. Transactions Vougioukas, S., Arikapudi, R., Munic, J., 2016. A study of fruit reachability in orchard
of the ASAE. 30 (2), 298–304. trees by linear-only motion. International Federation of Automatic Control -
Harrell, R.C., Adsit, P.D., Munilla, R.D., Slaughter, D., 1990. C, Robotic picking of citrus. PapersOnLine 49–16 (2016), 277–280.
Robotica 8, 269–278. Vougioukas, S.G. 2019. Agricultural Robotics. Annual Review of Control, Robotics, and
Harrell, R. C., Adist, P. D., Slaughter, D. C. 1985. Real-time vision-serving of a robotic Autonomous Systems, 2: 365-392. doi.org/10.1146/annurev-control-053018-
tree fruit harvester. ASAE Paper 85-3550, St Joseph, MI 49085. 023617.
Hayashi, M., Ueda, Y. 1991. Orange harvesting robot. “Kubota” Co., Sakai, Japan. Xiong, Y., Ge, Y., Grimstad, L., From, P.J., 2020. An autonomous strawberry-harvesting
Juste, F., Sevila, F., Citrus: An European Project to Study the Robotic Harvesting of robot: Design, development, integration, and field evaluation. J Field Rob 37 (2),
Oranges, in Proceedings of the 3. International Symposium on: Fruit, Nut, and 202–224.
Vegetable Harvesting Mechanization Jordbrugsteknik Agricultural Engineering, Zion, B., Mann, M., Levin, D., Shilo, A., Rubinstein, D., Shmulevich, I., 2014. Harvest-
no.67, 1992, 331-338. order planning for a multiarm robotic harvester. Comput Electron Agric 103, 75–81.
Karkee, M., Zhang, Q., 2012. Mechanization and automation technologies in specialty
crop production. Resource Magazine. 19 (5), 16–17.

17