Intelligent Soft Robotic Grippers for Agricultural and Food Product Handling: A Brief Review with a Focus on Design and Control

Advances in material sciences, control algorithms, and manufacturing techniques have facilitated rapid progress in soft grippers, propelling their adoption in various fields. In this review article, a comprehensive overview of the design and control aspects of intelligent soft robotic grippers tailored specifically for agricultural product handling is provided. Soft grippers have emerged as a promising solution for handling delicate and fragile objects. In this article, the recent progress in various gripper design, including fluidic and mechanical grippers, is elucidated and the role of advanced control approaches in enabling intelligent functions, such as object classification and grasping condition evaluation, is explored. Moreover, the challenges and opportunities pertaining to implementation of soft grippers in the agricultural industry are thoroughly discussed. While most demonstrations of soft grippers and their control strategies remain at the experimental stage, in this article, it is aimed to provide insights into the potential applications of soft grippers in agricultural product handling, thereby inspiring future research in this field.

conform to the shape of the object and distribute forces more evenly.Additionally, soft grippers offer versatility and can be easily adjusted to handle a wide range of products with minimal reconfiguration, making them ideal for managing seasonal crops and produces with varying sizes.Furthermore, the use of soft grippers can minimize produce damage, reduce waste, and consequently lead to increased yields and profits.Therefore, soft grippers hold the promise of revolutionizing the agriculture and food production industries by providing efficient, adaptable, and cost-effective solutions for crops and produce handling.
Drawing inspiration from natural creatures such as earthworms, [10] octopuses, [11] jellyfish, [12] elephant trunks, [13] and twining plants, [14] researchers have designed soft robotic actuators capable of fulfilling the grasping requirements for agricultural applications.For instance, multiple actuators can be combined into a gripper, as single soft actuators can lift weights successfully but may struggle to grasp objects [15] unless they possess the ability to twist and hold objects like a tentacle. [13,14]In contrast, envelope actuators [16][17][18] offer the advantage of a large contact area by enclosing objects within their body, eliminating the need for complex energy input strategies.[23] Similarly, hand structure actuators [24][25][26] demonstrate potential for improved grasping capabilities by emulating the dexterity of human hand.The attainment of complex grasping actions such as clam, pinch, and scoop may be accomplished by implementing bespoke actuators or interchanging between diverse configurations. [27]In comparison to rigid robotic grippers that primarily perform single clip and release motions, soft grippers have already achieved diverse motions such as extension, expansion, bending, and twisting. [28,29]This presents significant opportunities and high potential for harvesting crops of varying sizes, shapes, and stiffness.
In addition to structural considerations, soft actuator design has been enriched through the advancement of materials.32] Soft actuators can be further categorized based on their actuation method, encompassing fluidic-driven, mechanical-driven, electric-driven, thermal-driven, magnetic-driven, light-driven, and chemical-driven mechanisms (Figure 1).Among these, fluidic-driven mechanisms, particularly pneumatic-driven and hydraulic-driven systems, have gained widespread popularity owing to the accessibility to energy source and adaptability to various environments. [33]Actuators driven by pneumatical power have drawn particular interest in the field of soft robotics, as they offer lightweight, cost-effective, and easily fabricated solutions that deliver nonlinear motion with simple inputs. [22,25,34,35]1][42] Jamming, another versatile mechanism in soft robotics, enables new systems to be the development of systems with high stiffness variation with minimal volume variation, [43] which is mostly driven by vacuum as well.Mechanical-driven mechanisms, such as tendon transmissions, have been devised to transfer the pulling force directly to the actuator, thereby allowing the end-effector to hold targets while the controller and energy source are at a distance. [44,45]In addition, soft materials can undergo deformation through adhesion under relatively high voltages of approximately 3.5 kV, [46][47][48] allowing for the conversion of electrical energy into shape changes.
The functionalities of soft grippers can be further expanded by employing materials that respond to thermal, magnetic, light, and chemical stimuli, such as shape memory alloys, [49] shape memory polymers (SMPs), [50] and low-melting point alloys (LMPAs). [51]Although various actuation methods, including thermal-driven liquid-crystal elastomers (LCEs), [52] magneticdriven 3D-printed hydrogels, [53] and light-driven laser projection actuator, [54] have been constantly developed, their application in agriculture appears to be limited due to stringent usage conditions and a lack of substantial performance.
To handle fragile and delicate objects, such as agricultural products, advanced control methods are also essential to ensure the safety of targets during operation.The control system for soft grippers mainly relies on sensors to provide feedback (i.e., closeloop control) regarding the deformation state of the actuators and their contact state with targets and the surrounding environments. [55]Closed-loop control becomes crucial when there is uncertainty in the environment or task, such as when a robotic gripper needs to grasp objects of varying sizes, shapes, and stiffness.[70][71] These strategies are based on internal muscle perception, contact condition perception, target and surrounding temperature sensing, and vision feedback (Figure 2), respectively.By employing several sensors, soft robot systems can obtain various sources of information.The feedback signals from these sensors enable soft robots to perform functions such as deformation sensing, [72] touch sensing, [13] texture discrimination, [73] object recognition, [64] and targets classification. [69]Consequently, the control architecture is designed to accommodate multiple functions, different actuators, sensors, and devices.
In this review article, our main focus is on the investigation of the design and control methods of soft robotic grippers for agricultural product handling tasks.We emphasize essential characteristics required for functioning effectively in field environments, including ease of fabrication and operation, cost-effectiveness, and strong adaptivity.Consequently, we primarily delve into the realm of soft fluidic/mechanical actuators and sensors that have demonstrated reliable performance in grasping agricultural products.The "Soft Robotic Design" section highlights the potential of pneumatic-driven, vacuumdriven, tendon-driven, and jamming soft actuators in crop harvesting applications from their structural and functional perspectives.Moving to the "Soft Robotic Control" section, we explore the sensor structures and functions facilitated by soft actuators, as well as strategies for implementing control Figure 1.The diagram for diversified soft robotic actuators for agricultural product harvesting, especially with four actuation mechanisms: pneumatic [22] (Reproduced with permission.Copyright 2020, IEEE), vacuum [16] (Reproduced with permission.Copyright 2021, IEEE), jamming [39] (Reproduced with permission.Copyright 2010, NAS), and tendon [21,152] (Reproduced with permission.Copyright 2021, AAAS) (Reproduced with permission.Copyright 2022, IEEE); and five configurations: tentacle [14,140] (Reproduced with permission.Copyright 2020, Optica) (Reproduced with permission.Copyright 2022, ASME), finger [122,214] (Reproduced with permission.Copyright 2022, ASME) (Reproduced with permission.Copyright 2021, Mary Ann Liebert Inc.), hand [24,215] (Reproduced with permission.Copyright 2016, AAAS) (Reproduced with permission.Copyright 2021, Mary Ann Liebert Inc.), envelope [16,18] (Reproduced with permission.Copyright 2021, IEEE) (Reproduced with permission.Copyright 2021, Mary Ann Liebert Inc.), and suction [76,129] (Reproduced with permission.Copyright 2022, Wiley-VCH) (Reproduced with permission.Copyright 2021, Springer).
structures using various types of actuators and sensors.Subsequently, driven by the advancements in material science, mechanical design, and computing, emerging trends in this field aim to design grippers with multifunctionality, reliable and powerful grasping capabilities, and increased levels of intelligence.Hence, the "Emerging Directions" section discusses novel approaches to fabricating grippers using advanced materials, mechanisms, and control methods.Finally, we conclude the article by summarizing the development of soft grippers in their application to handling agricultural products in the last section.

Stages of Grasping Process
In the context of autonomous grasping, both pre-grasping (such as approaching, executed by robotic arm) and post-grasping manipulations (such as delivery, performed by mobile robot or assembly line) are essential, as illustrated in Figure 3.The grasping task involves multiple operations and requires comprehensive consideration.In this review article, we specifically focus on the handling task accomplished by soft grippers.The process of grasping involves four stages: 1) activating the gripper to grasp the target object, which is fed to it by a robotic arm, as the grasping force increases; 2) lifting and relocating the target object to the desired position; 3) manipulating either the gripper or arm to orient the target correctly, using their multi-degree of freedom systems; and 4) gently releasing the target.It should be noted that this entire process requires monitoring through various sensors, such as proprioceptive sensors, tactile sensors, and imaging sensors, which will be further detailed in Section 4. In the case of real harvesting tasks in the agricultural field, an additional process of detachment is necessary.This attachment can be accomplished through either pulling, bending, or twisting of the grippers, [74] or through cooperation with other end-effectors. [75]

Grasping Configurations
In the field of agricultural product handling, various gripper configurations have been developed to accommodate the diverse nature of different products.These state-of-the-art grippers can be equipped with either single or multiple actuators to effectively perform grasping tasks.Soft grippers, often drawing inspiration from bionics such as the human hand and octopus, have gained prominence in this area.Notable gripper configurations include tentacle, finger, hand, envelop, and suction, as depicted in Figure 1.Depending on the number of actuators used, these configurations can be broadly classified into two categories: 1) grippers with tentacle or envelope grasping configurations that typically operate using a single actuator to accomplish their grasping function, and 2) grippers with finger and hand configurations that usually require multiple actuators to collaborate and function as an integrated gripper.It should be noted that both single suction disk [16] and multiple suction mouth [76] can be employed for grasping tasks, making it challenging to categorize them into the aforementioned classifications.Suction configurations are often combined with other configurations, working together synergistically to effectively handle various targets. [77][80] Inspired by an octopus tentacle, these grippers twist helically and utilize the inner hole to grasp objects when activated.This configuration has proven effective in handling cylindrical agricultural products such as cucumbers, [81] flowers, [14,78] and stems.However, it encounters challenges when dealing with larger agricultural products of round and rectangular shapes, such as apples, oranges, and pumpkins.The size dimensions of these targets exceed the gripper's capability, limiting its effectiveness.Both soft fluidic and mechanical actuators have demonstrated tentacle-grasping ability.In contrast, the envelope configuration covers most of the targets' surface, enabling universal grasping and high output force.These grippers typically have only one actuator, which uniformly envelops targets from all circumferential directions.The most significant advantage of envelope grippers lies in their ability to grasp a wide variety of target shapes due to their compliance.For agricultural products, this configuration can firmly grasp broccoli, banana, mushroom, grapes, [40] apple, egg, [18] tomato, [16] and more, as long as they match the gripper's size capacity and do not have an extremely high weight.However, the weakness of this configuration is its limited flexibility due to the structure of a single actuator, which makes manipulation tasks challenging for them to perform.Both pneumatic and vacuum grippers can be designed with the envelope configuration.
The finger configuration stands out as the most popular choice in the field of soft grippers due to its flexibility, cost-effectiveness, and easy fabrication.The majority of grippers with this configuration consists of 2, 3, or 4 actuators, arranged in two-finger, three-finger, and four-finger configurations. [8]The fingers are often evenly distributed around the circumference of the gripper, enabling them to exert the same holding force from all directions.This configuration can grasp rounded and elliptical objects such as apples, oranges, and kiwis, [82] as well as objects with a cylindrical shape like bananas, cucumbers, and eggplants, through the parallel arrangement of fingers. [83]This multiconfiguration mechanism of finger grippers will be further explored in Section 3.3.3.Soft fluidic and mechanical actuators can both be designed to work with this configuration.
In contrast, the hand configuration mimics the structure of the human hand, comprising five fingers and one palm.Soft grippers with finger and hand configurations do not share the same inspiration.Finger grippers are mostly inspired from octopus, tentacles, and elephant trunks, while hand grippers are inspired from human hands.Different origin leads to different structure and components: grippers with finger configuration can be assembled with 2-4 fingers or more, while hand configuration only consists of 5 fingers.Moreover, grippers with hand configuration are always accomplished with a palm to support the target while grasping, which may decrease the flexibility but increase the stability.In contrast to the finger configuration, the fingers in a hand gripper are not identical, and each finger performs a specific task.For example, the middle finger is typically the longest, while the thumb generates the greatest force.Hand grippers have been successfully applied in handling various items, including bananas, oranges, [84] grapes, radishes, corn, [85] strawberries, [26] and more.Due to the nonsymmetric arrangement of fingers, hand grippers have the capability to grasp two objects simultaneously using a pair of fingers. [21,86]imilar to the finger configuration, both soft fluidic and mechanical actuators can be designed to work with this hand configuration.

Soft Robotic Actuators and Design
Actuators, serving as the "mover", play a crucial role in controlling and moving a wide range of machines or systems, which typically require a control device and energy source.Soft robotic actuators are designed to leverage the softness, flexibility, and compliance of materials, thereby enabling safe and gentle operation.With diverse designs and fabrication methods, soft robotic actuators have demonstrated remarkable performance across various application fields, particularly in agricultural products and food handling.Considering factors such as design cost, manufacturing complexity, and suitability for agriculture and food application, this review primarily focuses on two categories of soft robotic actuators, namely soft fluidic actuators (SFAs) and soft mechanical actuators (SMAs).These actuators can be further classified into distinct types based on their driving principles, including pneumatic-driven, vacuum-driven, jamming technology, and tendon-driven designs.While some single actuators can function as grippers (e.g., many vacuum grippers), most grasping tasks require multiple actuators to cooperate harmoniously.Multi-activated actuators have also been developed to respond differently based on the type of input energy.Additionally, advancements have been made in achieving tunable stiffness, [87] high reaction speed, [88] and other special functions.Although thermal, light, magnetic, or chemical-driven mechanisms show potential in certain applications, they are not covered in this review due to their relatively limited practical applications in agricultural and food product handling.Considering the design cost, manufacturing complexity, and potential applications in agriculture and food industries, all the actuators discussed in this review hold great promise.

Pneumatic Driven
Soft pneumatic actuators (SPAs) have been a long-standing and prevalent actuation technology in soft robotics, [8] offering significant potential for applications in agriculture due to their safety, durability, ease of fabrication, controllability, low cost, and high lifting weight ratio.These actuators function by applying air pressure to the internal chambers of highly deformable soft materials.Upon inflation, the actuators bend or even grasp due to their structural anisotropy, which favors bending along their low-stiffness direction.The performance of SPAs is evaluated from two perspectives: geometry response, which includes parameters like bending angle and curvature, and mechanical response, which involves measurements of block force and grasping force.It is worth noting that high block force may not always accurately reflect the gripper's actual grasping ability, [89] even though it is often correlated with high grasping force.
One significant advantage of SPAs is that they do not require heavy equipment or strict operational conditions to generate a relatively large grasping force, as this force is highly proportional to the active air pressure.Most finger grippers constructed using SPAs can achieve a bending angle ranging from 0°to 360°, providing a large workspace for grasping.However, the force response of these actuators can vary significantly depending on the fabrication method and material choice.For example, a high-force soft robotic gripper [90] fabricated using fused deposition modeling 3D-printing technology exhibits a maximum grasping force of approximately 50 N, while a similar structure soft pneumatic gripper made from Dragonskin30 material demonstrate a maximum pullout force of only 10 N. [91] These variations in force response emphasize the importance of carefully selecting the fabrication method and materials when designing SPAs for specific agricultural applications.

PneuNet Design
Numerous architectures have been developed for SPAs, with one classical structure being PneuNets, introduced by Ilievski. [34]neuNets consist of a series of chambers embedded in an elastomer.Upon inflation of the internal chambers, PneuNets generate bending motions owing to their asymmetrical design geometry and extensible hyperelastic material.As mentioned earlier, assembling multiple PneuNets actuators creates a soft gripper.While these grippers offer pliable bending motion, inherent compliance, and a simple morphological structure, they do have some disadvantages, such as insufficient force output, gaps between gripper and targets, single motion with one degree of freedom (DOF), and limited shape adaptivity.These limitations warrant further investigation into their working mechanism and performance improvement.
provide a variety of actuation strategies, enabling complex movements such as rotation and inclination motions (multi-activation unit).These movements can also be achieved by splicing different segments (multi-segment).Despite its advantages in low cost and easy fabrication, the PneuNet structure's limited DOF hampers the flexibility that most actuators strive to achieve.Table 1 presents major PneuNets grippers from the literature and compares their strengths and limitations.It is essential to note that the lists provided in this review are not exhaustive but represent the current state of knowledge to the best of our understanding.

Target Channel Design
The PneuNets structure is not the only method to achieve the desired structural anisotropy for SPAs.Instead of molding a series of inflation chambers, pneumatic channels with diverse sizes, shapes, and locations can also achieve the desired motions, as illustrated in Figure 5a,b.Researchers, such as Perez-Guagnelli [80] and Elsayed, [96] have investigated actuators with horizontal, vertical rectangular, and circular cross-sectional chambers, as well as with various channel shapes, size, and locations optimized using the finite-element method.With the aid of simulation data, Runge [97] proposes a machine-learning algorithm to learn the kinematic model of this structure, enhancing the fundamental understanding of its nonlinearity kinematic for gripper control.The kinematic theory of this specific design was presented by Xie, [98] and Gopesh [99] developed a microhydraulic soft navigation robot, which showed their promise working in a constrained environment like the aorta.
As mentioned in the introduction, SPAs with specific pneumatic channel designs can also function as tentacle, [13,14] finger, [100][101][102][103][104] hand, [102] and envelope [17] grippers.These actuators are activated by inflating selected channels, causing them to flex in the opposite direction of the channel location [103,104] (Figures 5e,f ).For actuators with only one-channel designs, inflating the channel causes them to expand to the thinner thickness orientation and bend toward another direction, then the thicker wall side can be utilized to grasp targets, [102,105] as shown in Figure 5c,d.However, these actuators may swell like a balloon (depending on material and cross-section design) on their thinner wall direction, which may bring difficulty when operating under extremely high pressure and condensed environments.Designing twisting channels can be used to implement tentacle-like twining motions to grasp targets with small radii, such as flowers or pencils.However, these grippers are not suitable for handling heavy objects unless a more complex fabrication method is introduced. [13]n the aforementioned study, the design spaces for SPAs are expanded by modifying their cross sections.This approach focuses on altering the size, shape, and location of channels within the actuator to induce bending motions upon activation.Additionally, other soft grippers with unique structures have also been developed, relying on expansion to grasp targets with specified channel designs.In contrast to the one-channel soft actuators discussed previously, which bend toward the thicker wall to grasp targets, some researchers have chosen to use the thinner wall to make contact with objects, resulting in slight deformation but strong support and clipping capabilities, This design approach has demonstrated good shape anastomosis capabilities, [100,101] as shown in Figure 5h,i.Moreover, researchers have explored the use of expanding deformation and circumferential arrangement of channels to achieve enveloping grasping, as PneuNets [34] Ecoflex30, PDMS 7-28 kPa Egg, live mouse Light weight Not suitable for heavy objects Fast PneuNets [35] Ecoflex30, PDMS,/ Elastosil M4601 0-60 kPa Playing keyboard

Hand rehabilitation Variable bending stiffness
Not equipped with feedback system Snake-like soft actuator [15] Dragon Skin 30 0-110 kPa 5.

N Lifting weights High output tip force Low robustness
Hybrid gripper with soft and rigid [20] Dragon Skin30, ABS-P430 0-80 kPa 30 N Tumbler, doll, tape, apple, heat gun, banana, box High fingertip force and fast actuation speed

Did not show a variety of configurations
Multi-segment PneuNet soft manipulators [28] E = 0.8-0.4MPa 0-50 kPa Diverse motions due to different section designs Grasp ability needs to be further investigated Soft robotic hand [25] Dragon Skin 10 0-100 kPa 5.

N Watering can, chair Diverse motions due to 4 chambers
Remaining gap exists between hand and target Soft robotic surface [94] Silicone rubber E625 0-50 kPa 8 N Apple, table tennis, bottle, beaker, keys, tape, banana Envelope grasping Single and simple motion illustrated in Figure 5j.This enveloping grasping capability allows the soft gripper to firmly hold objects by surrounding them with a compliant structure.A comprehensive list of SPAs with specific channel designs form the literature, along with their strengths and limitations, is provided in Table 2.

Fiber-Reinforced Design
Fiber-reinforced soft actuators (FRSAs) offer a promising solution for soft actuators, as they consist of an elastomer actuator wrapped with inextensible reinforcements.In contrast to earlier actuators that relied on inflating selected channels to achieve desired motions, FRSAs inflate purely through a single channel.When unrestrained, an actuator without fiber constraints would expand evenly in all directions.However, the orientation of the fiber (sometimes an inextensible layer was attached to force the actuator from expanding to bending) determines the diverse motion, as shown in Figure 6a.These fibers are placed on the surface of the elastomer, constraining deformation, and causing the actuator to extend longitudinally, expand circumferentially, twist, and bend. [29,106]FRSAs possess two main advantages over SPAs: 1) the orientation of the fiber is programmable, allowing for changes in local stiffness.Additionally, by arranging different fiber patterns segment by segment, FRSAs can match desired trajectories, [29,107] making the gripper more adaptable to the objects' shape, as shown in Figure 6b; 2) FRSAs can handle heavier tasks than SPAs, as the output force is proportional to the energy input (i.e., pressure).The same magnitude of pressure that enables SPAs to bend circumferentially and self-contact can make them challenging to grasp objects.However, the fabrication process of FRSAs may require more time and resources compared to other soft actuators. [108]he finger configuration remains the most traditional and widely used structure for FRSAs to function as a gripper.Researchers have explored various improvements and novel designs to enhance the performance of these FRSAs as effective grippers.Galloway [109] enhanced the shape-matching ability and holding strength by covering the FRSAs with a sleeve surface.Al-Ibadi [110] developed an extension-circular gripper capable of grasping diverse objects.Luo [111] utilized elastic stitches to construct a nonhomogeneous knitting structure hand, which is also considered as an FRSA in this review.The three-layered fabric knitting allowed for a programming design, leading to improved gripping capabilities.Furthermore, combining multiple FRSAs designs has shown potential in improving the overall performance of soft actuators.Zhang [112] employed two FRSAs to activate the bistable laminates gripper, enabling a fast response at the millisecond level under low pressure.Zhou [113] developed a gripper with special contact surface feature design and passive compliance, exhibiting three firm grasping modes.Table 3 lists the FRSAs grippers from literature, along with their respective strengths and limitations.
FRSAs have also demonstrated their suitability for soft continuum manipulators, [114] such as soft arms.These manipulators differ significantly from traditional soft actuators in terms of their lengths and are designed to achieve a broad range of motion to approach targets for various tasks.Some notable applications of FRSAs in soft continuum manipulators include COVID-19 testing robots, [115] neuroprosthetic hands, [26] and obstaclecrossing soft continuum arms. [116]However, as these applications fall outside the scope of this review, they are not extensively discussed in this article.
The three categories of SPAs discussed earlier all rely on positive pressure to achieve grasping functions.These actuators are known for their cost-effectiveness, ease of fabrication, gentle grasping mechanism, high grasp speed, and adaptivity to different environments.These characteristics have made them well suited for handling fragile objects in agricultural, food, and biosystem engineering.The actuator designs using PneuNets structure and specific channel designs share the common idea of creating asymmetric cross sections through channel design.Some grippers use thicker sides to grasp objects through bending motions, while others use thinner sides to grasp through expansion.In contrast, almost all FRSAs adopt a straight channel design, with their unique solutions provided by materials, fiber orientations, or other components like inextensible layers.The inclusion of fiber reinforcement significantly enhances the grippers' robustness and output force.These grippers exhibit high flexibility and have been successfully used to handle various agricultural products such as tomatoes, [117] apples, [118] mangos, pears, [119] eggs, [120,121] peppers, pomegranates, and even large items like watermelon and pumpkin using large grippers. [101]owever, there are two shortcomings that limit the use of pneumatic grippers for agricultural products.First, the limited contact area may result in an insecure grasping state, as these grippers typically only use their tips to hold the target.In fact, under general designs, larger contact areas can only be achieved when dealing with spherical objects.Second, the grasping force cannot be precisely loaded in the ideal direction to optimally support the targets.To address these issues, researchers have focused on programming the actuators' responses through sleeve design, [109] fiber orientation, [29] and channel design, [92] increasing output force and adjusting their stiffness through inextensible layers, [122] rigid component reinforcement, [20,123] and jamming technology.Given the diverse geometries, sizes, weights, and softness of agricultural products, specific gripper designs are required to handle different items effectively.For example, a gripper designed for apple, mango, and pear grasping may not be suitable for large items such as pumpkins and watermelons or different-shaped targets like cucumbers and bananas.Moreover, the generated force for grasping a tomato should be different from that for an apple to prevent damage.Recent developments in gripper design have focused on achieving universal Helical soft fabric gripper [13] Ecoflex30, PVDF 0-1.15 MPa 18 N Tripod, hammer, lemongrass, cucumber, grape.

High load carrying capacity, envelop grasping
Not suitable for tiny object Multidirectional bending pneumatic muscles [103] E625 0-100 kPa 1.83 N Tomato, bulb, banana, aluminum tube, carton Diverse motions due to 4 circular chambers

Low grasping force due to high air chamber volume ratio
Pneumatic actuators with bellows [104] Shore A scale from 35 to 95 A by mixing rubbery with rigid 0-60 kPa 3 N Apple Rapid fabrication due to directly printed 3-D print actuator Low grasping force grasping capabilities and precise control strategies to accommodate various agricultural products and optimize their handling.

Vacuum-Based Design
Agricultural products exhibit great diversity in terms of their shapes, sizes, and mechanical properties, owing to factors such as their semirigid or nonrigid nature, growth stage, and the presence of peel. [124]Dealing with such a wide variety of products requires grippers with versatile capabilities, but traditional pneumatic grippers have struggled to meet these demands.In this context, soft vacuum grippers have emerged as a promising alternative, providing both universal compatibility and gentle handling.
Vacuum actuation has revolutionized soft robotic grippers by employing negative pressure as the driving force instead of traditional positive pressure, both utilizing air as the energy source.Since we have already gone through soft grippers driven by positive pressure, this section will introduce the concept of multi-actuation methods, where an actuator can be activated by both positive and negative pressure.For example, an octopus-inspired SPA, featuring three vacuum suckers on its contact surface, is ingeniously developed.Utilizing a PneuNet structure, this unique design allows the gripper to flex toward the target, while the incorporated sucker securely held the object in place. [77]Similarly, Sandoval devised a remarkable three-fingered PneuNets SPA, where a central sucker added an extra layer of gripping capability [125] (Figure 7a).Other ingenious designs include PneuNets that exhibit bending in the opposite direction of traditional SPAs when subjected to negative pressure (Figure 7b).
Soft vacuum grippers, relying on the principle of envelope and suction grasping mechanisms, showcase exceptional versatility by adeptly adapting to various object geometries through squeezing and contracting actions.This novel approach led to the creation of remarkable grippers such as the cube-like pneumatically actuated soft cubical vacuum (PASCAV) gripper, [126] rubber-band gripper, [127] granary-shaped soft gripper, [128] origami "Magicball" gripper, [40] elastic-membrane gripper, [18] and multimodal envelope gripper. [16]Note that the elastic-membrane gripper requires a unique inflation-deflation process for object manipulation, while the multimodal envelope gripper exhibits the ability to hold subjects using both its inner and outer surfaces, even incorporating a specially designed disk for added versatility.Vacuum grippers displayed notable advantages, such as requiring minimal operating space [125,129] and demonstrating substantial output force due to negative pressure. [76]These grippers harness the power of vacuum to generate significant grasping force without incurring considerable deformations, which distinguished them from certain SPA designs that prioritize the conversion of energy input into two distinct aspects (shape and force response).The inherent compliance and softness of vacuum grippers facilitate their effectiveness in handling a variety of agricultural products, including bananas, [76] eggs, tomatoes, [16] broccoli, mushrooms, and lettuce. [40]However, vacuum grippers face certain challenges, including limited flexibility arising from their single-actuator design, and the complexity of adjusting grasping force through fully openclose control.To address these limitations, complementary integration robotic arms or other end-effectors is necessary.Additionally, novel inner surface designs inspired by origami teeth or gecko-skin structures are suggested to enhance frictional force and minimize slippage.Table 4 provides a comprehensive summary of the diverse vacuum grippers featured in the literature, along with their respective strengths and limitations.

Jamming-Based Design
Jamming technology, a specialized subset of pneumatic-driven mechanisms, has facilitated the development of novel systems capable of achieving remarkable stiffness variation while undergoing minimal volume changes. [43]It is worth notice that although jamming is not a special actuation method, it can be activated under both positive and negative pressure, leading to the stiffness change, and making it difficult to be simply categorized.The majority of grippers utilizing jamming technology are activated through vacuum, although positive pressure jamming also plays a crucial role in adjusting their stiffness. [87,130]enerally, jamming actuators can be classified into three categories: granular, layer, and fiber jamming.Furthermore, this section discusses various advanced jamming techniques, including tensile jamming and honeycomb jamming.
Jamming technology, an intriguing phenomenon observed in granular materials such as sand or coffee, bestows the ability to transform compliant, low-density packing into rigid, high-density packing under the influence of external stress (Figure 8a).This unique capability has found extensive Radius Programmable fiber-reinforced soft actuators [109] Elastosil M4601 0-414 kPa 61 N Rigid foam block High shape adaptability Arrangement of the sleeve location needs to be further investigated Active soft end effectors [110] Rubbery tube, braided sleeve 0-400 kPa 240 N Calculator, tape, card High load capacity Single motion Sleeved bending actuators [217] Silicone (McMaster-Carr #5236K525/ 5236K234) 0-350 kPa 200 N Broccoli, rice noodles, a bag of carrots, cucumber, a 10 pound potato bag High load capacity, durability

Computationally designing and digitally fabricating
Complex design due to machine knitting Figure 7. Vacuum-driven soft grippers.a) A sucker in the middle of a three-fingered PneuNets soft pneumatic actuators (SPA) [125] (Reproduced with permission.Copyright 2022, IEEE).b) Soft vacuum gripper with PneuNets structure [218] (Reproduced with permission.Copyright 2018, Mary Ann Liebert application in developing pneumatic grippers with variable stiffness.Starting from the classical PneuNets structure, Hu [131] developed a PneuNets SPA with a constrain layer composed of granular materials, endowing it with variable stiffness functionality.Furthermore, granular jamming has also been harnessed to empower FRSAs for achieving high surface texture adaptability. [132]Notably, Mitsuda [133] fabricated a six-finger vacuum gripper with foam as granular jamming, which enhances its ability to hold heavy objects with relatively low holding force.Envelope grippers based on granular jamming have also demonstrated remarkable promise. [39,134]In this approach, a cohesive mass of granular material replaces individual fingers and smoothly conforms to the shape of target object when pressed onto it.The granular material contracts and hardens quickly upon application of a vacuum, pinching and holding the object without necessitating sensory feedback.Elaborate investigations have explored the relationship between gripper response and diverse granular materials (coffee, sawdust, glass spheres, and diatomaceous earth), [135] membrane materials (vitrile, vinyl, nitrile, latex, and polythene), [136] and grain geometries (ellipsoid, sphere, superellipsoid, and superball). [134]Notably, granular jamming actuators can also work as a tentacle gripper. [135]ayer jamming is another method of achieving stiffness variation (Figure 8b).This method utilizes planar sheet packings and exhibits significant advantage in scenarios with limited volumes and weights. [137]Compared with granular jamming, layer jamming commonly serves as a constrain layer and is predominantly found in tentacle and finger gripper [87,[138][139][140] due to its reduced malleability.Notably, Crowley [87] developed a positive pressure jamming technology that greatly improved the stiffness of the actuator upon activation of the jamming channel. [139]This technology enabled a shape-locking function similar to LMPAs. [141]In contrast, fiber jamming leverages bundles of threads, enabling enhanced versatility compared to layer jamming by facilitating tunable bending stiffness in multiple directions (Figure 8c). [142]It should be noted that fiber jamming actuators differ from FRSAs as they require specific jamming channels.145] Liu [145] fabricated a three-fingered gripper with FRSAs and fiber jamming to achieve joint-tuning capability.Tensile jamming can be considered a specialized form of fiber jamming, exhibiting significant tensile stiffness changes with minimal alterations in bending stiffness (Figure 8d). [146]Lastly, honeycomb jamming is utilized to control not only stiffness but also the positions and areas of the stiffening regions (Figure 8e). [147]Table 5 presents a comparison of the strengths and limitations of the various jamming grippers described in the literature.
The jamming method exhibits remarkable versatility, serving as both the primary framework of a gripper and a mechanism to vary its stiffness.In certain instances, like layer and fiber jamming, the jamming channel and activation channel are separated, while in others, such as granular and tensile jamming, they are integrated within the same design.The grippers based on layer and fiber jamming exhibit a compact and lightweight nature, surpassing their granular counterparts, owing to their longitudinal configurations. [43]Moreover, granular jamming grippers possess an amorphous structure, which allows for exceptional adaptability to various shapes.In summary, granular jamming demonstrates its potential in envelope grasping configurations, while layer and fiber jamming excel in enhancing the grasping force of finger grippers.However, it should be noted that granular jamming can also be utilized in grippers with finger configurations, serving as an inextensible layer to enable structural anisotropy. [131,132,148,149]Grippers featuring jamming Bioinspired 3D printable soft vacuum actuators [218] Thermoplastic polyurethane, TPU, (NinjaTek) À80-0 kPa 1.64 N Cup, kiwifruit, mandarin, apple High actuation speed, long lifetime

Uniformity envelope grasping, multiple grasp configuration
Require specific orientation of items Sucker octopus-inspired soft gripper [76] Silicone from Beijing Sanxin Jingde Co., Ltd.technology have achieved successful grasping performance on diverse objects, including apples, eggs, [121,148] pears, pitayas, and pineapples, [138] effectively providing a firm grip and adjustable force through their stiffness variation functions.

Tendon Driven
The tendon-driven mechanism, also known as cable-driven mechanism, empowers the motion of soft-bodied actuators Berry harvesting soft robotic gripper [152] Dragon Skin FX Pro 18.94 N Can, pear, strawberry, jar, screwdriver High safety to targets, high harvesting reliability Single motion, potential damage induced by force sensor 3D-printed soft monolithic fingers [153] Soft thermoplastic polyurethane (TPU) 1.93 N Potato, avocado, strawberry, steel rod

Stable grasping with PID control
Low output force Spherical self-adaptive gripper [45] Elastic membrane, water 10 N Cup, paper, glass Self-adaptive grasping Not suitable for tiny object Soft robotic finger with variable effective length [156] NinjaFlex 12 N Plate, wooden block, knife, bottle Shape adaptability due to tendon constrain

Reconfiguration design Single motion
High-speed soft grippers [88] Polyurethane material 3 N Apple, milk can, cooked egg yolk High speed, tendon, pneumatic, vacuum involved

Single motion
Layer jamming structure and tendon-driven gripper [138] Ecoflex50, NASIL4230 40 N Apple, pear, pitaya, pineapple, bowl, screwdriver, plastic water bottle High output force due to layer jamming and tendon
by retracting tendons embedded in the structure and anchored at specific points. [150]Soft actuators driven by a single tendon can achieve a single inward-bending motion [151][152][153] (Figures 9a-c), while some designs such as the PneuNets-like tendon-driven actuator [154,155] (Figure 9e), and the one-tendon soft actuator (Figure 9d) can also achieve shape conformity. [156]However, a variety of motions can be realized through mechanical design, including a four-bar mechanism with the parallelogram mechanism [157] (Figure 9j), multi-joint design, [21,158] and multiple tendons. [21]These innovative designs have demonstrated impressive results, such as the Farm Hand with gecko-skin adhesives developed by Ruotolo, [21] which successfully lifted a pumpkin with high contact area, shear load sharing, and evenly distributed normal stress (Figure 9i).To further augment the capabilities of soft grippers, some designs synergistically combine multiple methods mentioned in this review.For example, a high-speed soft gripper [88] was developed with both tendon and pneumatic-driven strategies, showcasing an ultrafast reaction speed when catching a flying baseball, similar to a flytrap.Moreover, the combination of layer jamming and tendon-driven mechanisms enabled compliant grasping and multimode grasping modes, including enveloping grasping, clamping, sucking, and hooking. [138]Table 6 lists the tendon-driven grippers from the literature, providing a comprehensive comparison of their strengths and limitations.
The transmission of pulling force to soft actuators is facilitated through embedded tendons, which are often generated by electrical motors.In certain cases, only the fingertip of a tendon-driven gripper requires compliance characteristics to apply force gently to targets.The incorporation of flexible mechanisms and soft fingertip material enables the gripper to deform and conform to the shape of the object, thereby reducing the risk of damage.The continuous pulling of the tendon grants soft grippers the capability to apply a wide range of grasping forces, making them versatile enough to handle agricultural products spanning from small blackberries [152] to larger items like pumpkins, carrots, zucchinis, oranges, and broccoli. [159]Notably, multiple tendons can be integrated into a single actuator, allowing for the activation of different tendons to switch bending directions, [160] similar to the specific channel design discussed in Section 3.1.2.These designs illustrate how a single actuator can achieve various motion modes by utilizing multiple activation units, which will be elaborated upon in the following section.

Other Complex Actuators and Grippers Applied in Agriculture
In the realm of soft robotics, achieving a wide range of complex movements often requires the implementation of multi-activation methods, where different activations lead to distinct modes of operation. [161]For instance, a high-speed soft gripper demonstrated its versatility by being able to swiftly clamp a can in 100 ms using a tendon-driven mechanism, while also exhibiting gentle control to pick up an egg yolk through a pneumatic-driven mechanism. [88]However, for the purpose of this review, our focus primarily lies on soft grippers that consistently perform well in handling agricultural and food products, and thus, other complex actuation mechanisms have not been considered in depth. [162]However, in subsequent sections, we will delve into the exploration of various motions possibilities attainable by connecting segments of actuators (multi-segment), incorporating multiple active units (multi-activation unit), and adjusting the cooperation strategy of multiple actuators (multi-configuration) to design soft grippers capable of fulfilling a diverse array of task.
PneuNets structures to enable the actuator to execute twisting, bending, and helical motions, resembling the functionality of a tentacle. [28]Furthermore, the development of an inverse design algorithm has facilitated the manufacturing of these actuators with desired trajectories.Another example of a multi-segment soft pneumatic actuator is presented in Figure 10b, where both fiber jamming and fiber reinforced technologies are utilized.The unique cross-section geometry design grants the actuator varying bending angles and forces under the same pressure.Consequently, the tip unit of the actuator is tailored for tasks requiring lower force but higher bending angles, while the root unit provides greater stiffness to support the entire actuator.As identical segments are progressively assembled, the total length of the actuator increases, leading to a decrease in overall bending stiffness.As a consequence, manipulating the actuator using a hand or finger configuration becomes challenging, and it becomes primarily suited for functioning as a tentacle.
Soft grippers incorporating a suction disk can also be regarded as a type of multi-segment. [16]The implementation of such multisegment designs has expanded the potential applications of soft grippers, making them more effective at handling agricultural products that were previously difficult to grasp using the original gripper configuration.For instance, these designs have empowered grippers with a tentacle configuration, which is primarily suited for cylindrical objects, to effectively grasp pears with varying cross-section radii. [28]Moreover, the integration of different segments can endow the gripper system with new abilities, such as the capability to lift flat objects. [16,76]These multi-segment designs facilitate the development of a collection of soft actuators that can be connected and detached like Lego blocks, exhibiting diverse motions and user-friendly operation.Another potential application of this concept is the development of a soft continuum arm that can traverse obstacles and reach targets, although this aspect will not be extensively explored within the scope of this review.

Multi-Activation Unit Actuator
Soft grippers often require the cooperation of several actuators, each actuator typically consisting of only one activation unit (i.e., one inflation channel, one tendon, etc.), resulting in a bulky design with limited DOF.However, the development of multiactivation unit actuators has revolutionized soft grippers, enabling them to perform complex tasks like the in-hand manipulation mentioned earlier. [164]For example, Jiang [165] engineering an actuator with concentric, eccentric, and helical channels, each inducing distinct motions upon activation.By selectively activating specific combinations of channels, the gripper gains enhanced flexibility.As depicted in Figure 10c, Abondance [22] designed a four-finger soft pneumatic gripper, with each finger equipped with two channels, enabling it to rotate objects by activating opposite fingers and sequentially inflating the channels of one finger.Similarly, Wang [25] developed a soft hand with four channels in each finger, displaying a range of movements through the selective activation of channels, while exhibiting distinct movements under the same energy input.Tendon-driven soft actuators also possess this ability, as evident by a three-finger soft gripper, where each finger is actuated by three tendons, resulting in a more than 150% increase in bending stiffness. [160]Another multi-fingered soft gripper exploited an antagonist tendon to activate a rigid phalange and an agonist tendon to activate a soft buckling rib, thereby achieving a high Universal robotic gripper [39] Granular jamming Single nonporous elastic bag filled with granular matter 90 N Bulbs, LEDs, bottle caps, plastic tubing.

Multiple task applicability Not suitable for large object
One-shot 3D-printed soft gripper [219] Granular jamming Elastic membrane and grain fill with various designs 9.563 N Cube, ball, star, coin High grasping ability in various environments

Not suitable for tiny object
Soft robotic gripper with a variable stiffness [87] Layer jamming Soft TPU bellows 40 N Aluminum cylinder, block, cup bucket

Variable stiffness Single motion
High performance pneumatically actuated soft gripper [140] Layer jamming TPU, PETG 110.2 N Cylinder, 3D-printed material, cup High load capacity Require specific orientation of items Variable stiffness devices using fiber jamming [144] Fiber jamming Tube, fiber 4.28 N Cup, rigid hook Multiple motions, variable stiffness due to fiber jamming

Diverse motions Specific design required at fingertip to increase friction
Honeycomb jamming soft gripper [147] Honeycomb jamming Silicon carbide grit papers, stereolithography

N Cylinder Variable stiffness mechanisms
Only showed the grasping ability of cylindrical objects contact area, shear load sharing, and evenly distributed normal stress.The incorporation of multi-activation units significantly enhances the flexibility of the grippers.In cases where the workspace is constrained, such as when dealing with branches in the field, multi-activation units enable the gripper to adapt to varying gripping modes effectively.

Multi-Configuration Gripper
The universal grasping capability demonstrated by soft grippers with an envelope configuration may encounter certain limitations, such as easy slippage, [40] size constraints, [16] and workspace limitations. [166]However, this ability can also be achieved by continuously switching the gripper configuration.Soft grippers based on finger grasping mechanism with multiple grasping modes have been developed, as depicted in Figure 11, where the switch of configurations is mainly reliant on motors [27] and sliding mechanical channels. [83]Motors can dynamically adjust the distance and orientation of each finger, allowing the gripper to adapt and manipulate a greater variety of objects.For example, a four-finger gripper with tactile sensors (GTac) exhibits diverse grasping modes, including caging, parallel pinch, thumb-three-finger, clasped, and T-shape grasping (shown in Figure 11a). [157]Additionally, multi-object grasping is achieved by using two pairs of fingers to grasp different objects.Another three-finger 3D-printed actuator presented claw, pinch, and scoop positions (shown in Figure 11b), [27] showing successful grasping of various items such as sausage, potato, broccoli, tomato, tangerine, long bean, tofu, pudding, and noodle.Sliding mechanical channels serve a similar function by mechanically adjusting the distance between actuators, [83] making the gripper suitable for objects with various sizes and shapes, as shown in Figure 11d.The concept of multi-configuration not only enhances the universal grasping ability but also preserves high flexibility, exhibiting significant potential in the handling of agricultural products.

Soft Robotic Sensors and Control
In recent years, soft robotic grippers have gained immense popularity due to their simplicity in operation, using energy sources such as air pumps and motors.For effective performance, soft robots require knowledge about the robot itself and the operating environment, which can be obtained through analytical [167] or  [28,115] (Reproduced with permission.Copyright 2021, AIP) (Reproduced with permission.Copyright 2021, Cell PRESS).c,d) Multi-activation unit soft actuators [22,25] (Reproduced with permission.Copyright 2020, IEEE) (Reproduced with permission.Copyright 2021, IEEE).
numerical models. [168]However, in situations where the environment or task is uncertain of varied, closed-loop control becomes essential.This control approach relies on sensors to provide feedback about the deformation state of the actuators and their interactions with the surrounding environment. [55]Moreover, when handling agricultural and food products, additional challenges arise due to the varying sizes, shapes, and colors of the products.The challenges encountered in handling agricultural products are mainly manifested in the following two aspects: 1) the constantly changing mechanical characteristics of agricultural products; and 2) the complexity of the operating environment.As discussed earlier, the application prospect of soft robotic grippers in the field of agricultural products handling is significant.To be specific, the diverse range of crops exhibit varying sizes, shapes, weights, softness, and even different mechanical properties at different stages of maturity.This necessitates soft grippers possessing high versatility, strength, and adaptability.Meanwhile, the operational settings in agricultural product handling vary from experimental benches to industrial assembly lines and agricultural fields.This demands soft robotic grippers that are highly flexible and are integrated with more intelligent control systems.From the perspective of design, handling robotic systems needs to be soft, safe, cost-effective, environmentally friendly, and robust.Innovative designs, such as multi-segments, multi-activation units, and multi-configuration actuators, have been developed to meet the requirements for flexibility and universality, coupled with complex control methods.From the perspective of control, having a fundamental sense of contact allows better monitoring of the grasp process to avoid damaging the products.Visual feedback also aids in analyzing the target and environment, thereby enabling the development of proactive grasping strategies. [169]To tackle these challenges, a diverse range of sensing methods and control strategies are required, considering the different actuation mechanisms and unique designs of soft robots.
In addition, there are four fundamental attributes that play a crucial role in defining the quality of grasping control: 1) accuracy: requiring grippers to grasp targets with precision;  [27,157,158] (Reproduced with permission.Copyright 2021, IEEE) (Reproduced with permission.Copyright 2022, IEEE) (Reproduced with permission.Copyright 2021, SAGE).d) A vacuum gripper attached with a suction disk to grasp flat object [16] (Reproduced with permission.Copyright 2021, IEEE).
2) flexibility: demanding grippers to agilely manipulate targets; 3) reliability: necessitating targets not to be damaged during grasping; and 4) stability: mandating targets not to slip from the grippers when disturbances arise.Of these, accuracy and reliability can be achieved through the use of imaging and tactile sensors, while flexibility and stability require cooperation between multiple components and a sophisticated control approach.These aspects will be elaborated in the subsequent sections.

Type of Sensors
As mentioned earlier, soft robotic grippers require real-time feedback related to the targets, circumstance, contact state, and grasping performance to perform complex varied tasks under uncertain conditions.However, due to their compliance and stretchability, selecting suitable sensors to cooperate with soft robotic grippers can be challenging.While few strategies have been developed to fabricate sensors with high sensitivity at ultrahigh strains, [170] traditional sensors may impact the performance of soft robots.Biological systems provide valuable insights into four sensing functions during object grasping: 1) proprioceptive sensors for perceiving intrinsic deformation through flex and strain information; 2) tactile sensors that mimic skin to detect the stiffness changes, force changes, and temperature variations upon contact with the target surface; 3) optical sensors that offer contact force, shape sensing, and object/material/stiffness recognizing capabilities; [24,66,67] and 4) imaging sensors that provide visual information to plan grasp strategies in advance.Figure 12 shows eight typical sensors used by various soft grippers, with some having multiple functions that can be classified into more than one category.

Proprioceptive Sensors
Proprioceptive sensors primarily focus on the activation status and intrinsic deformation of the actuator such as bending or elongation, rather than contact and grasping performance.An important characteristic of proprioceptive sensors is their ability to function without direct contact with the target and independently of the environment.In the context of SFAs, pressure sensors are also considered proprioceptive sensors as they monitor internal variables.
Commercial flex sensors, like the Long Flex Sensor from Adafruit in the USA and the Flexpoint Sensor Systems Inc., UT, USA, are already employed to provide immediate feedback signals and achieve tasks such as measuring bending angle, curvature, and size recognition upon actuator activation. [60]ustomized flex sensors, such as TENG sensors (a triboelectric sensor, shown in Figure 12a), [58] offer self-powered ability without an external power supply, enabling size sorting and capturing continuous motion and tactile information for soft grippers.These customized sensors can be considered both proprioceptive and tactile, as they measure electrical resistance variations with Figure 12.Typical sensor structures for soft grippers.a) Triboelectric "TENG" sensor consists of silicone, PET, and Ni-fabric [216] (Reproduced with permission.Copyright 2021, Wiley-VCH).b) Tactile sensor consists of fabric traces and piezoresistive fabric [73] (Reproduced with permission.Copyright 2021, Mary Ann Liebert Inc.).c) Air pressure sensor ABPDANT015PGAA5, Honeywell Inc. [177] (Reproduced with permission.Copyright 2022, IEEE).d) Asus Xtion PRO LIVE camera. [69]e) Quadruple tactile sensor [64] (Reproduced with permission.Copyright 2020, AAAS).f ) Nanoscale flexible temperature pressure tactile sensor [62] (Reproduced with permission.Copyright 2021, ACS).g) Soft strain gauge [172] (Reproduced with permission.Copyright 2021, Springer).h) Optical waveguides [24] (Reproduced with permission.Copyright 2016, AAAS).
changes in strain when mounted perpendicularly to actuator's direction of deformation. [171]Alternative solutions involve the use of stretchable strain gauges, which exhibit stable signals and can be mounted in the same location as flex sensors, [172] as depicted in Figure 12g.Another approach is to embed multiple soft strain sensors discretely within the actuator to measure local strains and estimate tip displacement and angle in real time, [173] with optimized sensor placement.Furthermore, the recorded sensor data can be utilized as a dataset for machine-learning algorithms.For instance, Thuruthel [174] utilized pressure inputs and current impedance values from three soft strain sensors (customized polydimethylsiloxane (cPDMS) sensors) as inputs to an LSTM network, effectively modeling the kinematics of a soft continuum actuator in real time while remaining robust to sense nonlinearities and drift.
Pressure sensors are also employed to measure the air pressure applied to SFAs, serving a purpose like a barometer that can monitor the air pressure inside the actuators.For instance, an air pressure sensor with power, ground, and signal pins has been utilized to detect internal chamber pressure. [56,175]This technology has been integrated with bend deformation for size recognition [60] or teleoperation via strain sensors.Another innovative application of pressure sensors involves the use of soft pneumatic sensing chambers, which can provide position and touch information by detecting minimum volume change, [153,176,177] as illustrated in Figure 12c.It should be noted that some studies refer to sensors that measure contact pressure as pressure sensors, but in this review, sensors associated with contact are considered tactile sensors, while all pressure sensors refer to air pressure.

Tactile Sensors
Tactile sensors, also referred to as contact or touch sensors, have become increasingly ubiquitous in the realm of soft robotic grippers.These sensors offer the capability to capture essential contact information such as contact force, [132] area, [148] and grasping success rate, [40] while also providing more advanced features like texture analysis (as shown in Figure 12b) [71] and object classification. [126,141]When dealing with delicate targets in conjunction with aggressive grippers, the potential for catastrophic failure exists.In such scenarios, tactile sensors serve as a vital tool for researchers to evaluate the safety and reliability of the grasping process.Unlike proprioceptive sensors that focus on the internal state of the soft actuator itself, tactile sensors act as a bridge between actuators and targets, providing the means to adjust strategies to accommodate various tasks.
During the process of grasping objects, there are three physical parameters that require attention: mechanical variables such as force, thermal variables such as temperature and thermal conductivity, and optical variables such as wavelengths and refractive indices.Force sensors offer a straightforward way to assess the grasping status by converting mechanical loads into electrical output signals, [41] providing precise force feedback along three directions. [157]While thermal sensor is not commonly embedded in soft grippers due to the predominance of morphological and mechanical parameters, they can offer unique solutions such as material classification based on temperature [62] and thermal conductivity [64] when in contact with targets.Conventional optical sensors operate on the principle of optical reflection between mediums with different refractive indices.As light propagates through the sensor, some of it radiates to the environment, [24] and the amount of light lost is proportional to the magnitude of the pressure. [178]This allows them to function in a manner similar to tactile sensors.For instance, optical fibers [14,66,67] have been utilized for movement and force sensing.Optical waveguides, [24,65,179] a combination of soft optical fiber and cladding, can also detect shape and texture through actuator touching and scanning, as well as softness via actuator contact or pressure and curvature sensing.Sensors within this category have good spatial resolution, sensitivity, high repeatability, and immunity from electromagnetic interference. [180]Proximity sensors, in contrast, can monitor the contact status by measuring the distance between fingertip and object, allowing for the adjustment of closing speed. [19]Notably, some classification methods with tactile sensors were realized by machine-learning algorithms, which utilize sensor signals with labeled targets as a dataset to recognize objects based on the established relationships.

Imaging Sensors
Apart from finger muscle and skin sensing shape, force, and temperature change, visual feedback plays a critical role in human grasping strategies and is equally essential for enabling soft robotic systems to function at a higher level of intelligence.In indoor settings such as a laboratory or food handling factory, high-performance imaging sensors are required for automated grasping of targets, while agricultural products transferred through industrial assemble lines also require robotic systems to detect, recognize, and localize them accurately.The presence of obstacles in agricultural fields, such as other plant parts or other crops, further complicates the grasping process.From the target crop perspective, agricultural products with a variety of shapes and sizes and softness, necessitate different grasping strategies that should be planned to use the information from the "eye" of soft robotic gripper system.The accuracy of grasping is achieved through advanced control strategies based on visual systems, which ensure that the end-effector precisely reaches the detected products without deviation or damaging the produce.
Soft robots can perform single tasks such as pushing manipulation and navigation with the aid of simple visual information. [181,182]In the case of agricultural product detection, imaging sensors can generate 2D or 3D information from the scene.Objects can be detected from 2D images with bounding boxes and the 3D coordinates of the objects in the world frame can be obtained by camera calibration and robotic arm kinematics.As such, soft grippers can be employed for grasping. [68]eep-learning techniques have been used to achieve target recognition and adjust grasping strategies accordingly.For example, with a neural network (NN) and reinforcement learning (RL) frame, Ficuciello [69] provided the robot with a solid foundation to start the learning process that improves its abilities through trial and error via imitation learning.Similarly, Low [27] used you only look once, version 3 (YOLOv3) for object detection and successfully switched between three gripper poses based on detection results.Other networks, including convolutional neural network (CNN) [183] and Transformer, [184] have also provided reliable solutions for object detection, thus enhancing grasping capabilities.The performance of sensors in the aforementioned four categories is documented in Table 7-10, along with their respective strengths and limitations.

Control Strategies
As mentioned earlier, achieving specific tasks through open-loop control can be facilitated by appropriately designing the robot's structure. [55]However, for soft actuators, closed-loop control mechanisms that monitor the actuation of different actuator mechanisms are essential.For instance, open-loop control for SPAs is usually achieved by switching the on/off state of air pumps and valves.However, due to the bulkiness and expense of this equipment, it is challenging to precisely control the pressure or flow rate of air inside each actuator and achieve desired motions for various tasks.A more effective method is to implement a closed-loop control system based on external variables such as strain, [57,170] force, [185,186] and flex. [56,187,188]For fluidic-driven actuators, sensor-based closed-loop control yields higher performance, since the relationship between deformation status (e.g., bending, grasping, inflating, deflating) and activation source has been thoroughly investigated.This relationship is shown in Figure with pressure, [189] [56] and optical sensors, [14] respectively.By employing imaging sensors, researchers have been able to cumulatively measure the deviation angle between the locomotion direction and the object direction, [182] as shown in Figure 13e.Furthermore, teleoperation can be achieved by transferring signals from strain sensors located on human fingers to soft hydraulic actuators, with the pressure measured by pressure sensors and controlled by valves, as shown in Figure 13d.
SMAs driven by tendons are controlled in a different manner compared to those driven by pneumatic systems.While pneumatic systems can easily quantify energy input by measuring air pressure, tendon-driven systems require sensors to collect data on deformation or force.These parameters provide valuable insight into the grasp status, indicating whether it is becoming tighter or looser.Rather than directly measuring quantizable parameters such as tendon displacement or pulling force, the collected data is used to calculate the required grasping force and adjust the control module accordingly.This approach ensures safe and efficient grasping, firmly holding the object without causing damage.Simple deformation sensor data can be directly converted to the control module to make decisions without the need for post possessing, as shown in Figure 14a.To achieve safe and robust grasping, force data, including support force (perpendicular to the contact plane) and friction force (parallel to the contact plane), are collected to calculate the grasping score, [157] as depicted in Figure 14c.Another method involves proportional-integral-derivative (PID) control enabled by force sensors, as shown in Figure 14b.
Gerez [183] investigated a camera mounted on a soft exoskeleton glove to capture scene images and select grasp types via a CNN, as depicted in Figure 15c.Additionally, vision-based grasp systems have been developed through machine-learning methods, such as YOLOv3 [27] or an algorithm combining NNs and RL, [69] both enabling grasping pose modifications according to targets' shape, as illustrated in Figure 15b,d.

Control Methods in Manipulation and Failure Prevention
As noted in Section 3.3, there is an urgent need to develop reconfiguration and manipulation strategies that enhance the flexibility and stability of soft grippers for handling diverse objects, especially agricultural products in the field.These products come in various orientations, necessitating flexibility to adapt to complex environments.Furthermore, damage can occur not only during grasping but also during conveying, requiring a stable robotic grasp to prevent crops from slipping out of the gripper.While envelope and suction grasping configurations offer good stability, they have limitations in improving flexibility.Tentacle configurations that rely on twining motions also provide limited benefits for handling agricultural products.The enhancement of efficiency and flexibility for grippers with envelope, suction, and tentacle configurations predominantly relies on robot arms with  [189] (Reproduced with permission.Copyright 2019, Mary Ann Liebert Inc.), b) flex [56] (Reproduced with permission.Copyright 2018, MDPI), c) optical fiber [14] (Reproduced with permission.Copyright 2021, Optica), d) a soft hydraulic actuator which realized teleoperation via strain, pressure sensor, and dielectric elastomer actuators valves [220] (Reproduced with permission. Copyright 2021, IEEE), e) and vision [182] sensors (Reproduced with permission.Copyright 2021, Elsevier), respectively.
large workspace and multiple degrees of freedom.such, the successful grasping rate is highly dependent on the gripper's position, which falls beyond the scope of the present discussion.In contrast, finger and hand grasping configurations require effective collaboration between multiple actuators to achieve complex motions for both flexibility and stability.In this section, two control methods, in-hand manipulation, and failure prevention, are discussed to further improve the abilities of soft robotic grippers in terms of flexibility and stability, respectively.In-hand manipulation refers to the ability to manipulate small objects within a single hand, [191,192] which can be expressed through two forms: translation and rotation.Soft grippers with finger and hand configurations are the only ones that possess this potential.To achieve in-hand manipulation, single bending  [221] (Reproduced with permission.Copyright 2021, IEEE); b,c) and force sensors [153,157] [190] (Reproduced with permission.Copyright 2022, Mary Ann Liebert Inc.), b) neural networks, and reinforcement learning [69] (Reproduced with permission.Copyright 2019, AAAS), c) convolutional neural network [183] (Reproduced with permission.Copyright 2022, IEEE), and d) you only look once, version 3 [27] (Reproduced with permission.Copyright 2021, IEEE).
motion actuators with only one DOF are insufficient.Soft grippers with effective control can perform complex tasks such as holding pencils, fastening small buttons, twisting cups, and rotating cupcakes.In the agriculture context, soft grippers with strong object-manipulation abilities are required for tasks such as placing products in the correct orientation and twisting to detach fruits.The implementation of multi-segment, activation unit, and configuration structures, as described in Section 3.3, has greatly improved the manipulation ability of soft grippers.Actuators designed with three joints, inspired by the structure of human fingers, are more flexible and can achieve translation, rotation, and twisting tasks.Both pneumatic-driven and tendondriven soft grippers have demonstrated the ability for in-hand manipulation.For example, an FRSA-based gripper with three chambers located at the joints was able to shift a rod by activating specific chambers under a set pressure. [85]Similarly, a soft-rigid hybrid gripper with four bellows was successful in rolling a ball and pouring water, with the rigid part serving as the joints. [193]ndon-driven soft grippers have also demonstrated translation and rotation tasks, with three pulleys representing joints, [158] as shown in Figure 16a,b.Notably, twisting tasks can be accomplished without joint designs as well.For instance, PneuNets structure-based actuators with two parallel channels can twist a cupcake by differentially inflating the channels. [22]Another soft mechanical gripper unscrewed a cap by rotating the tendon embedded inside. [155]he control strategy for each actuator does not significantly differ from the methods mentioned in Section 4.2.The key to achieve in-hand manipulation lies in the effective cooperation between different control units (i.e., single actuators).Both open-loop control with potentiometers and motors and closedloop control with pressure sensors providing feedback, have shown good in-hand manipulation ability in real time.Additionally, a promising approach of vision-based control has enabled complex finger gaiting and multi-model planning, updating plans online to adaptively recover from trajectory Figure 16.In-hand manipulation: a) a fiber-reinforced multi-joint soft hand shifting a rod [85] (Reproduced with permission.Copyright 2018, IEEE), b) a tendon-driven multi-joint griper translating and rotating a cup [158] (Reproduced with permission.Copyright 2021, SAGE), c) a three-finger soft gripper twisting a cap enabled by rotating tendon [155] (Reproduced with permission.Copyright 2021, IEEE), and d) complex finger gaiting and multi-model planning enabled by vision-based control [23] (Reproduced with permission.Copyright 2022, IEEE).
deviation. [23]Moreover, a graph convolutional network was also employed to acquire tactile and geodesic features of a robot hand, achieving dexterous in-hand manipulation with synchronized finger movements. [194]o ensure reliable grasping of objects, mitigating relative motion between the gripper and the object is a critical aspect that can be investigated through intricate control methods.The gripper must not only firmly hold the target during grasping, but also prevent slippage when disturbance occurs.Recent advances in origami and gecko-skin-inspired structures have improved friction to avoid relative slippage.From a control perspective, the system achieves stable grasping by constantly evaluating the grasp status, thereby increasing energy input, and grasping force, or reestablishing the balance of the grasped object after external disturbance.To prevent slippage, two methods are commonly employed: 1) directly measuring the contact forces through force sensors, where regulated controllers can restore grasp forces by updating motor current profile, grasp pose, and stiffness, as demonstrated by Ajoudani. [195]Song [196] identified frictional properties and predicted the breakaway friction ratio to prevent object slip before its occurrence.2) Combining various sensory sources with algorithms: for instance, Averta [197] proposed a solution that combines distributed tactile sensing and RNNs to detect sliding conditions for a soft gripper, achieving a correct prediction of the failure direction in 75% of cases and an 85% success rate for regrasping.
In conclusion, advanced manipulation and failure prevention mechanisms are highly essential for the entire process of handling agricultural products, including flexible approaching, grasping, stable transferring, and placing.Control strategies have been developed to improve the performance of soft grippers, enabling the achievement of complex motions and the prevention of damage.

Emerging Directions and Outlook
Significant advancements have been made in the performance and versatility of soft grippers, thanks to progress in material sciences, mechanics, and computing over the last decade.As novel concepts continue to emerge, the field of soft robotics has undergone significant improvement.In this review, we primarily focus on the functionalization of soft robotic actuators as an end-effector for grasping tasks.[212][213] The future of soft robotic grippers holds promise in three critical areas: 1) material sciences, 2) mechanical design, and 3) perception and control.Emerging research in these directions will further expand the potential of soft robotics and lead to even more exciting possibilities in the future.
Soft actuators rely heavily on advanced materials for their functionality.The responses of output variables such as force, speed, and shape are greatly influenced by the material properties employed.While soft actuators made from single materials already exhibit differentiated responses, the utilization of composite materials has enabled them to perform even more advanced tasks.Dielectric elastomers, for example, can be activated by a high voltage and also exhibit reliable adhesive grasping capabilities.Phase-change materials have expanded the range of working environments for soft actuators.LCEs exhibit diverse motions under varying temperatures, while direct ink writing allows for actuators to be activated by light.Magnetic powder has opened up the possibility of actuation driven by magnetic fields.Furthermore, SMPs and alloys can adapt to the geometry of the target, maximizing the contact area during grasping.However, the frequent use of elastomers can lead to concerns about long-term robustness.Therefore, the progress in self-healing materials is of utmost importance for the development of reliable soft grippers.
Special mechanisms and designs have significantly enhanced the performance of soft grippers.Gecko-inspired structures, for instance, exhibit mechanical adhesion with multiple lines of micro-wedges.Meanwhile, origami and kirigami structures have allowed grippers to achieve a wider stiffness range, enabling them to adapt to different tasks.Grippers can switch between various modes by incorporating specific patterns into their design.Additionally, researchers have explored inverse design methods based on desired outputs.By thoroughly investigating the border workspace of their single deformable unit, diverse motions and outputs can be achieved through the use of optimization methods or machine learning.These advancements in special mechanisms and design approaches have significantly expanded the functionality and versatility of soft grippers, making them a promising technology with a wide range of potential applications.
In contrast to the previous perspectives that emphasize the "soft" aspect, the development of computer vision and machine learning has significantly enhanced the intelligence of soft robots.Recently, vision-based machine-learning algorithms such as CNN have enabled the grippers to accurately locate and identify targets in cluttered and constrained environments.These techniques empower the system to make informed decisions regarding grasp pose and configuration.Regression methods like multilayer perceptron (MLP) and support vector machines (SVM) have further improved the system's understanding of sensor data, allowing for object classification based on tactile information and precise adjustment of force output.Additionally, RL has also shown promising results, as it can be trained through iterative grasping operations and learn from failure cases to achieve a high success rate.The integration of computer vision and machine learning has opened up new possibilities for soft robots, enabling them to operate intelligently and efficiently in complex real-world scenarios.
The fundamental goal of soft robotic grippers is to handle objects with a prominent level of "safety".However, to meet the growing demands, additional requirements such as increased operating speed, high efficiency, high output force, easy fabrication, and low cost have been identified.Nevertheless, several challenges remain to be addressed to fully realize the potential of soft robotic gripper: 1) robustness: continuous operation can result in abrasion and smooth the surface of soft materials, affecting their grasping ability; 2) efficiency: soft robotic grippers have a slow activation speed, making it challenging to operate efficiently in labor-intensive scenarios; 3) insufficient force: the output force is still weaker compared to traditional rigid grippers, limiting their capability to handle certain targets; 4) toughness: soft materials may be punctured by sharp parts of the target, affecting the gripper's durability; 5) limited sensors: the development of flexible, stretchable, and nonintrusive sensors is essential for soft grippers, as existing sensors may add unwanted stiffness and hinder the overall performance.Addressing these challenges will pave the way for the wider adoption and application of soft robotic grippers in various industries and fields.

Summary
Soft robotic grippers have emerged as a promising solution to address the challenges in the agricultural industry, particularly in tasks like robotic harvesting and handling of specialty crops and food products.This technology offers a gentle yet powerful approach to handle delicate agricultural items efficiently while maintaining product quality.The impact of soft grippers in agriculture is significant and has the potential to revolutionize the industry.
This comprehensive review presents various soft gripper designs, including fluidic and mechanical grippers, and introduces five configurations suitable for agriculture product handling: tentacle, finger, hand, envelope, and suction.The review also explores sensor-based control methods, such as proprioceptive, tactile, and imaging sensors.Advanced control approaches have been developed to achieve intelligent functions, such as object classification and real-time grasping condition evaluation using different sensor categories.As the complexity of soft grippers increases, machine-learning methods become increasingly essential for predicting, designing, recognizing, and controlling their movements.
The control section highlights the importance of grippers that can adapt to complex circumstances while ensuring a stable grasp to prevent crops from slipping out.Future developments in soft grippers are likely to focus on enhancing safety, flexibility, and intelligence, thanks to advancements in material science, mechanical design, and computing.The potential benefits of using soft grippers in agriculture include increased efficiency, reduced labor costs, improved product quality, and enhanced safety.In summary, soft robotic grippers have tremendous potential to address the challenges faced by the agricultural and food industry.Continued innovation and collaboration between researchers and industry stakeholders will pave the way for a sustainable and prosperous agricultural and food industry.

Figure 3 .
Figure 3. Motions for handling agricultural products.a) Robotic system comprising a robotic arm, gripper, and mobile robot for approach, handling and delivering tasks.b) Four fundamental grasping motions denoted by human hand.

Table 1 .
Summary of PneuNets structure-based soft actuators.

Table 2 .
Summary of soft pneumatic actuators with various chamber designs.

Table 3 .
Summary of fiber-reinforced soft actuators.

Table 4 .
Summary of vacuum-driven soft actuators.

Table 5 .
Summary of tendon-driven soft actuators.

Table 6 .
Summary of soft actuators with jamming technology.

Table 7 .
Summary of morphological sensors in soft harvesting robotics.

Table 8 .
Summary of thermal sensors in soft harvesting robotics.

Table 9 .
Summary of mechanical sensors in soft harvesting robotics.

Table 10 .
Summary of optical sensors in soft harvesting robotics.