Intelligent Batch Harvesting of Trellis-Grown Fruits with Application to Kiwifruit Picking Robots

Abstract

This study aims to help researchers quickly understand the latest research status of kiwifruit picking robots to expand their research ideas. The centralized picking of kiwifruit is confronted with challenges such as high labor intensity and labor shortage. A series of social issues including the decline in agricultural population and population aging have further increased the cost of its harvest. Therefore, intelligent picking robots replacing manual operations is an effective solution. This paper, through literature review and organization, analyzes and evaluates the performance characteristics of various current kiwifruit picking robots. It summarizes the key technologies of kiwifruit picking robots, from the aspects of robot vision systems, mechanical arms, and the end effector. At the same time, it conducts an in-depth analysis of the problems existing in automatic kiwifruit harvesting technology in modern agriculture. Finally, it is concluded that in the future, research should be carried out in aspects such as kiwifruit cluster recognition algorithms, picking efficiency, and damage cost and universality to enhance the operational performance and market promotion potential of kiwifruit picking robots. The significance of this review lies in addressing the imminent labor crisis in agricultural production and steering agriculture toward intelligent and precise transformation. Its contributions are reflected in greatly advancing robotic technology in complex agricultural settings, generating substantial technical achievements, injecting new vitality into related industries and academic fields, and ultimately delivering sustainable economic benefits and stable agricultural supply to society.

1. Introduction

Kiwifruit has become a popular fruit among consumers worldwide due to its excellent nutritional value (rich in vitamin C) and potential health benefits (such as anti-cancer health benefits) [1]. According to FAO data, 23 countries around the world produce kiwifruit, and with the improvement of living standards, its annual consumption is growing rapidly at a rate of 26%, showing great market potential [2].

In 2023, the top five countries in terms of global kiwifruit production (

Table 1. Global kiwi cultivation status as of 2023.

Countries	Planting Area (in Ten Thousand Hectares) [5]	Yield (Kilograms) [3]	Main Varieties	Export Volume (Metric Tons) [6]
China	19.9138	2362.66	Hayward, Hong yang, Xu Xiang	——
New Zealand	1.5523	662.74	Hayward, Abbott	0.57
Italy	2.4850	391.1	Hayward, Zesy002	0.27
Greece	1.2570	317.08	Hayward, Soreli	0.17
Iran	0.9760	295.14	——	0.15

) were China (2,362,660 tons), New Zealand (662,740 tons), Italy (391,100 tons), Greece (31,708 tons), and Iran (29,514 tons) [3]. The market is projected to grow annually at a compound annual growth rate (CAGR) of 5–6% from 2022 to 2026 [4]. Globally, the majority of revenue comes from China ($89 billion in 2022; Statistics, 2022). However, it can be seen from exports that most of China’s fruit is produced and consumed locally. At present, kiwi consumption is growing from a global perspective.

Take China as an example. According to statistics from the Kiwifruit Branch of the Chinese Society for Horticultural Science [7], as of August 2024, the total planting area in the country was 292,000 hm², of which the fruiting area was about 199,000 hm², accounting for 68% of the total planting area. The largest planting area was in Shaanxi Province (73,000 hm²), followed by Sichuan Province (50,000 hm²) and Guizhou Province (43,000 hm²), then Hunan Province (24,000 hm²) and Jiangxi Province (20,000 hm²). In Shanxi Province, there are two counties each with a planting area of more than 20,000 hectares. Additionally, there are 10 provinces with a combined planting area of more than 7000 hectares. In terms of national output, the total was 3.826 million tons in 2023.

With the growth of consumer demand and the expansion of planting areas, agricultural picking operations are facing efficiency bottlenecks: reliance on manual or mechanical assistance for harvesting leads to high labor intensity [8], an insufficient agricultural labor force [9], and severe challenges to the traditional model [10]. Population aging [11] (e.g., the global number of people aged 65 and above is projected to soar from 761 million in 2021 to 1.6 billion in 2050) has led to a sharp decline in the labor force. This intensifies the transfer of labor to other high-yield industries, resulting in seasonal labor shortages and rising costs, which ultimately restrict the development of the kiwifruit industry. For this reason, scholars have proposed agricultural robots as an alternative solution to compensate for the shortcomings of manual harvesting [12]. Their wide application is regarded as a key way to reduce picking costs and increase fruit farmers’ income [13] and is also expected to drive smart agriculture toward becoming a landmark breakthrough [14].

Since the 1980s, Japan and the United States have been researching picking robots and have achieved fruitful results [15]. However, there are huge challenges in achieving full agricultural automation in production, making it extremely urgent to develop efficient automatic picking robots [16]. So far, researchers from all over the world have developed various fruit and vegetable picking robots, such as apple picking robots [17], tomato picking robots [18], kiwifruit picking robots, and strawberry picking robots [19], among others. There have also been phased research results in kiwifruit picking [20]. However, compared to harvesting robots for other fruits, kiwifruit presents unique challenges (

Table 2. Kiwifruit versus other fruits: a comparison in harvesting.

Challenge Dimension	Kiwifruit	Other Common Fruits (Apple, Citrus, etc.)
Growth Environment	Vine canopy structure, dense foliage, severe occlusion	Relatively regular tree shape, higher fruit exposure
Fruit Distribution	Grows in tight clusters, often squeezed together	Mostly solitary or spaced growth
Fruit Physical Properties	Extremely soft and easily damaged, requires highly precise gripping force control	Thicker peel, higher firmness, relatively resistant to squeezing
Stem and Detachment	Long and frequently entangled stems, require complex “untangling-pulling-cutting” combined actions	Short straight stems, typically detached by “twisting-pulling” or direct cutting
Visual Identification	Similar color between fruit and background, making identification difficult	Distinct color contrast between fruit and background, easy to identify

).

Precisely because kiwifruit harvesting necessitates addressing extreme challenges across three core robotic domains, identification (locating fruit amidst similar colors and severe occlusions), planning (navigating safely within densely entangled environments), and manipulation (executing complex motions with highly compliant force control), the technological spillover from solving these challenges is expected to significantly advance perceptual and operational capabilities of robotics in unstructured environments. The kiwifruit picking robot is mainly divided into five parts, including the visual recognition [21] and positioning system, the mechanical arm system, the end execution control system, and the mobile platform.

This paper conducted extensive literature retrieval and compared the performance characteristics of various current kiwifruit picking robots from the perspective of different picking objects (

Table 3. Comparison of different kiwifruit picking robots.

Equipment Pictures	Recognition Rate	Harvest Success Rate	Harvest Time	Test Environment	Separation Methods	Strengths	Limitations	References
	95.31%	75%	5.8 s per fruit	Orchard	Snap off	The system enables the robot to execute a seamless cycle of grasping, picking, and unloading	Picking accuracy and running speed need to be improved	[22]
	89.6%	51.0%	5.5 s per fruit	Orchard	Snap off	A new kiwifruit harvesting mechanism for effectively harvesting kiwis from the canopy	Accurate identification in complex environments	[23]
	~	Single fruit 100%; 7 fruits 67%	~	Indoor testing/orchard	Pull off, Snap off	The success rate of picking in laboratory tests was high (100%)	There is a significant difference between orchard picking and indoor simulation	[24]
	~	~	1 s per fruit	Orchard	Pull off	~	~	[25]
	98.90%	90.96%	14.89 s per fruit	Orchard	Pull off	The independent and continuous harvesting of kiwifruit in natural environment was realized	Autonomous obstacle avoidance and path planning must be enhanced	[26]
	~	76.78%	~	Orchard	Snap off, Pull off	Indoor test results showed that the identification accuracy and picking success rate were 97.9% and 92%, respectively	Slippage that exists during picking	[27]
	97.7%	86.2%	3.01 s per fruit	Orchard	Snap off, Pull off	By controlling the air pressure, non-destructive harvesting of kiwifruit was achieved	Darker recognition needs improvement	[28]
	~	90.89%	2.49 s per fruit	Orchard	Cutting	A new method for picking kiwifruit: navigation path planning method	The problem of undamaged unloading remains to be improved	[29]
	~	88.7%	6.5 s per fruit	Orchard	~	Effectively grasp the clustered fruits and avoid the interference of bending movements on adjacent fruits	The harvest time has increased compared to before	[30]

). A comprehensive assessment was conducted on the current research status and existing problems of fruit and vegetable picking robots, with the aim of helping researchers quickly understand the latest research progress of kiwifruit picking robots and expand their research ideas.

Literature Search Methodology

The databases searched included (CNKI, Science Direct, Web of Science). The time span of the literature ranged (2010–2023). The keywords used included (“kiwifruit” OR “kiwi”) AND (“harvest” OR “pick”) AND (“robot” OR “automat”) AND “machine vision” AND “mechanical arm” AND “end effector”. The clear inclusion criteria included the following: (1) studies primarily focused on the design, development, or testing of kiwifruit picking robots; and (2) peer-reviewed journal articles, conference papers, and patents. Exclusion criteria included the following: pure image processing studies without robotic applications.

2. Overview of Kiwi Growth

Fruit trees can be classified as trees, vines, or shrubs based on their growth habits. Trees and shrubs can stand on their own, while fruits that grow on vines and shrubs often have flexible herbaceous stems that can bend without breaking, such as berries, kiwis, tomatoes and grapes. Vines need physical support for attaching as shown in Figure 1.

China has abundant kiwifruit resources and diverse varieties (

Table 4. Characteristics of major kiwifruit varieties in China [32].

Varieties	Fruit Heart Color	Fruit Shape	Fruit Hairs	Single Fruit Weight	Vitamin (mg/kg)	Stability in Storage	Maturity Period in the Main Production Area
Hayward	Green Heart	Ellipse	Brown hard hair	100	936	Late-maturing, High Storage Stability	Late September
Qin Mei	Green Heart	Short ellipse	Brown hard hair	100	672	Late-maturing, Storage Stability	Mid to Late October
Cui Xiang	Green Heart	Oval	Yellowish-brown fuzz	92	1850	Early-maturing, Short Storage	Early September
Xu Xiang	Green Heart	Cylinder	Yellowish-brown hard bristles	67.7	1049	Medium-maturing, Storage Stability	Mid-September
Kiyochang	Green Heart	Long cylinder	Long, rough, grayish-brown hair	85	1022	Late-maturing Storage Stability	Mid to Late October
Red Sun	Red Heart	Cylinder	Hairless	95	1358	Early-maturing, Short Storage	Late August
Donghong	Red heart	Long cylinder	Hairless	95	1243	Medium-maturing, Storage Stability	Late September
Golden Peach	Huang Xin	Long cylinder	Short hairs	90	1476	Medium-maturing, Storage Stability	Late September

). Kiwifruit cultivation features are highly concentrated, with Shaanxi Province and Sichuan Province being the core production areas: the combined planting area of the two provinces accounts for 68.3% of the national total, and their output accounts for 94.8% of the national total [31].

In these major production areas, the two most widely used trellis patterns (

Table 5. Trellis cultivation of vine plants.

Trellis Frame	Type	Illustration	Field Application Scenarios	Specification	Strengths	Limitations
Fence Frame Trellis	Single fence frame			Row spacing 2.5 to 3.0 m, frame height 2.0 m. Add a crossbeam to a single fence frame, 1 m wide, and pull two strands of wire at each end of the crossbeam	Low cost, easy to assemble. Easy to manage. Good ventilation and light transmission.	Yield and quality are limited. Later management is difficult. The fruit part is prone to diseases.
Trellis Espalier	Horizontal Scaffolding			The length of the trellis is mostly 5 to 6 m, the height of the trellis roof (close to the plant) is 1.2 to 1.5 m, and the height of the trellis tip is 1.8 to 2.2 m	High yield and quality, with high consistency of fruit. Strong wind resistance, suitable for large-scale cultivation. Extend economic life.	High investment, slow formation The requirements for micro-environmental regulation are high. High humidity under the trellis during the rainy season can easily induce diseases. The temperature of the trellis is high during the hot summer season.

) are the vertical fence trellis and the flat-topped shed trellis [33]. There are various types of trellises, such as single trellises, double trellises, and wide-top trellises, which are suitable for dense planting and early high yield; both sides receive light and have good ventilation and light penetration. It has a large nutrient area, a large number of branches, and a high yield. Berries are good in color and quality. It is easy to operate, conducive to mechanized operation, and suitable for large-scale park construction and large-scale production. Its drawback is that the fruiting part tends to move upward, which cannot meet the growth requirements of high-node varieties. Pergolas are suitable for both buried and non-buried cold areas, with a wide variety of types, large and high pergolas, heavy loads, full utilization of various terrains, good ventilation and light transmission, and few pests and diseases. It is suitable for varieties with vigorous growth. The trellis (Figure 1) is costly in terms of trellis materials and shaping, and it is not easy to achieve high yield early. The fruit is only exposed to diffused light, the coloring is not as good as that of the fence trellis, and the management is more complicated. In addition, there are also combinations of trellises and hedge trellises, which control the growth and distribution of vines to form trellises with both types of trellises. Modern cultivation increasingly requires trellises that are suitable for local climate and production needs.

3. Key Technologies of Kiwifruit Picking Machines

3.1. Detection of Fruit Ripening

To accurately pick ripe kiwifruits, enable batch harvesting, and improve economic benefits, it is crucial to detect their maturity. The techniques for maturity detection mainly include appearance quality inspection, firmness measurement, soluble sugar content analysis, and titratable acid detection [34]. The maturity of kiwifruit is primarily determined by its soluble solid content (SSC) and firmness.

Shang Jing et al. [35] established a PLS-DA recognition model by combining hyperspectral imaging with pattern recognition. The correct recognition rate for kiwifruit maturity reached 100%, demonstrating the feasibility of using hyperspectral imaging technology combined with pattern recognition for maturity assessment. Wang Wen et al. from Southwest University of Science and Technology [36] developed a haptic perception system to evaluate kiwifruit firmness as an indicator of maturity. The contact force during robotic grasping was collected in real time using an 3D tactile sensor, and nonlinear mapping between the feature model and kiwifruit firmness was established to estimate maturity grade. Experiments showed that this method achieved 85% accuracy on a self-built dataset and effectively captured kiwifruit ripening information.

During fruit picking, Dong et al. [37] also determined fruit ripeness using acoustic recognition principles. In addition, Yan Zhenghong et al. [38] combined Principal Component Analysis (PCA) with the K-Nearest Neighbor (KNN) algorithm to propose a method for identifying kiwifruit firmness. Their approach acquires grasping contact information through a tactile sensor integrated into the manipulator, preprocesses the contact data to extract high-dimensional feature sequences, reduces their dimensionality using PCA, and finally performs firmness recognition by training a KNN classifier. This method achieved a firmness recognition accuracy of 90.03%, enabling the picking robot to intelligently perceive kiwifruit firmness.

To enhance the economic value of fruit picking, Zhou et al. [39] summarized two modes: selective harvesting and batch harvesting. The former relies on mobile platforms equipped with machine vision and end effectors to achieve precise picking, though its core challenge lies in coordinating the varying degrees of fruit ripeness within the orchard to prevent unripe or overripe fruits from being harvested together. The key to the latter lies in determining the optimal harvest time to minimize yield loss caused by unripe and overripe fruits as much as possible. To address these challenges, a relationship diagram between picking environment challenges and user demands was constructed, offering new directions for future research.

3.2. Visual Recognition Technology

In recent years, although mechanized fruit-picking technology has matured, it still causes significant damage to branches and fruits, has low efficiency, and manual participation is still required to perform fruit picking. Therefore, the height of the robotic arm is sufficient, the arm span is long, the working space is large, and the operation is stable. There is an urgent need to develop an efficient, non-destructive fruit-picking technique. This section introduces the principles of machine vision technology, the research progress of kiwifruit picking equipment applying machine vision technology, and path planning.

3.2.1. Traditional Recognition Techniques

Traditional digital image processing techniques include color recognition, geometric feature recognition, kiwi texture feature recognition, and feature defect fusion recognition (

Table 6. Traditional recognition techniques.

Technical Routes	Strengths	Limitations	Applicable Scenarios	Reference
Color characteristics	Fast segmentation, strong background contrast	Sensitive to light and subject to significant environment interference	Structured orchards, single-color fruits	[40]
Geometric features	High precision (>95%), excellent sphere fitting	Poor adaptability to complex scenarios	Ball-like fruits (apples/citrus)	[41]
Texture features	Strong surface detail recognition and good night vision adaptability	Noise is prone to interference	Kiwi, fruit with a distinct skin texture	[42]
Fusion of multiple features	Anti-occlusion, anti-shock	High computational complexity	Scenes with dense branches and leaves or mechanical picking	[43]

) of fruits. The object recognition method based on color features uses color spaces (such as RGB, his, and HSV) to segment images and identify target fruits by predicting pixel information. It is particularly suitable for situations where the fruit color is significantly different from the background. This method can effectively identify target fruits in images containing background objects such as leaves. However, the method for identifying color features is susceptible to the influence of lighting conditions and is prone to interference in natural environment. Therefore, it is more suitable for an artificial structured environment.

Xing W et al. [44] from Australia proposed the geometric acquisition network A3N. This method acquires color and geometric data through RGB-D coordinates and applies workspace geometric modeling to assist robot operation. Experimental results show that the average computing time of the network is 35 ms, and the average accuracy of grasping estimates is 0.61 cm (in the center direction) and 4.8° (in the Angle direction). The harvesting success rate of this robot system in field picking experiments reached 70% to 85%. It offers a solution to situations where many fruits are covered by branches and leaves during picking and where the fruits shake during picking. Zhao Yang et al. [40] conducted image recognition based on image features, using feature extraction and SVM classification methods to process multiple watermelon flower body images, successfully achieving a 70% recognition rate for the forward pollination state and flower body traits.

Wang Wei et al. [41] proposed a projection algorithm based on the geometric surface expansion of fruits and vegetables. Through geometric expansion, it broke through the area distortion problem of traditional projection and achieved remarkable results in a controlled environment, with an accuracy rate of 95.47%. However, in complex agricultural and forestry scenarios, it is still necessary to integrate three-dimensional point clouds or deep learning segmentation to improve adaptability. In the future, real-time motion compensation and neural network optimization can be combined to further enhance practical value. Meng Dawei et al. [45] adopted a recognition method that combines the color and texture features of fruits. By extracting texture features for simulation, combining color and texture features can effectively improve the accuracy of fruit classification. However, this method is susceptible to environment noise and thus is suitable for a relatively stable environment.

In terms of night-time identification of kiwifruit, Fu Longsheng et al. [42] proposed an identification method for night-time picking, which involves taking images from the bottom and using a computer to identify and locate the calyx. Under conditions of light ranging from 50 to 400 lux at night, the recognition accuracy reached 94.3%. In addition, for the positioning and recognition under night conditions, Jiao Yingxue et al. [46] segmented the image through PCNN, extracted the target model by boundary detection, and combined the improved three-point circle determination method to locate the fruit. When close to less than 50%, the recognition rate reached 94.3%, and when close to greater than 50%, the recognition rate was 89.05%, significantly improving the accuracy of night-time recognition and positioning.

Chu Guangli et al. [43] proposed a method for recognizing spherical fruits based on machine vision. First, the color of the collected image is normalized and then the fruit is segmented from the background. Then, by extracting the edges of the fruits and using the linear projection method to eliminate the corner points, the least square method is used to detect circles and identify fruits. The experimental results show that the algorithm takes an average of 83.548 ms to process medium-sized images, and the recognition accuracy rate is above 95%.

3.2.2. Deep Learning Recognition Technology

Deep learning utilizes deep neural networks and stochastic gradient descent algorithms to automatically learn and extract target features. Deep networks can learn complex features from a large number of samples, enabling models to achieve better recognition results. Deep learning can be applied to speech recognition, image recognition, natural language processing, etc., in picking robots [47]. According to the vertical growth characteristics of kiwifruit, kiwifruit clusters are mainly divided into single clusters, linear clusters, and other clusters. The extraction dataset includes depth images, color images, and extraction labels.

Hua et al. [48]’s research ranges from digital image processing to deep learning, applied to the three-dimensional reconstruction of fruits and branches for obtaining accurate three-dimensional information. Stereo vision technology has high accuracy in the measurement of close-range targets, but it faces the challenge of feature point matching, especially when fruits are occlusions, and the reconstruction effect is easily affected by changes in lighting conditions. To address the problems caused by the occlusion of leaves, branches, and other fruits in image recognition, Tao et al. [49] proposed a single-stage deep learning segmentation network for locating apple fruits on RGB images. The method can reduce the median error and average error of fruit position by 59% and 43%, respectively. The three-dimensional reconstruction technology based on RGB-D vision sensors measures depth information through structured light or flight time, which can effectively eliminate the influence of illumination changes. However, it still has problems such as low detection accuracy of edge positions and low image resolution. In addition, vision-based fruit reconstruction technology does not rely on mechanical movement and is mainly limited by the performance of hardware equipment, which provides greater potential for future technological development.

Huang Yufu et al. [50] studied a fruit image recognition algorithm based on multi-pixel feature fusion. First, data supplementation and gray-scale normalization were performed on the fruit images in the dataset. By training through the established multi-scale convolutional neural network, the best fruit image recognition model is obtained. The recognition accuracy of this method is as high as 99.4%, which is superior to the existing mainstream algorithms and can effectively support the automatic fruit picking technology.

Gao Jianbo et al. [51] proposed a detection method based on the lightweight YOLOv4-GhostNet network. This method effectively enhances the detection ability of kiwifruit by replacing the original CSP-Darknet53 backbone network model with GhostNet and introducing a feature layer suitable for small target detection in the fusion feature layer. The generalization ability of the model is enhanced. Based on this, in response to the multi-classification recognition requirements in complex field environments, Song Zhenzhen [52] adopted the YOLOv5s model for optimization as shown in Figure 2a,b. To address the interference problem of multiple detection boxes caused by fruit overlap, a method based on “category priority” is proposed to remove the levels of low-value detection boxes, thereby improving the recognition accuracy.

Furthermore, Mali et al. [30] applied the above-mentioned target positioning technology to the actual picking system and developed a kiwifruit harvesting robot based on GG-CNN2 as shown in Figure 2c,d. This scheme integrates the fruit position information provided by YOLOv4 with the grasping angle prediction generated by GG-CNN2. It achieves a picking success rate of 88.7%, a fruit drop rate of 4.8%, and an average single-fruit operation time of 6.5 s in densely arranged scenarios, verifying the efficiency and practicability of the lightweight detection model in the closed-loop control of agricultural robots.

In response to the higher real-time requirements of the picking system, Xu et al. [26] proposed the optimized YOLOv4-GS model through comparative experiments with Faster R-CNN, YOLOv4, and SSD-300 models. Experiments show that this model demonstrates significant advantages in both detection accuracy and recognition speed and is particularly suitable for scenarios with high timeliness requirements in kiwifruit picking, further strengthening the performance boundaries of the aforementioned lightweight architecture.

Extending to a wider range of application scenarios, Peng Hongxing et al. [53] extended this technical framework to the detection of multiple types of fruits such as apple, Lychee, navel orange, and emperor tangerine. By improving the SSD model—replacing the VGG16 of the infrastructure with ResNet-101 and optimizing the parameters—an average detection accuracy of 88.4% was achieved (86.38% for the original SSD); the accuracy reached 89.53% after data augmentation, and the F1 value could reach 96.12% when the occlusion area was less than 50%. It fully demonstrates the cross-species applicability and generalization potential of deep learning models in fruit-picking robots.

To address the challenges of limited computational resources in harvesting robots and detection accuracy issues caused by fruit occlusion, Chen et al. [54] introduced an improved lightweight EDT-YOLOv8n model with optimization methods based on YOLOv8n. They employed techniques such as the Effective Mobile Inverted Bottleneck Convolution (EMBC) module, the dynamic upsampler DySample, and the Task-Aligned Dynamic Detection Head (TADDH) combined with Group Normalization, achieving higher detection accuracy.

Faced with the challenge of recognizing and locating small, dense kiwifruit in complex field environments, Li et al. [55] explored a stereo imaging approach based on binocular cameras and deep learning. They compared the performance of models such as Im-YOLOv7, YOLOv7, YOLOv5s, YOLOv4, and SSD, and found that Im-YOLOv7 improved detection accuracy by 6.87% to 8.17%, increased detection speed by 12.55% to 35.16%, and achieved precise spatial localization with an average three-dimensional positioning error of less than 3.1 mm.

For the development of future smart agriculture, identification technology based on deep learning is an inevitable choice and core technical foundation for building reliable, efficient, and autonomous harvesting robots.

3.3. Positioning and Routing Planning

3.3.1. Fruit Spatial Positioning Techniques

In the picking robot, the most important aspect of the visual system positioning is to obtain the fruit-picking point to guide the mechanical arm to move towards the target. The three-dimensional location of the picking point goes through two stages: First, based on the recognition and segmentation of the fruit, the position of the target picking point on the two-dimensional image is determined. After obtaining the two-dimensional information, three-dimensional coordinates, and three-dimensional posture, other information of the target is obtained through algorithms such as stereo matching and three-dimensional reconstruction. While stereoscopic positioning of the picking point is the ultimate goal, planar positioning of the picking point also has a significant impact on the accuracy of stereoscopic positioning. To achieve three-dimensional positioning for fruit picking, the appropriate three-dimensional positioning hardware system and software algorithm should be selected based on the environment, the agronomic characteristics of the crops, and the precision requirements of the visual system.

Wang Bin et al. [56] developed a spatial coordinate positioning method based on Kinect sensors in the field of kiwifruit picking positioning. Depth images were obtained through Kinect, and a coordinate system was established with infrared rays as the origin. The experimental results showed that the positioning height was less than 2 mm, and it had the ability to precisely locate the fruit. In addition, Gu Xinyun et al. [57] adopted a multi-fruit picking mode of moving in opposite directions, approaching the picking area from the bottom of the fruit and pushing the fruit clusters into rows. This process causes the kiwifruit to rotate around its center of mass, where the fruit stalk breaks under shear force at the separation layer of the fruit stalk, achieving multi-fruit picking. This pattern significantly improves the efficiency of picking kiwifruit.

Luo et al. [58] studied a method for grape recognition and location based on binocular vision for the location and recognition of fruits on trellis vines. The three-dimensional positioning and stereo matching of fruits were achieved through three-dimensional vision technology, demonstrating the feature matching of robot detection and picking points. However, due to the complexity of the line of sight and interference factors such as lighting and shadows, the stereo matching mode is easily affected by noise, resulting in an increase in the positioning phase. In addition, in a dynamic environment, the inconsistency between the shape of the effect image and the original shape simultaneously reduces the accuracy of three-dimensional positioning.

In robotic harvesting vision systems, achieving precise three-dimensional pickup point acquisition is the core objective. The accuracy of this process relies on high-quality two-dimensional plane positioning as its foundation. While technologies like binocular vision enable three-dimensional reconstruction, real-world orchard environments pose significant challenges. Factors such as lighting fluctuations, leaf obstruction, and dynamic disturbances continue to severely impact the stability and precision of stereo matching and positioning systems.

3.3.2. Mobile Platform Path Planning and Picking Path Planning Techniques

In the research of the mobile platform for kiwifruit picking, Chen Zixiao et al. [59] first analyzed the key parameters (such as the relationship between the height adjustment range and the connecting rod angle) through Mat Lab and determined the optimal installation of the hydraulic cylinder. Meanwhile, ANSYS Workbench was used to check the strength of the key mechanisms to ensure the reliability and safety of the platform. Subsequently, Ye Zhenhuan et al. [60] conducted motion simulation through ADAMS software, verifying the stability of the platform under both no-load and full-load conditions, and meeting the requirements of translation and rotational operations. To ensure the accuracy and safety of the picking task, Yang Xiaohui [61] further utilized the least square method to calculate the intersection point between the fruit trees and the ground as shown in Figure 3a, thereby optimizing the picking navigation path and keeping the path travel within the maximum range. Through simulating human visual analysis, the fault information of kiwifruit can be explored to meet the requirements of customized picking design.

For navigation path planning, Li Wenyang [29] proposed a navigation method based on the parameters of the intersection point between kiwi tree trunks and the ground. Firstly, the feature target of kiwi tree trunks is marked with a minimization box, and the midpoint of the bottom edge of the minimization box is taken as the intersection point between the tree trunk and the ground. The support navigation path at this coordinate point is calculated. In addition, Wang et al. [62] proposed a path planning method for kiwifruit picking robots based on deep reinforcement learning. Compared with the traditional DQN algorithm, the re-DQN algorithm can converge to a better solution more quickly. The experimental results show that this method has increased the path length by 31.56%, and the overall navigation time of the robot has increased by 35.72% in total.

When the robotic arm is performing picking tasks, reasonable motion trajectory planning is the key to ensuring the success of picking [63]. In the complex environment of the greenhouse structure, to meet the picking requirements of kiwifruit, Zhang Shasha et al. [64] studied the coordinated picking of double mechanical arms, formulated a method for determining the picking sequence of multi-objective kiwifruit, and proposed a self-obstacle avoidance method for double mechanical arms based on the position layer and speed layer for the coordinated picking of double mechanical arms. At the same time, the singularity problem of the robotic arm was avoided through design. In the research on the kiwifruit picking robotic arm and the end execution control system, Duan Jieli [65] proposed an optimization scheme for the motion planning problem of the fruit-picking robotic arm, aiming to liberate human labor, improve picking efficiency, and reduce harvesting costs.

Aiming at the problem of picking clustered kiwifruit in orchards, Min Fu [66] developed a multi-fruit picking robot as shown in Figure 3b, which adopted an automated continuous picking process (including identification, packaging, separation, and collection). Laboratory tests showed that the average picking time was 9.7 s per string, the success rate was 88.0%, and the damage rate was 7.3%. Picking failures occurred when more than three fruits were processed. To enhance efficiency, Fu Longsheng et al. [67] designed a system based on precise camera recognition and bionic mechanical arms to achieve efficient and non-destructive picking, providing a feasible solution for orchard automation as shown in Figure 3c.

Current research presents diverse solutions and clear trade-offs in terms of real-time performance and computational efficiency. On one hand, traditional methods, such as least-squares-based navigation intersection calculation and minimal bounding box feature extraction, have lower computational load, high determinism, and are easy to implement in embedded systems for real-time response, suitable for more structured environments. On the other hand, emerging deep reinforcement learning methods, such as re-DQN, although showing significant advantages in path convergence speed and global optimization (reducing path length by 31.56% and improving time efficiency by 35.72%), have high computational costs during the training phase, and their real-time online inference heavily depends on hardware capabilities, posing challenges for deployment in large-scale dynamic orchards.

3.4. Picking Technology of Robotic Arms

Currently, in the field of agricultural harvesting robotics, manipulators can be classified into two types based on their degrees of freedom (DOF): low-DOF and high-DOF arms. The manipulator adjusts its spatial posture to quickly transport the end effector to the target position while avoiding obstacles. Low-DOF arms primarily include three-DOF and four-DOF configurations.

Cui Yongjie et al. [68] designed a three-DOF robotic arm with two bionic end effectors equipped with fingers. The picking time for this arm typically exceeds 10 s per fruit, indicating issues of low efficiency and poor flexibility. Feng Keru et al. [69] established kinematic and dynamic models for a four-DOF kiwifruit harvesting robot (as shown in Figure 4a) and conducted trajectory planning aimed at reducing braking. The four-DOF arm has a relatively simple structure, high control accuracy, and fast response speed. However, its obstacle avoidance capability in complex orchard environments is relatively poor. Due to its limited DOF, this type of arm requires a larger physical size to operate in an extended workspace.

To address these issues, Lu Hailin [70] studied a six-degree-of-freedom collaborative robot, constructed a kinematic model of the robotic arm using the DH parameter method, and analyzed its kinematics. The results have proved that the six-degree-of-freedom collaborative robot has obvious advantages in flexibility and growth and can effectively replace manual kiwifruit picking, especially performing better in complex orchard environments. In terms of the robotic arm picking system, Li Zhen et al. [71] established the rod coordinate system of the robotic arm by using the DH method and constructed the kinematic equation by applying the homogeneous transformation matrix. In the Mat Lab environment, the mathematical model of the mechanical arm was established with the help of the robot toolbox. The research shows that this mechanical arm has the advantages of height and arm span, a large working space, stable operation, and can meet the requirements of automated kiwifruit picking under the greenhouse planting mode.

To enhance the fruit recovery rate and accuracy, Au et al. [24] designed a system consisting of four identical robotic arms (as shown in Figure 4b,c), each equipped with a stereo vision system. This system has addressed the problem of robotic arm positioning error through precise calibration, achieving a 100% success rate in the laboratory. However, due to the visual system’s lack of dynamic perception capability, it is unable to cope with fruit occlusion, resulting in an insufficient actual recovery rate. In the future, the vision system requires upgrading to a real-time one to achieve dynamic environmental perception.

The kiwifruit harvesting robot developed by Williams et al. [23] (as shown in Figure 4d) consists of a mobile platform, four three-degree-of-freedom articulated robotic arms, a stop actuator, a vision system, and a kiwifruit storage unit. In field trials, the picking success rate was 51.00%. While the basic picking function has been demonstrated, the results suggest that a parallel robot architecture could be introduced in the future to optimize performance.

He et al. [72] developed a dual-arm picking robot. Simulation results showed that the average picking success rate of this platform was 86.67%, and the collision detection time for a single workpiece was 3.95 ± 0.83 s. These results indicate that the system can effectively plan operational tasks, ensure collision-free performance for the dual-arm picking robot, and thereby enhance the stability and efficiency of the harvesting process.

To improve picking efficiency, shorten the harvesting time, and reduce the fruit damage rate, researchers have also optimized the robotic arm. For instance, Williams et al. [73] integrated a new vision system into a kiwifruit harvesting robot and made key upgrades to two gripper variants. The effectiveness of these improvements was validated through a large-scale real-environment assessment involving over 12,000 kiwifruits. In initial tests (conducted in the four bay regions of New Zealand), commercialization bottlenecks were identified: the picking success rate for 1456 samples was only 51.0%, the average cycle time was 5.5 s per fruit, and the fruit damage rate was as high as 23.4%. The improved system achieved significant breakthroughs, with a harvesting success rate of 86.0% for accessible fruits, an overall harvesting success rate increased to 55.8%, and an average cycle time reduced to 2.78 s per fruit. This makes it one of the most efficient selective fruit harvesting devices currently available.

Current kiwifruit harvesting robotic arm designs face significant trade-offs in degrees of freedom, complexity, and performance. Low-degree-of-freedom (3–4 DOF) solutions offer simple structures, lower costs, and efficient control but suffer from limited flexibility and obstacle avoidance capabilities. This results in low picking efficiency (e.g., over 10 s per fruit) and poor environmental adaptability. High-degree-of-freedom (6 DOF) or collaborative robotic arms, while providing greater flexibility and workspace for handling complex environments, require significantly increased system complexity, modeling costs, and control difficulty. Moreover, they still depend on high-precision vision systems for support.

Taking the classic Mitsubishi RV-2AJ 5-DOF robotic arm as an example [74], its performance was validated by comparing it with a three-dimensional CAD Simscape model in MATLAB/Simulink. The validation process incorporated inverse kinematics, velocity control, trajectory planning, and obstacle avoidance. By integrating a machine homo sapiens dynamics model, the efficiency and accuracy of the robotic arm in executing complex tasks can be significantly improved. The nonlinear equations governing the behavior of the machine homo sapiens system can be expressed in matrix form, as shown in the following equation:

$$\mathrm{D}(\mathsf{\theta })\ddot{\mathsf{\theta }}+\mathrm{C}(\mathsf{\theta },\dot{\mathsf{\theta }})+\mathrm{G}(\mathsf{\theta })=\mathrm{T}\mathrm{i}-\mathrm{T}\mathrm{f}\mathrm{i}$$

$\mathsf{\theta },\dot{\mathsf{\theta }}$, and $\ddot{\mathsf{\theta }}$ are the vectors of joint angle positions, angular velocities, and angular accelerations, respectively. $\mathrm{D}\left(\mathsf{\theta }\right),\mathrm{C}(\mathsf{\theta },\dot{\mathsf{\theta }}),\mathrm{G}(\mathsf{\theta }),\mathrm{T}\mathrm{i}$, and $\mathrm{T}\mathrm{f}\mathrm{i}$ are mass matrix, the Coriolis and centrifugal term, the gravity term, vectors of applied torques, and the friction torques, respectively. The friction force at each joint can be dynamically modeled as non-conservative, linear, and nonlinear friction models. The accuracy of the model was verified by illustrating the Simscape model of the machine homo sapiens in Figure 5.

However, even high-performance robotic arms (such as six-degree-of-freedom or dual-arm systems) that demonstrate excellent performance in simulations and laboratory settings (with nearly 100% success rates) still show generally low fruit harvesting efficiency in real orchards (approximately 51–86%). This highlights practical challenges: dynamic occlusion, positioning errors, and fruit damage rates (up to 23.4%) remain technical bottlenecks. While system upgrades like multi-arm coordination and visual algorithm optimization can improve efficiency (reducing cycle time to 2.78 s per fruit) and success rates, they inevitably increase costs and deployment complexity. Therefore, a balance must be struck between flexibility, efficiency, reliability, and cost, as no single solution currently exists that meets all requirements simultaneously.

3.5. End Effector Picking Technology

The end effector is a key component of the manipulator, primarily composed of a drive unit, a clamping device, a cutting device, and sensors. Existing picking end effectors can be primarily classified based on the method of fruit separation into the following types: laser cutting [75], tool shearing [76], air suction, and torque breaking [77].

The grasping methods for picking robots are mainly based on three principles: suction, clamping, and a combination of both [78]. The suction method uses negative pressure to grasp fruits via suction cups or tubes. It causes less damage and has a fast response time, but the grasping force is limited. It is suitable for spherical or hemispherical lightweight fruits (e.g., strawberries, mushrooms, and tomatoes) and is typically used for single-fruit picking. The clamping method achieves a stable grasp through a clamping mechanism, with its gripping capacity increasing with the number of fingers. It is suitable for heavier fruits, but the design of the drive mechanism is complex. The combined method integrates the advantages of both suction and clamping, first sucking and then clamping the fruit to improve stability and efficiency. Additionally, given the thin skin of kiwifruit, the pressure applied by the end effector should not exceed 15 N to avoid damage [79].

Wang et al. [80] proposed a design principle for kiwifruit picking, suggesting that the end effector should approach the fruit from bottom to top. Based on the growth and physical characteristics of the fruit, four design schemes were compared and analyzed (

Table 7. Four types of kiwifruit grasping mechanisms.

Comparison of Items
Picking principles	Clamping twist	Clamping rotation	Spatial rotation	Space rotation
Power source of the mechanism	A power source	Three power sources	Three power sources	One power source
The complexity of the institution	Simple structure	Simple structure	Complex structure	Simple structure
Fruit stress damage	Less damage to the fruit	Less damage to the fruit	The fruit is more damaged	The fruit is less damaged
The operating time of the institution	Short	Short	Long	Short
The stability of the institution	Unstable	Stable	Stability	Stability
Environment impact	Impact	No effect	Influence	No effect
Effects on adjacent fruits	No effect	No effect	Influence	It does not affect

). Experimental results showed that for single-fruit picking, efficiency was optimal when the inertial axis formed a 30° angle with the fruit stalk. The clamping success rate was 94.87%, the picking success rate was 79.49%, and the average picking time per fruit was 4.86 s. However, the damage rate was as high as 19.4% (6 out of 31 fruits were damaged). For multi-fruit picking, the overall success rate was 77.5% (81.25% for double fruits and 75% for triple fruits), the average picking time per fruit was reduced to 1.63 s (4 s for row picking), and the damage rate was significantly reduced to 6.5% (2 out of 31 fruits were damaged). Consequently, the multi-fruit picking scheme was selected as the optimal design, as it balanced both efficiency and a low damage rate.

A common challenge in mechanical harvesting is fruit damage. To address this, researchers have proposed various solutions. For instance, to mitigate damage caused by direct contact, the development of flexible end effectors [81] can be considered. Furthermore, kiwifruit picking robots can be integrated with flexible adaptive bionic manipulators to enhance adaptability during the grasping process, thereby reducing damage.

The primary method for harvesting kiwifruit employs a clamping–rotary end effector. Among existing designs, Chen Jun et al. [82] developed one for a kiwifruit picking robot (as shown in Figure 6a). By testing the compressive resistance and other relevant physical properties of kiwifruit, they developed corresponding mechanical devices, perception systems, and control systems. Experiments showed a grasping success rate of 100%, a picking success rate of 90%, and an average picking time of 9 s.

Meanwhile, Mu et al. [83] designed another end effector (as shown in Figure 6b) for harvesting clustered kiwifruit grown under scaffolding conditions. Their design utilizes bionic fingers and cam mechanisms to achieve an integrated operation of “picking-unloading”. It approaches the fruit from below for grasping and separates it by bending the stem downward. This structure is simple and efficient, with an average picking time of 4–5 s and a success rate of 94.2%.

Liang Yong et al. [84] designed an end effector for red-fleshed kiwifruit. Tests verified that red-fleshed kiwifruit has good compressive resistance and that the end effector can effectively pick the fruit. When equipped with a Delta ECMA 200 W servo motor, the load current remained stable at 1.3 A, achieving a 100% success rate, with each picking cycle taking approximately 8 s.

Liu Yadong [85] developed a trajectory groove for the end effector based on an anti-cam mechanism. By converting linear motion into oscillating motion and optimizing the groove structure using geometric principles, he achieved an integrated operation of “grasping-picking-unloading”. The final structural parameters of the actuator were determined through a simulation analysis of the machine’s motion, significantly enhancing the stability and operational efficiency of the picking machinery.

To address the problem of kiwifruit cluster distribution, Min Fu et al. [86] designed a multi-fruit cutting linkage picking end effector (as shown in Figure 6c). The manipulation range was determined based on the spatial distribution parameters of the fruit cluster. A multi-fruit stable support mechanical model was constructed to analyze the support force, and the stem-shearing device was optimized through a kinematic study. A dual-sensor fusion recognition method was employed to improve positional discrimination accuracy. Tests showed that the average picking time was reduced to 8.28 s per cluster, with a net picking rate of 87.5% and a damage rate of only 7.5%, demonstrating efficient and reliable performance for clustered kiwifruit.

He, Z. et al. [87] designed a kiwifruit picking end effector (as shown in Figure 6d) that separates the stem through a rotational action. Separation is achieved when the stem bending angle reaches a preset threshold, enabling continuous and efficient picking. The study used fruit separation time, damage rate, and harvest success rate as evaluation indicators and employed the Response Surface Methodology (RSM) for parameter optimization. Experimental results showed that under optimal conditions—a clamping force of 3.05 N, a separation angle of 65.75°, and a separation speed of 60.03°/s—the picking performance was optimal, balancing efficiency and damage control.

Future research should not be content with merely reducing its own damage rate; it should aim explicitly at achieving this commercial threshold (<5%) and proposed possible technical directions (such as force control, flexible materials). Additionally, Ma L. et al. [30]’s study added that the separation force between the fruit stem and the fruit is 3–10 N.

For the practical application and commercialization of end effectors, future research needs to focus on the following three aspects: (1) Given the complex and diverse working environment encountered in fruit picking, it is very important for end effectors to avoid trees and other obstacles in real time. This requires the development of software and powerful algorithms to enhance the success rate and efficiency of the harvesting process. (2) The mechanical structure of the robot needs to be enhanced; for instance, by combining a highly adaptable articulated manipulator and an end effector to select kiwifruits of different shapes and sizes. (3) The control system can apply multi-target recognition and multi-manipulator collaboration to enhance the picking efficiency of the robot, thereby increasing the robot’s versatility, reducing the overall cost, and promoting the commercialization of the robot.

4. Challenges and Trends

Kiwifruit harvesting robots hold significant potential for alleviating labor shortages and improving efficiency, yet several critical challenges impede their widespread commercialization. Currently, a primary limitation is the low recognition accuracy for clustered and occluded fruits in unstructured orchard environments. While deep learning-based vision systems have made substantial progress, their performance suffers when dealing with dense foliage and overlapping fruits. Enhancing these algorithms through expanded datasets, advanced feature extraction modules, attention mechanisms, and model simplification is crucial for improving both the speed and accuracy of fruit detection.

Beyond perception, issues related to execution also pose considerable obstacles. The damage rate during picking remains unsatisfactorily high, often due to the lack of sophisticated end effectors capable of handling delicate kiwifruit. The development of sensor-rich end effectors—integrating tactile, force, and proximity sensors—is essential to enable adaptive grasping and minimize mechanical harm. Furthermore, the breeding of new kiwifruit varieties with easy-separation traits may also offer a biological solution to reduce harvesting damage.

System-level inefficiencies further compound these problems. The total time per pick remains high, mainly due to computational delays in visual processing. Here, the integration of high-speed communication technologies such as 5G can facilitate offloading computationally intensive tasks to the cloud, thereby reducing latency and improving real-time performance. Moreover, the inherent complexity and variability of orchard environments necessitate a holistic approach that combines robotics with agronomic practices. Optimized pruning, trellising, and fruit thinning can significantly improve fruit visibility and accessibility, thus simplifying the robotic harvesting process.

Finally, the high manufacturing cost and limited versatility of current platforms hinder large-scale adoption. Most existing robots are designed for specific tasks and environments, particularly the challenging conditions of mountainous kiwifruit orchards. To address this, a modular design strategy—enabling the interchange of end effectors and perception algorithms for different crops—presents a promising path toward multi-functionality and cost reduction.

In summary, future efforts must focus on developing more robust and adaptive perception algorithms, designing sensitive and reliable end effectors, leveraging high-speed communication for computational efficiency, promoting agronomy–robotics integration, and pursuing modular and scalable robot architectures. These advancements together will pave the way for the practical deployment of intelligent and economical kiwifruit harvesting robots.

5. Conclusions

Research on kiwifruit harvesting robots is considerable, primarily focusing on machine vision for precise identification and positioning as well as on end effector design. Deep learning has significantly enhanced the perceptual capabilities of these systems: for fruit recognition, it offers high-precision feature extraction and strong anti-interference capabilities, representing a major breakthrough. In terms of positioning, while current technologies can achieve millimeter-level accuracy in static conditions, their robustness in dynamic environments and adaptability to complex fruit formations remain significant bottlenecks. However, the overall technology still faces challenges such as sensitivity to environmental variations, heavy reliance on large datasets (leading to weak cross-species generalization), and high commercialization costs. While end effectors can achieve efficient picking within seconds with success rates exceeding 90%, their performance is constrained by difficulties in controlling damage to individual fruits and insufficient adaptability to densely clustered fruit. Future efforts should promote the deployment of lightweight models, the fusion of multi-modal sensors, the optimization of adaptive algorithms, and bio-inspired flexible design in a coordinated manner. Integrated with strategies for multi-robot collaboration and parameter self-learning, these advancements will drive the evolution of the perception–execution closed loop, ultimately advancing harvesting robots from single-point, high-precision operations toward fully autonomous systems for entire orchards.