You are on page 1of 12

Computers and Electronics in Agriculture 204 (2023) 107552

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture


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

Original papers

Development and evaluation of a robust soft robotic gripper for apple


harvesting
Xing Wang a,b ,1 , Hanwen Kang a ,1 , Hongyu Zhou a , Wesley Au a , Michael Yu Wang a , Chao Chen a ,∗
a
Department of Mechanical and Aerospace Engineering, Monash University, Clayton, Australia
b
Robotics and Autonomous Systems Group, Data61, CSIRO, Brisbane, Australia

ARTICLE INFO ABSTRACT

Keywords: Fruit harvesting is facing challenges due to the labour shortage, which has been more severe since the
Soft robotics rapid pandemic. Robotic harvesting has been attempted in autonomous fruit harvesting tasks, such as apple
Soft gripper harvesting. However, current apple harvesting robots show limited harvesting performance in the orchard
Fruit harvesting
environment due to the inefficiency of the robotic grippers. This research presents a fruit harvesting method
High force
that includes a novel soft robotic gripper and a detachment strategy to achieve apple harvesting in the natural
Apple grasping
orchard. The soft robotic gripper includes four tapered soft robotic fingers (SRF) and one multi-mode suction
cup. The SRF is customised to avoid interference with obstacles during grasping, and its compliance and
force exertion are comprehensively evaluated with FEA and experiments. The multi-mode suction cup can
provide suction adhesion force, show active extrusion/withdrawal, and present passive compliance mode.
The simultaneously twist-pulling motion is finally proposed and implemented to detach the apples from the
trees. The proposed robotic gripper is compact, compliant with apple grasping and generates a large grasping
force. Our proposed method is finally validated in a natural orchard and achieves a detachment, damage and
harvesting rate of 75.6%, 4.55%, and 70.77%, respectively.

1. Introduction robot, whose end-effector consists of a clasp mechanism made of food-


grade silicone cast on the 3D printed rigid pad (Williams et al., 2019).
Fruit harvesting is facing increasing challenges due to its labour- The comprehensive in-field trial suggested that a successful harvesting
intensive nature and the workforce’s diminishing work passion in agri- rate of 51% and an average cycle tie of 5.5 s per fruit are achieved. The
culture (Zhou et al., 2022; Zheng et al., 2021; Ma et al., 2014). It has low success rate mainly dues to the accident drop-off and knock-off by
become even more severe since the rapid pandemic caused by COVID- the gripper during the harvesting process, as the kiwi fruit are highly
19 (Tougeron and Hance, 2021; Larue, 2020). Robotic harvesting can clustered.
address the labour shortage issue and prevent fruit waste from being Autonomous apple harvesting also attracts increasing attention in
unharvested or rotting (Junge and Hughes, 2022). industry and research fields (Hu et al., 2022; Ji et al., 2022b, 2021;
As a research hot spot, some pioneering research has recently been Wang et al., 2022; Kang et al., 2022a). Industry trials prefer non-
conducted on autonomous and intelligent fruit harvesting robots (Tang fingered end-effectors, such as the suction tube mechanism. Abundant
et al., 2020; Bogue, 2020; Hua et al., 2019; Ji et al., 2022a). Some robotics and Ripe robotics utilised a large fluid pipe, where the vacuum
robots mainly aim for selective harvesting, which uses end-effectors
can be generated to suck the apples into the pipe (Zhang et al., 2020a).
associated with the robotic arm to detach the fruits selectively (Kurtser
There is no official data available to evaluate the overall harvesting
and Edan, 2020; Liu et al., 2020; Gao et al., 2022; Zhang et al., 2020b;
performance. However, the suction pipe may suffer from frequent jams
Kang et al., 2020). Simon et al. designed a lettuce harvesting robot with
due to the leaves and multiple occluded fruits. Robotic grippers as
a pneumatic actuated knife to cut the lettuce off for deposit (Birrell
an effective solution for selective harvesting were also investigated to
et al., 2020). Although a high harvest success rate of 88.2% was
grasp and detach the fruits from the trees. Soft robotic fingers, the
achieved, the damage rate was poor at 38% due to the rigid mechanical
essential component of the robotic gripper, are typically made from soft
cutting mechanism, which prevents its broader application in precision
harvesting. Henry et al. developed an autonomous kiwi fruit harvesting hyperelastic materials with a Young modulus of less than 1 GPa (Rus

∗ Corresponding author.
E-mail address: Chao.Chen@monash.edu (C. Chen).
1
These authors contributed equally.

https://doi.org/10.1016/j.compag.2022.107552
Received 1 August 2022; Received in revised form 10 November 2022; Accepted 6 December 2022
Available online 14 December 2022
0168-1699/© 2022 Elsevier B.V. All rights reserved.
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

and Tolley, 2015). They can generate compliant, flexible, and highly Orchard, located in Drouin, Australia, and the system is shown in
adaptive bending motion from the pre-defined geometry (Peng et al., Fig. 1(b).
2018; Elfferich et al., 2022). Some research has been conducted to It can be seen that there is occlusion on the apples caused by
apply soft robotics to agriculture applications due to these advantages leaves, branches, or clustered apples, which should be considered in
over the traditional rigid grippers (Navas et al., 2021; Hohimer et al., the robotic gripper design to maximise the performance of the robotic
2019; Zhao et al., 2011). Davidson et al. proposed an apple harvesting grasping. Leaves can potentially cause slipperiness between the finger
robot, which utilised a cable-driven three-finger robotic gripper to and the contact surface of the apple, so it is crucial to secure the
detach the apples from the trees. An overall success rate of 84% with grasping pattern and lock the apple in place once it is grasped. This
a cycle time of 6 s per apple was reported (Davidson et al., 2016). can be tackled by designing grippers with multi-fingers. The contact
Even though the soft pads are applied at the tip of the robotic finger, surface is maximised, preventing the potential slippery caused by leaf
the compliance of the robotic gripper is greatly limited by the rigid interference while grasping. In addition, the branches can be thick and
linkages. In addition, the trial orchard was pre-processed by thinning stiff, which needs to be avoided during grasping. This requires the
and reducing the complexity of the environment. The robotic grasp- gripper to limit the contact area with the environment but keep the
ing, detachment, and overall harvesting performance were expected large contact surface with fruits. Occluded apples also pose challenges
to drop in the more complex and clustered environment. Hohimer to conformal grasping. Solutions reducing the interference from adja-
et al. utilised a 3D printed soft end-effector to harvest the ‘‘Envy’’ cent apples should also be explored to minimise the potential failure
apple (Hohimer et al., 2019). The soft end-effector provides some grasping rate and risk of knocking down these apples. The design will
improvements over previous grasping end-effectors. The compliant soft investigate the tapered soft robotic finger with a gradually decreasing
fingers were unharmed by collisions with the tree canopy and trellis cross-section area to minimise the potential contact at the gripper tip.
wires and improved significantly from their last generation of tendon- The diameters of the apple mainly fall into a range of 67–82 mm,
driven fingers. Their gripper can provide 49.54 N pulling force under with an average of 75.6 mm. The average apple weight comes to 162 g.
103.4 kPa input pressure. However, the gripper design is bulky in These data are calculated from the harvested apples. The number of
size to maximise the grasping force. It relies on a symmetrical grasp clustered apples is also identified as an estimation of the grasping diffi-
with simultaneous activation of fingers (no suction) to grasp and direct culty index, as the clustered apple is more difficult to achieve conformal
pulling to detach, which causes significant damage to adjacent fruit grasping than the isolated and hanging ones. All the information needs
while detaching in an orchard environment. to be considered in the design of soft pneumatic grippers to achieve a
Despite the attempts and advances in industry and research to higher detachment rate.
develop autonomous apple harvesting robots, the current robotic grip-
per is still not robust or reliable enough to deal with the complex 2.2. Gripper development and modelling
orchard environment, and their overall system performances are still
unsatisfactory from being fully deployed in the natural orchard. This 2.2.1. Design consideration
is because the current under-developed gripper designs, such as the From the orchard environment and fruit data, we propose a four-
bulky, rigid gripper, unreliable soft end-effectors etc., cannot meet finger soft pneumatic gripper with an additional suction component, as
the requirements of being compliant, compact in size, forceful, and shown in Fig. 2. The four-finger design is a compromised solution for
dexterous simultaneously. These disadvantages make it hard to harvest performance and operation difficulty. It can guarantee the contact area
the apples in the natural orchard. with the target fruit to prevent the potential slippery while retracting
This study proposes a novel dexterous soft robotic gripper with and does not over-complicate the grasping subsystem of the robot.
tapered soft robotic fingers and an adaptive suction cup. The tapered To reduce the interference with branches and adjacent apples during
robotic finger shows excellent conformability and large force exertion, grasping, SRF is designed to have a tapered shape to reduce the body
while the suction cup assists the grasping with active suction and size while keeping the forceful output. The feasibility of the tapered
passive compliance. This works also enables the suction cup with multi shape and structure parameters are evaluated through finite element
modes: active extrusion/retraction and passive adaption. This work modelling. The detailed study is presented in the next section.
then introduces a detachment strategy to simultaneously combine the Additionally, a suction cup with an active extension/retraction func-
twist and pulling motion after grasping the target. Unlike the direct tion and novel passive compliance is proposed and integrated with the
twist or pulling that cause knocking down adjacent apples or shaking soft gripper. The active motion is achieved by the extrusion/withdrawal
of mature apples respectively, the proposed detachment strategy is an of the pneumatic cylinder, as shown in Fig. 2(c). The stoke of the
efficient way to reduce fruit waste during harvesting. Lastly, this work suction cup is 0.091 m, and the dimension of the suction cup is also
integrates the proposed soft robotic gripper with the system to conduct shown in Fig. 2(d). It has a length of 0.036 m. It can suck objects under
in-field experiments to validate their harvesting performance. negative pressure. The passive mode is realised by setting both ends of
The rest of the paper is organised as follows. The materials and the pneumatic cylinder to atmosphere pressure. In this case, the suction
methodologies of our approach are presented in Section 2. The experi- cup is free to move together with the shaft of the pneumatic cylinder.
ment results and discussion are presented in Section 3, followed by the The rest of the parts are designed and manufactured from 3D printing
conclusion in Section 4. with a consumer-level printer, Prusa MK3S.

2. Materials and methodology 2.2.2. Soft robotic finger


The SRF is designed with a hollow fold-based structure, extending
2.1. Orchard environment and design consideration along its longitudinal direction once under pressurisation. The addi-
tional strain limiting layer is added to constrain the extension at the
Different fruit tree arrangements, including traditional 3D complex bottom side (see Fig. 3). In this case, the elongation at the top will
trees and 2D planar tree structures, are managed to keep different apple be larger than the bottom side upon inflation, generating the bending
varieties. The harvesting mainly focuses on the 2D planar structures, motion for each SRF. It is then 3D-printed with hyperelastic material
with the apple variety being ‘‘pink lady’’. These are wire trellis in the named Ninjaflex. This tapered SRF is an updated version of the bistable
orchard. The apples are grown and arranged in a planar manner, which soft pneumatic fingers from our previous research work (Wang et al.,
is ideal for autonomous robotic harvesting. The apples are reasonably 2020, 2021). The past work utilised a finger design with constant height
dense and occluded for our test spots, as shown in Fig. 1(a). The to generate a constant curvature bending profile. The energy storage
autonomous apple harvesting tests are conducted in Fankhauser Apple mechanism from our previous design enables the soft finger with

2
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 1. (a) 2D apple orchard, and (b) Monash Apple harvesting system.

Fig. 2. (a) Simplified model for the overall system, (b) Detailed view of the proposed soft robotic gripper, (c) the stroke of suction cup (two SRFs are visualised in this plot), (d)
dimensions of the suction cup.

Fig. 3. (a) 3D view of the tapered soft robotic finger, (b) The detailed cross-section view of the SRF.

3
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 4. (a) Dog-bone test sample, (b) Instron Universal Tester, (c) The average stress–strain curve and curve fitting with Mooney–Rivlin model.

the binary bending states. Those grippers pose two bending profiles: Table 1
Mooney Rivlin model for NinjaFlex.
open and closing, corresponding to two local minimum energy states
Hyperelastic model Material constant Value
during the operation. They can rapidly close within less than 0.2s by
releasing the pre-stored strain energy. The disadvantage is that this C01 2.90 MPa

high-speed closing only stops when blocked by external objects. This C10 −0.045 MPa
process cannot be controlled with the current control method, making Mooney Rivlin C11 −0.017 MPa
apple grasping difficult. This can cause fruit damage like bruises, which C20 0.0023 MPa
needs to be avoided for future sales in the market. In addition, their C02 0.18 MPa
constant height, especially the large tip size, generates a constant incompressible Parameter D 0 MPa−1
bending curvature increasing the chance of interference with adjacent
obstacles. A tapered shape, in this case, can better extend to the deep
canopy for fruit grasping. This design with a tapered shape on similar
2.2.4. Finite element modelling
soft robotic fingers also suggests that the tapered design can benefit
grasping tasks (Dilibal et al., 2018). The structure optimisation of FEA is utilised as a pilot study to study the bending behaviour,
the soft gripper can also be implemented because soft grippers from force exertion, and contact with apples. The simulation is conducted in
human design, bio-inspired design, or even some local computation Abaqus/CAE with the material properties found in the previous section.
design algorithms are likely not to be globally optimised in terms of The SRF, block, and apple are imported into the Abaqus in step format.
performance (Xavier et al., 2021; Pinskier and Howard, 2022; Chen The block is to simulate the load cell. Then the material properties
and Wang, 2020). In this case, further structure/parameter optimisation are assigned for each object. The SRF, block and apple are meshed
research can be conducted to find an optimal gripper design for apple with Tetrahedral quadratic hybrid elements (element type C3D10H)
harvesting. However, it is not within the scope of this research stage. for faster convergence. The SRF instance is first created, whose inner
The interactions between a robot gripper and the deformable fruit cavity is subjected to various static input pressures and the proximal
are highly complex and thus challenging to model. So this work ex- end is enabled with the ENCASTRE boundary condition. The instance,
plored the bending profiles, force output, and physical contact with including SRF and block, is then created to find the blocked tip force
fruit by finite element modelling simulation and tested from experi- of the SRF. The last instance includes SRF, and the apple is created to
ments. The material property of Ninjaflex is found from a tensile test, find the finger’s conformability when grasping the apple.
while that of apple fruit is from Fan et al. (2021).

2.3. Detachment strategy and control circuit


2.2.3. Tensile test and material characteristic
The material property of Ninjaflex is evaluated from the tensile
Once the target fruit is grasped or secured within the proposed
test following ISO 37 standards. The dog-bone samples in Fig. 4(a)
gripper, the following detachment motion becomes the dominant role
are printed longitudinally and stretched by Instron Universal Tester
E3000 with a speed of 200 mm/min. The average stress–strain data in removing the fruit from the tree. Selecting the proper strategy is
fit into Mooney–Rivlin modes in Abaqus, as shown in Fig. 4(c). The one of the critical factors in improving the operation efficiency, and
Mooney–Rivlin model is below with the material constants shown in harvesting rate (Fan et al., 2021). We investigate the most efficient de-
Table 1 (Kim et al., 2012; Kumar and Rao, 2016). tachment strategies from the literature, including direct twist, pulling,
lifting, and bending (Navas et al., 2021). Another recent research on
𝐸𝑠𝑡𝑟 = 𝐶10 (𝐼1 −3)+𝐶01 (𝐼2 −3)+𝐶11 (𝐼1 −3)(𝐼2 −3)+𝐶20 (𝐼1 −3)2 +𝐶02 (𝐼2 −3)2 human three-finger grasp experiments concludes that the best pattern
was determined to be rotation–pull. This result is from evaluating the
(1)
separation force during a dynamic simulation (Fan et al., 2021). Our
𝐸𝑠𝑡𝑟 is the overall stored strain energy per unit volume. previous in-field harvesting experiments also suggest that combining
The results from the tensile test are utilised in the finite element twist and pulling motion is efficient (Wang et al., 2022). This study
modelling section below. Note the polynomial is selected in Abaqus involves adjusting and synchronising the twist and pulling motion for
settings. detachment efficiency.

4
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 5. The overall control circuit for the system.

2.3.1. Pneumatic control circuit and suction component set to the predefined home positions for the following motions. Visual
This section details the overall control circuit of the Monash Apple perception performs fruit detection using the RGB image received from
Retrieving System (MARS), as shown in Fig. 5. MARS consists of 4 com- the Realsense camera. Then LiDAR-RGB fusion algorithm matches the
ponents: mobile base, LiDAR-RGB vision, UR5 manipulator, and soft corresponding targets from RGB to LiDAR frames and records their 3D
robotic gripper. A central laptop is the master of communication, con- locations. A picking sequence is defined based on the depth distance
nected to LiDAR-RGB (Livox-Realsense), a Robotic manipulator (Uni- of each detected apple. After that, the final path is generated from
versal Robot 5), and an Arduino controller via Ethernet and USB the IK solver and path selection. UR5 manipulator moves to a closer
Ethernet and USB cables, respectively. Only LiDAR is directly connected position to the first apple. Next, grasping is enabled starting from the
to a 24 V power supply, while the others are connected to an inverter suction force generation of the suction cup. It is then pushed out to
that generates 220 V AC from the 24 V battery. The communication contact the target apple by a pneumatic cylinder. Once it sucks the
between each component is conducted through ROS Melodic, while the apple, the suction cup retracts together with the apple. The fingers close
laptop serves as the ROS master and others as the ROS nodes. Fig. 6 immediately to secure the apple within the closure. It must be noted
shows the pneumatic control circuit. The Arduino controller controls that the suction cup only provides initial contact and suction to allow
the pneumatic components. It sends signals to relays and activates the soft robotic finger to conform to the apple promptly and thoroughly.
certain solenoid valves. There are six solenoid valves, including one Purely relying on a suction cup is not efficient due to the complex
five-port-three-way solenoid valve and five five-port-two-way solenoid orchard environment. Then wrist three of the UR5 arm twists 270
valves. degrees, and the arm pulls out of the canopy for 20 cm simultaneously
The pneumatic control cycle is to control certain solenoid valves to detach the fruit from the tree. The manipulator moves the apple
and the downstream pneumatic components with compressed air. Once to the deposit position and opens the gripper to collect the apple.
the soft robotic finger opens, the five-port-two-way solenoid valve that The grasping-to-deposit processes finish once the last planned apple is
controls the suction cup is firstly activated to provide suction force for harvested (see Fig. 8).
adhesion, as shown in Fig. 7(a). The five-port-three-way valve is then
activated to drive the pneumatic cylinder to extrude the suction cup. 2.4.2. Performance evaluation
Thirdly, the pneumatic cylinder sucks the apple and instantly switches The performance of the soft robotic gripper is evaluated through the
to passive and compliant mode, as shown in Fig. 7(b). Next, the suction individual finger test, the grasping force test of the four-finger gripper,
cup withdraws, and four solenoid valves that control the soft pneumatic as well as the performance evaluation of the integrated system. The
finger will be activated to close the finger, as shown in Fig. 7(c) and individual finger tests involve the test of bending profile and tip force.
(d). The object is secured inside the gripper. Further twist and pulling The former indicates different bending states under various inflation
motion is enabled by the UR5 manipulator, as shown in Fig. 7(e). Once pressures, while the latter means different force exertion under various
the apple is picked from the tree, the end-effector and the picked apple input pressures. The grasping force test evaluates the force required to
are manipulated by the robotic arm to reach the predefined deposit pull the object out of the closed gripper. This estimates the maximum
position. This deposit position sits right at one end of the sliding tubes. effective grasping force the gripper can provide during the detachment
process.
2.4. Overall system setup and test The performance evaluation of the integrated harvesting system
includes the following indexes:
2.4.1. Hardware setup and workflow
The soft robotic gripper is integrated with the MARS to perform au- • Detection rate: the number of detected apples with 3D location
tonomous apple harvesting in the Fankhauser apple orchard in Drouin, per total number of apples within the RGB image. The apple
Australia. Aside from the four main components, a trailer that drives picking is conducted for mature apples as directed by the farm
the pump to provide the pressure source and an apple deposit subsys- owner, so we do not apply maturity classification in this section.
tem with a plastic basket are also included. The integrated system is • Path planning success rate: the number of successfully planned
as shown in Fig. 1(b). The overall flow of the harvesting robot can paths to each detection apple per total required paths.
be summarised as follows. Once the harvesting starts, the mobile base • Detachment rate: the number of successfully detached apples per
navigates to a position at a certain distance from the fruit plane. This total apples planned to detach.
distance guarantees that the target apples are within the workspace • Damage rate: the number of damaged apples per total detached
of the UR5 manipulator. Next, the robotic arm and soft gripper are apples.

5
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 6. The overall pneumatic control circuit for the gripper.

Fig. 7. (a) Gripper open and suction force enabled to approach target apple, (b) Suction cup extrudes and adheres target, (c) Suction cup withdraws, (d) Finger closes instantly,
(e) Robotic arm twists and pulls the apple simultaneously to detach the target. Note one finger is hidden from (b) for better visualisation.

3. Results and discussion The finite element analysis is initially performed to evaluate the
design feasibility of the proposed robotic finger design, which involves
the study of design parameters and tapered chamber structure. After
3.1. Bending curvature of SRF
this pilot study, the physical robotic finger is printed and evaluated. The
finger is fixed at the base and inflated from 0 kPa to 80 kPa, as shown in
The bending curvature is utilised to evaluate the bending state of the Fig. 9. During the inflation, a step of 20 kPa increment is implemented,
robotic fingers under specific input pressures. This can help understand while the deformation of the fingers is recorded and saved for further
the bending state when the SRF is inflated in the free space. processing. The performance under 10 kPa instead of 20 kPa is recorded

6
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 8. The workflow for the apple harvesting, including system initialisation, visual perception, path planning, grasping, detachment, and final deposit.

Fig. 9. (a) Bending motion under 10 kPa, (b) Bending motion under 40 kPa, (c) Bending motion under 60 kPa, (d) Bending motion under 80 kPa.

at the beginning due to its rapidly changing deformation. The deformed reality, as shown in Fig. 11. Once the robotic finger is pressurised, the
bending profiles provide an initial estimation of physical bending. tip bends towards the hypothetical or actual load cell and generates the
Afterwards, a physical experiment is performed to provide an accurate tip force. The force is recorded 3 times with each soft robotic finger
bending profile of the soft robotic finger under various inflation inputs. inflated by the pressure ranging from 0 kPa to 140 kPa. A pressure
The same setup in the simulation is arranged with a high-resolution step of 20 kPa is implemented as well. Additionally, the distance 𝑑
camera to record the bending curvature corresponding to each input. between the SLL and the load cell is set as 0 mm, 10 mm, and 20 mm,
The FEA and experimental results are shown in Fig. 9. respectively, to evaluate the force exertion.
It can be seen that most of the bending occurs at the root of the Detailed experimental results with tip force measured at different
SRF due to a larger contact area with compressed air. The bending distances of 20 mm, 10 mm, and 0 mm are shown in Figs. 12 (a)–
propagates along the longitude direction with the increase of input (c). The tip force increases with the increases of inflation pressures
pressure. The maximum inflation pressure is set to 80 kPa due to the under all three configurations. The tip force decreases when the dis-
physical constraint in the experiment setup. The SRF bends to a near- tance 𝑑 increases because the SRF needs to deform to generate a tip
circle shape and is sufficient to wrap around the contour of the apple displacement before exerting the force. The maximum tip force is under
surface under 80 kPa. 140 kPa inflation pressure when the distance equals 0 mm. A tip force
Both FEA and experiment are further conducted to test its conforma- of 4.68 N is generated on the load cell. In addition, it can be seen
bility and flexibility. The robotic finger is inflated to contact the apple that the simulated and experimental results match each other well.
in both the simulation and experimental setup. Only one finger is set Both results indicate that the inflation pressure and initial configuration
up in FEA to save computation time. The contact shape verifies the highly affect the tip force.
conformability of the robotic finger. The deeper colour at the tip of the
simulated SRF indicates a compression force that tries to compress the 3.3. Grasping force evaluation of gripper
apple and secure its position inside the gripper.
The tapered SRF shows its superior performance in conformability
3.2. Tip force evaluation of SRF and force exertion and is utilised to create the four-fingered soft robotic
gripper. This section further evaluates the grasping performance of the
The tip forces exerted by the robotic fingers under various inputs are assembled soft robotic gripper. It is orientated with the suction cup
simulated and experimentally measured. To achieve this, the fingers are facing downwards to grasp a sphere with a diameter of 80 mm, as
fixed at the base, with their tip end free to move under inflation. The shown in Fig. 13(a). The apple is attached and fixed to a load cell.
finger guarantees initial face-to-face contact between the SLL and the The envelope grasp first secures the apple after the gripper is inflated.
modelled hypothetical load cell in Abaqus or the six-axis load cell in The critical force pulling the apple out of the gripper is recorded as

7
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 10. (a) The simulated finger bending to conform to the apple model, (b) The experimental results for apple grasping.

Fig. 11. (a) The setup for tip force test, with load cell, (b) Different distances between the tip of the SRF and the load cell are set up.

Fig. 12. Tip force for SRF at a distance of (a) 20 mm, (b) 10 mm, (c) 0 mm.

8
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 13. (a) Experiment setup for pull-out force, (b) Results for the pull-out force to pull spheres with different diameters and under different pressures.

an evaluation index, called pull-out force. This force is tested under the target apples are received from the point clouds and mapped to
various pressure inputs ranging from 20 kPa to 140 kPa and on spheres the world frame. The detection of apples within the scene is saved as
with diameters of 70 mm, 80 mm, and 90 mm. The results are shown masked images to evaluate the accuracy of visual detection. There are
in Fig. 13(b). 71 apples captured in the common FoV of the RGB camera and Lidar
The experiment results suggest that a higher pull-out force is re- sensor. Among all those, 65 apples are labelled manually to evaluate the
quired for higher inflation pressure. The SRF can exert a higher force on performance of visual detection. This is because some of the occluded
the apple’s surface to secure it. The highest force is required for a sphere or boundary apples are not feasible to grasp. The average number of
with an 80 mm diameter. Because the soft robotic finger is too long or apples within one scan is 36. Our one-stage visual network work can
too short for smaller or larger spheres. When the sphere is smaller, the detect 92.3% of all those 65 pickable apples. The average inference
inner surface of the SRF cannot fully conform to the apple’s surface. The time for 2D segmentation is 53 ms per image. The total inference time is
finger intends to bulge with the tip contacting the apple only, especially 213 ms to receive the 3D location of all detected apples from colourised
under high pressure. On the contrary, the gripper cannot fully contain point clouds.
the larger sphere within its grasping range, resulting in a lower pull-
out force. What is essential, this pull-out force is high because the apple 3.4.2. Fruit harvesting process
tries to unbend the robotic finger. A higher force is needed to unbend The 3D locations of the apples are utilised to plan the approaching
the finger to a straight configuration and compensate for the friction path for the robotic arm, and the planning success rate is also recorded
force during pulling. The maximum pull-out force at 140 kPa is 65.76 as an evaluation index. Robot Operating System’s MoveIt motion plan-
N. The gripper can exert high detach force due to an average apple ning method is utilised to implement constrained task space planning
diameter of 75.6 mm with a flexible and conformable grasp. and configuration space trajectories. The motion planning algorithm
In addition, the suction cup is enabled to evaluate the suction RRT connect and PRM∗ are combined for configuration space planning.
force. The vacuum generator component provides the suction force, After the path plans, the robotic arm is manipulated to approach the
generating up to −80 kPa vacuum pressure. The maximum suction force first target apple with the gripper open and suction force enabled, as
goes up to 12 N to adhere to apples with an average size of 75.6 mm. shown in Fig. 15(b). The gripper closes immediately once the suction
This suction force is not included in evaluating the overall pull-out cup withdraws with the apple. Fig. 15(d) demonstrates the twist and
force because the vacuum suction is disabled once the gripper closes pulling off of the overall gripper to detach the apple from the apple tree.
and secures the apple. This early disablement of suction force is due to The robotic arm then moves to the deposit position to drop the apple
the physical limitation as keeping the suction force consumes a massive into the collection bin, as shown in Figs. 15 (e) and (f). The average
amount of airflow. planning time is 2.61 s while executing the manipulation, grasping,
moving to deposit, and releasing the gripper take an average time of
3.4. Overall performance in apple orchard 10.87 s. This can be broken down to 3.13 s, 2.08 s, 3.65 s, and 2.01 s for
each process. The suction cup extension and withdrawal take about 0.2
3.4.1. System initialisation and visual sensing s individually, and the SRF takes about 2 s to open due to the limitation
The system is initialised and stays at the home position before of our current vacuum generator. All the apples removed from the tree
operation. Then the mobile base is navigated to the predefined position versus all the planned apples are recorded as the detachment index,
in the orchard and gets ready for harvesting. ROS nodes are published, while the damage to apples, including bruises and stalk damage, are
including vision, arm manipulation, and gripper control nodes. also recorded.
The vision node is called to perform the fruit detection, and in-
stance segmentation on the orchard (Kang et al., 2022b), as shown 3.4.3. Experimental results
in Fig. 14(a). The calibrated LiDRA-RGB sensor set also captures the In total, there are four spots where 65 apples were detected as the
colourised apple scene, as shown in Fig. 14(b). The 3D locations of test samples from the recorded orchard images, as shown in Table 2 .

9
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Fig. 14. (a) Visual perception to detect and segment the apples (sample image), (b) LiDAR-RGB fusion to receive 3D locations of apples.

Fig. 15. The robotic harvesting process, including (a) system initialisation at home position, (b) gripper opens with suction enabled to approach the apple, (c) suction cup withdraws
and gripper closes to secure the apple, (d) Arm twists and pulls to detach the apple, (e) Arm move to deposit position, (d) Gripper releases to deposit the harvested apple.

The visual perception successfully detects and returns the 3D location 3.5. Discussions
of 60 apples. Some of the apples are highly occluded by the leaves or
branches, which are not classified as the picking target and are not This study proposes a soft robotic gripper for apple harvesting in the
considered in calculating vision detection accuracy. The path planning natural orchard. The gripper consists of four soft robotic fingers and one
algorithm generates and selects a valid approaching path for 60 target suction cup with active extrusion/withdrawn and passive adaptivity. By
3D locations. The arm knocks down two apples during the planning. comparing the results of experimental and finite element modelling, the
As a result, 58 paths are successfully executed to detach the apples. proposed taper SRF exhibits excellent compliance and forceful output
In total, 44 apples successfully detached from the apple trees, which from the tests of bending and force, respectively. The adaptivity when
is about 75.86% of the planned target apples. The detachment failure grasping the apple model also demonstrates its potential for compliant
mainly comes from the clustered apple, thick branch interference and grasping. The tests on both tip force and pull-out force demonstrate an
relative-lower torsion stiffness of the SRF. Forty-six apples were de- extensive range of force exertion under various input pressures. The
tached from the tree, two of which were damaged during harvesting. suction cup also benefits the initial adhesion of apples, which makes
Those two apples are considered damaged due to the bruise and stalk the following finger close and apple securing more efficient.
damage. They are also considered in the calculation of the harvesting A downside of the pneumatic-actuated fingers is the potential failure
rate. An average cycle time of 14.69 s is required to harvest one apple. under long-time operation. As a 3D-printed SRF, it can last more than

10
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Table 2
Overall performance index for the proposed soft robotic gripper with the system.
Performance index Value Percentage Average time /s Note
Total number 71 100.00% – Apples within recorded data
Pickable 65 91.54% – Pickable apples as labelled manually
Detection 60 92.3% 0.213 Detection results from vision
Path planning 58 96.67% 9.39a Two apples knocked down during planning
Detachment 44 75.86% 2.08 All the apples removed are counted
Damage 2 4.55% – Including bruise and stalk damage
Harvesting 46 70.77% 14.69 Including all detached apples
a
Including the time for planning (2.61 s), executing (3.13 s), and deposit (3.65 s).

five thousand cycles. This number can increase with the decrease of CRediT authorship contribution statement
normal operating pressure. However, this life cycle is still an issue for
autonomous fruit harvesting and needs to be addressed for the proposed Xing Wang: Conceptualization, Methodology, Software, Simula-
soft robotic gripper. tion, Validation, Writing – original draft. Hanwen Kang: Conceptu-
The bending curvature of the SRF can be optimised to maximise the alization, Methodology, Data curation, Validation, Writing – original
contact area. From Fig. 10, the tip of the SRF conforms to the apple draft. Hongyu Zhou: CAD design, Validation. Wesley Au: Conceptual-
ization, Validation. Michael Yu Wang: Writing – review & editing, Su-
well, but the contact area decreases towards the root direction. This
pervision. Chao Chen: Conceptualization, Writing – review & editing,
is due to the uniform design parameters and material selection of the
Supervision, Project administration.
strain limiting layer (SLL). Inverse design and multilateral fabrication
of SLL based on the contour of the apple can be an efficient way to
Declaration of competing interest
maximise the contact area, grasping stability and detachment rate.
Even though the suction cup show advantages of generating active The authors declare that they have no known competing finan-
motion and passive compliance, the suction force cannot be applied cial interests or personal relationships that could have appeared to
to continuously adhere apples due to a massive amount of air con- influence the work reported in this paper.
sumption. Adding an extra vacuum pressure source can be utilised as
a compromised solution to increase the complexity of the pneumatic Data availability
control circuit and then increase the suction force’s working time. The
increased vacuum pressure value and operating time can potentially Data will be made available on request.
increase the detachment and overall harvesting rates. Optimising the
pneumatic circuit is also expected to reduce the response time of the Acknowledgements
soft gripper.
The accidentally knocked-off rate needs further reduced to prevent We gratefully acknowledge the financial support from the Aus-
undesired fruit waste. Potential optimisation on the finger design may tralian Research Council (ARC ITRH IH150100006). Additionally, Dr
involve varying taper angles, fold structure and material selection. Xing Wang would like to thank the funding support from CSIRO’s CERC
Enabling tactile sensing capacity for the robotic fingers can also provide program. We also thank Mr Ruijie Zhang at Monash University for his
help with the experiment.
feedback and assist with the current open-loop twist-pulling motion to
prevent potential knocking-down.
References

Birrell, S., Hughes, J., Cai, J.Y., Iida, F., 2020. A field-tested robotic harvesting system
4. Conclusion
for iceberg lettuce. J. Field Robotics 37 (2), 225–245.
Bogue, R., 2020. Fruit picking robots: has their time come? Ind. Robot Int. J. Robot.
Res. Appl.
This work proposes a novel soft robotic gripper with tapered soft Chen, F., Wang, M.Y., 2020. Design optimization of soft robots: A review of the state
robotic fingers and a suction cup with active motion and passive of the art. IEEE Robot. Autom. Mag. 27 (4), 27–43.
compliance. The tapered SRF is a customised soft bending actuator from Davidson, J.R., Silwal, A., Hohimer, C.J., Karkee, M., Mo, C., Zhang, Q., 2016. Proof-of-
concept of a robotic apple harvester. In: 2016 IEEE/RSJ International Conference
single-step 3D printing. It shows excellent compliance and adaptivity in
on Intelligent Robots and Systems. IROS, IEEE, pp. 634–639.
apple fruits’ free bending and grasping. Further tip and pull-out force Dilibal, S., Sahin, H., Celik, Y., 2018. Experimental and numerical analysis on the
tests indicate promising force exertion. A maximum pull-out force of bending response of the geometrically gradient soft robotics actuator. Arch. Mech.
65.76 N can be generated on an 80 mm sphere under 140 kPa inflation 70 (5), 391–404.
Elfferich, J.F., Dodou, D., Della Santina, C., 2022. Soft robotic grippers for crop handling
pressure. To assist the grasping sequence, the suction cup can also show or harvesting: A review. IEEE Access 10, 75428–75443.
different modes, such as suction adhesion, active extrusion, passive Fan, P., Yan, B., Wang, M., Lei, X., Liu, Z., Yang, F., 2021. Three-finger grasp planning
compliance and active withdrawal. An effective detachment method and experimental analysis of picking patterns for robotic apple harvesting. Comput.
Electron. Agric. 188, 106353.
of simultaneously twist-pulling is implemented to remove the apples
Gao, J., Zhang, F., Zhang, J., Yuan, T., Yin, J., Guo, H., Yang, C., 2022. Development
from the trees. Comprehensive field tests have been conducted with and evaluation of a pneumatic finger-like end-effector for cherry tomato harvesting
the proposed soft robotic gripper and overall system to validate the robot in greenhouse. Comput. Electron. Agric. 197, 106879.
efficiency. The experiment results suggest a 75.86% detachment rate Hohimer, C.J., Wang, H., Bhusal, S., Miller, J., Mo, C., Karkee, M., 2019. Design and
field evaluation of a robotic apple harvesting system with a 3D-printed soft-robotic
and around 4.55% damage rate are achieved. Further improvements are end-effector. Trans. ASABE 62 (2), 405–414.
planned to advance the design, material, and structure of the proposed Hu, G., Chen, C., Chen, J., Sun, L., Sugirbay, A., Chen, Y., Jin, H., Zhang, S., Bu, L.,
soft robotic fingers, together with the control of the pneumatic system. 2022. Simplified 4-DOF manipulator for rapid robotic apple harvesting. Comput.
Electron. Agric. 199, 107177.
In addition, more experiments should also be conducted to provide a
Hua, Y., Zhang, N., Yuan, X., Quan, L., Yang, J., Nagasaka, K., Zhou, X.-G., 2019.
more comprehensive evaluation of the harvesting performance of the Recent advances in intelligent automated fruit harvesting robots. Open Agric. J.
overall system. 13 (1).

11
X. Wang et al. Computers and Electronics in Agriculture 204 (2023) 107552

Ji, W., Pan, Y., Xu, B., Wang, J., 2022a. A real-time apple targets detection method Pinskier, J., Howard, D., 2022. From bioinspiration to computer generation:
for picking robot based on ShufflenetV2-YOLOX. Agriculture 12 (6), 856. Developments in autonomous soft robot design. Adv. Intell. Syst. 4 (1), 2100086.
Ji, W., Tang, C., Xu, B., He, G., 2022b. Contact force modeling and variable damping Rus, D., Tolley, M.T., 2015. Design, fabrication and control of soft robots. Nature 521
impedance control of apple harvesting robot. Comput. Electron. Agric. 198, 107026. (7553), 467–475.
Ji, W., Zhang, J., Xu, B., Tang, C., Zhao, D., 2021. Grasping mode analysis and adaptive Tang, Y., Chen, M., Wang, C., Luo, L., Li, J., Lian, G., Zou, X., 2020. Recognition and
impedance control for apple harvesting robotic grippers. Comput. Electron. Agric. localization methods for vision-based fruit picking robots: A review. Front. Plant
186, 106210. Sci. 11, 510.
Junge, K., Hughes, J., 2022. Soft sensorized physical twin for harvesting raspberries. Tougeron, K., Hance, T., 2021. Impact of the COVID-19 pandemic on apple orchards
In: 2022 IEEE 5th International Conference on Soft Robotics. RoboSoft, IEEE, pp. in Europe. Agric. Syst. 190, 103097.
601–606. Wang, X., Kang, H., Zhou, H., Au, W., Chen, C., 2022. Geometry-aware fruit grasping
Kang, M., Fan, Z., Yu, X., Wan, H., Chen, Q., Wang, P., Fu, L., 2022a. Division- estimation for robotic harvesting in apple orchards. Comput. Electron. Agric. 193,
merge based inverse kinematics for multi-DOFs humanoid robots in unstructured 106716.
environments. Comput. Electron. Agric. 198, 107090. Wang, X., Khara, A., Chen, C., 2020. A soft pneumatic bistable reinforced actuator
Kang, H., Wang, X., Chen, C., 2022b. Accurate fruit localisation using high resolution bioinspired by Venus Flytrap with enhanced grasping capability. Bioinspiration
LiDAR-camera fusion and instance segmentation. Comput. Electron. Agric. 203, Biomim. 15 (5), 056017.
107450. Wang, X., Zhou, H., Kang, H., Au, W., Chen, C., 2021. Bio-inspired soft bistable actuator
Kang, H., Zhou, H., Wang, X., Chen, C., 2020. Real-time fruit recognition and grasping with dual actuations. Smart Mater. Struct. 30 (12), 125001.
estimation for robotic apple harvesting. Sensors 20 (19), 5670. Williams, H.A., Jones, M.H., Nejati, M., Seabright, M.J., Bell, J., Penhall, N.D.,
Kim, B., Lee, S.B., Lee, J., Cho, S., Park, H., Yeom, S., Park, S.H., 2012. A com- Barnett, J.J., Duke, M.D., Scarfe, A.J., Ahn, H.S., et al., 2019. Robotic kiwifruit
parison among Neo-Hookean model, Mooney-Rivlin model, and Ogden model for harvesting using machine vision, convolutional neural networks, and robotic arms.
chloroprene rubber. Int. J. Precis. Eng. Manuf. 13 (5), 759–764. Biosyst. Eng. 181, 140–156.
Kumar, N., Rao, V.V., 2016. Hyperelastic Mooney-Rivlin model: determination and Xavier, M.S., Fleming, A.J., Yong, Y.K., 2021. Finite element modeling of soft fluidic
physical interpretation of material constants. Parameters 2 (10), 01. actuators: Overview and recent developments. Adv. Intell. Syst. 3 (2), 2000187.
Kurtser, P., Edan, Y., 2020. Planning the sequence of tasks for harvesting robots. Robot. Zhang, Z., Igathinathane, C., Li, J., Cen, H., Lu, Y., Flores, P., 2020a. Technology
Auton. Syst. 131, 103591. progress in mechanical harvest of fresh market apples. Comput. Electron. Agric.
Larue, B., 2020. Labor issues and COVID-19. Can. J. Agric. Econ./Revue Canad. 175, 105606.
D’Agroeconomie 68 (2), 231–237. Zhang, B., Xie, Y., Zhou, J., Wang, K., Zhang, Z., 2020b. State-of-the-art robotic
Liu, J., Peng, Y., Faheem, M., 2020. Experimental and theoretical analysis of fruit grippers, grasping and control strategies, as well as their applications in agricultural
plucking patterns for robotic tomato harvesting. Comput. Electron. Agric. 173, robots: A review. Comput. Electron. Agric. 177, 105694.
105330. Zhao, D., Lv, J., Ji, W., Zhang, Y., Chen, Y., 2011. Design and control of an apple
Ma, S., Karkee, M., Scharf, P.A., Zhang, Q., 2014. Sugarcane harvester technology: a harvesting robot. Biosyst. Eng. 110 (2), 112–122.
critical overview. Appl. Eng. Agric. 30 (5), 727–739. Zheng, W., Xie, Y., Zhang, B., Zhou, J., Zhang, J., 2021. Dexterous robotic grasping of
Navas, E., Fernández, R., Sepúlveda, D., Armada, M., Gonzalez-de Santos, P., 2021. Soft delicate fruits aided with a multi-sensory e-glove and manual grasping analysis for
grippers for automatic crop harvesting: A review. Sensors 21 (8), 2689. damage-free manipulation. Comput. Electron. Agric. 190, 106472.
Peng, Y., Liu, Y., Yang, Y., Yang, Y., Liu, N., Sun, Y., et al., 2018. Research progress Zhou, H., Wang, X., Au, W., Kang, H., Chen, C., 2022. Intelligent robots for fruit
on application of soft robotic gripper in fruit and vegetable picking. Trans. Chin. harvesting: Recent developments and future challenges. Precis. Agric. 1–52.
Soc. Agric. Eng. 34 (9), 11–20.

12

You might also like