Conventional fruit sorting methods are based on visual assessment. Various factors such as ripeness, quality, decay, disease, and injury are considered for sorting (Wendel et al., 2018; Gabriëls et al., 2020; Kang and Gwak, 2021). On-farm sorting requires a considerable amount of labor and low productivity is prone to human fatigue, and is biased by inspector experience, resulting in product variability and failure to meet quality standards (Rysz and Mehta, 2021). Given these challenges, recent studies have focused on the application of AI models to on-farm sorting to increase harvest automation and reduce labor costs. Table 1 shows examples of on-farm sorting tasks solved by AI models.

On-farm sorting tasks that can be addressed by AI models mostly include maturity evaluation, quality identification, injury detection, and decay and disease detection for fruits such as apples, mangoes, coconuts, blueberries, etc. (Kuzy et al., 2018; Caladcad et al., 2020; Gabriëls et al., 2020; Kang and Gwak, 2021). On-farm sorting aims to remove fruit that is defective and not suitable for the fresh market and to grade fruit immediately after harvest so that it can be stored differently to achieve cost savings. AI systems have been deployed to replace human workers, overcoming visual variability, inaccuracy, and fatigue in quality grading. Machine vision is a potentially viable solution for improving harvest efficiency by providing an alternative way to automatically and non-destructively grade harvested fruits in the field. The sorting machine is highly dependent on an integrated mechanism, which is mainly composed of an image processing unit, a fruit conveyor, a paddle, and a bin filler, as shown in Figure 2 . Implementing a computer-controlled hydraulic system for on-farm sorting machines has been greatly beneficial to achieve significant cost savings and meet commercial needs.

Pickers spend a considerable amount of time walking to carry full containers, i.e., trays and bags, to the collection center at the end of the row and walking back with the empty container to resume picking. Workers who use the conventional manual picking method are at risk of musculoskeletal problems due to carrying heavy loads and repetitive hand and upper body activities. Faced with the challenges of labor costs, low harvesting efficiency, and occupational injuries, harvesting aids such as on-farm transportation co-robots ( Figure 3 ) have been introduced to assist in harvesting fruits such as apples, strawberries, and grapes (Ye et al., 2018; Faheem et al., 2021; Peng et al., 2021; Lu et al., 2022).

Previous research on on-farm transportation has focused on co-robot scheduling, maneuvering strategies, and minimizing damage to products during transportation. With the goal of developing a co-robot to carry trays and transport fruit in the harvest field, human-robot collaboration models that include picking and walking have been investigated for co-robot scheduling and dispatching policies (Seyyedhasani et al., 2020; Peng et al., 2021; Rysz and Mehta, 2021). Faced with the challenges of bin handling in commercial orchards, navigation, and maneuvering strategies for bin carriers have been proposed to achieve autonomous navigation between tree rows to perform bin handling tasks (Ye et al., 2018; Blok et al., 2019). A fleet of robotic N-trailer vehicles has also been introduced to assist with crop transportation as an alternative to the ultimate mechanization of manual harvesting activities to improve field efficiency (Guevara et al., 2021). To minimize the mechanical damage induced by vibration during post-harvest handling, the vibration mechanism of hanging grape clusters was analyzed to achieve low vibration of the fruit during on-farm transportation (Faheem et al., 2021).

Multiple sensors have been used for image acquisition in on-farm sorting and transportation, including a red-green-blue (RGB) camera, a charge-coupled device (CCD) camera, a hyperspectral camera, a near-infrared (NIR) sensor, visible and near-infrared spectroscopy, and a thermal camera. According to previous studies, the RGB camera is currently the most widely employed for on-farm sorting, especially for surface damage detection, color grading, mass and volume estimation of apples, and ripeness of avocados (Jaramillo-Acevedo et al., 2020; Lu et al., 2022; Mansuri et al., 2022). A CCD camera was utilized for the size and color grading of apples and mass grading of mangoes (Momin et al., 2017; Zhang et al., 2021). A thermographic imaging system combining a thermal camera (7.5–13 µm) with heat lamps was built for internal bruise detection in blueberries (Kuzy et al., 2018). A hyperspectral camera mounted on a ground-based mobile platform was used to detect ripeness in mango and apples (Wendel et al., 2018; Çetin et al., 2022; Wang et al., 2022). A NIR spectrometer (850–1,888 nm) was used for spectral measurements to determine the total soluble solid content in “Medjool” dates to predict fruit quality (Ben-Zvi et al., 2017). VNIR spectroscopy (200–1,000 nm and 673–1,100 nm) was employed for internal browning in mangoes (Gabriëls et al., 2020; Raghavendra et al., 2021). In addition to imaging and spectral techniques, a tapping system using an omnidirectional microphone to record the tapping sound was developed for ripeness detection in coconuts (Caladcad et al., 2020). Table 2 represents the key instruments required for data acquisition for on-farm fruit sorting.

The RGB camera is the most employed sensor for automated on-farm sorting and grading. It should be noted that using a machine vision-based AI algorithm to sort and grade fruits and vegetables need high-quality and decipherable images. Hence, an RGB camera is the most appropriate sensor for the job. RGB cameras are low-cost and can be used to measure texture, color, geometric shape, and defects in fruits and vegetables. The main disadvantage of an RGB camera is its sensitivity to changes in illumination under different lighting conditions. Besides RGB cameras, CCD cameras are also widely used for on-farm sorting because of their low cost. Hyperspectral, NIR, and thermal cameras are more expensive sensors and provide different spectral ranges from visible light, which are universally utilized for internal defect and ingredient detection due to their sensitive spectral range for internal damage and composition. Hyperspectral imaging includes a large amount of spectral data, providing more information for defect and contaminant detection in on-farm sorting. However, the cost is the major drawback of hyperspectral cameras.

With regards to cameras and sensors employed for on-farm transportation, LiDAR sensors, and positioning systems are more commonly used than vision-based sensors due to their robustness for robot localization and navigation in outdoor environments. A 2D LiDAR scanner was placed on the front of the orchard transport robot to observe tree growth on each side (Blok et al., 2019). Positioning systems such as GNSS, GPS, RTK-GNSS, and RTK-Global Positioning System (RTK-GPS) were used to obtain ground truth position information and schedule the path for the bin-handling machines (Ye et al., 2018; Blok et al., 2019; Peng and Vougioukas, 2020; Peng et al., 2021). An inertial measurement unit (IMU) was employed for the measurement of the location and heading angle of on-farm harvesting-aid robots (Ye et al., 2018; Blok et al., 2019). A wheel encoder was used to measure wheel odometry (Blok et al., 2019). A load cell was used to sense the load on the tray to calculate the amount of fruit collected for transportation path scheduling (Anjom and Vougioukas, 2019; Peng and Vougioukas, 2020; Peng et al., 2021). A uniaxial force sensor was inserted between the grape clusters to calculate the peak hanging force and realize the vibrations generated in the grape clusters during postharvest handling (Faheem et al., 2021). Table 3 shows the key instruments required for data acquisition during on-farm transportation.

In previous studies, LiDAR sensors have been most commonly used for on-farm transportation. They have been widely used for yield estimation and geometric measurements of canopy value at the harvest stage due to their ability to be unaffected by illumination (Westling et al., 2020; Chen et al., 2022; Yu et al., 2022). For on-farm transportation, the main ability of LiDAR is that it can directly deliver 3D localization data to co-robots navigating in the field and orchard. Positioning systems such as GNSS, GPS, RTK-GNSS, and RTK-GPS have been employed to obtain the locations of co-robots for bin and tray scheduling. Other sensors, such as the IMU, wheel encoder, load cell, and force sensor, have been adopted to control on-farm vehicles and quantify harvesting efficiency for co-robot route planning. The accuracy of the sensors is crucial to the localization and navigation of on-farm transport vehicles. With the development of sensor technology, the accuracy of LIDAR scanners and positioning systems has been limited to the centimeter level, which is within the tolerance of precision agriculture (Ye et al., 2018; Blok et al., 2019). The accuracy of the IMU used to measure the heading angle of the on-farm vehicle was 0.1–5°. The accuracy and details of other sensors, such as wheel encoders, load cells, and force sensors, employed to evaluate harvest efficiency were not mentioned in the relevant literature, but they can be considered suitable for on-farm vehicles based on the efficiency of the experimental results.

AI models have been used for on-farm sorting along with the significant development of machine learning, deep learning, and machine vision techniques. On-farm grading is mostly aimed at identifying the size, mass, ripeness, surface and internal defects, and diseases in fruit for further transportation, storage, and packaging. Surface and internal defect detection for fruit classification is useful to improve product quality and boost profits. Moreover, on-farm defect detection could analyze product degradation and damage and take preventative actions to halt the deterioration process. Fruit firmness analysis can assist in improving the packaging process by maintaining consistency. AI-based techniques provide a solution for automatically unloading graded fruit into bins to improve harvesting efficiency.

Various image processing methods and deep learning models have been introduced to achieve automated on-farm sorting. Artificial neural network (ANN) and support vector machine (SVM) models have been developed to estimate the mass and volume of apples and mangoes (Utai et al., 2019; Mansuri et al., 2022). Classification of bruises in blueberries was performed using linear discriminant analysis, SVM, random forest (RF), K-nearest-neighbors, and logistic regression classifiers (Kuzy et al., 2018). ANN, RF, SVM, convolutional neural network (CNN), multiple linear regression (MLR), and partial least squares (PLS) were used to classify the ripeness and firmness levels of apple, coconut, kiwi, and mango (Torkashvand et al., 2017; Torkashvand et al., 2017; Wendel et al., 2018; Caladcad et al., 2020; Wang et al., 2022). ANN and PLS were used for identifying browning and total soluble solids in “Medjool” dates and mangoes for quality measurement (Ben-Zvi et al., 2017; Gabriëls et al., 2020). The performance of the AI models is presented in Table 4 .

Spectroscopy techniques such as hyperspectral spectroscopy and visible and near-infrared spectroscopy, which include large amounts of data, have been used for the chemical detection of various fruits due to their sensitivity to nutritional and constituent content that can be detected by external inspection (Ben-Zvi et al., 2017; Jaramillo-Acevedo et al., 2020; Çetin et al., 2022). ANN models, which are robust for pattern recognition in a large amount of data, are used for hyperspectral imaging analysis in chemical detection, such as total soluble solids content, mineral nutrient content, and dry matter content for mangoes, bananas, blueberries, etc. Visible image size and external defects. ANN models have shown great potential in grading, which are superior to other models such as PLS, SVM, and RF (Wendel et al., 2018; Gabriëls et al., 2020). Visible imaging has been employed to determine physical features such as size, mass, color, and external defects and bruises in fruits. CNN models have been introduced to replace manual inspection in fruit sorting by extracting physical features from images and using them to detect the size, mass, color, and external defects and bruises (Wendel et al., 2018; Kang and Gwak, 2021). ANN and CNN models provide solutions for making quality decisions and improving postharvest efficiency through on-farm sorting, which will be soon widely adopted for automated on-farm sorting systems, especially in small commercial farms and orchards.

Workers’ picking speeds vary based on fruit distribution and the individual’s work pace. To address this issue, crop transport harvesting-aid robots have been developed to manage full and empty bins/trays as harvesting aids to increase the productivity, safety, and aesthetics of manual harvesting (Fei and Vougioukas, 2021; Guevara et al., 2021). Since harvesting-aid robots are subject to the constraints of space between crop rows and the cooperation with pickers, maneuvering strategies are required to determine the transport path and operation to deliver full and empty bins in the field. Modeling frameworks for the combined tasks of manual harvesting and robot-assisted crop transport have been proposed to support specialty crops that require labor to distribute and pick fruit. Models describing the interactions between robots and humans during fruit harvesting, such as deterministic predictive dynamic scheduling (Peng and Vougioukas, 2020), a stochastic model (Seyyedhasani et al., 2020), and a risk-averse optimization approach (Rysz and Mehta, 2021), have been utilized to identify optimal harvester servicing policies. Evaluating the efficiency of pickers is vital for scheduling co-robots to transport empty and full bins in the field. Mechanistic gray box models were addressed to predict the time when a picker will fill up the tray being currently harvested to calculate tray transport request time, with the goal of increasing harvesting efficiency (Anjom and Vougioukas, 2019). Field navigation is another core feature for on-farm transport harvesting-aid robots, which includes co-robot localization, a vehicle steering scheme, and path scheduling. Probabilistic localization algorithms, i.e., the Kalman filter and the particle filter using a 2D light detection and ranging (LiDAR) scanner, were developed for in-row robot navigation in orchards (Blok et al., 2019). Maneuvering strategies were created for four steering vehicles to complete different bin-handling tasks in commercial orchards (Ye et al., 2018). To support cooperation between pickers and harvesting-aid robots, cooperative navigation strategies were used to generate a feasible path to arrange and allocate vehicles and trailers for transporting empty and full trays between field rows (Guevara et al., 2021; Peng et al., 2021).

Small crews of harvesting-aid robots that assist big picking crews by transferring full and empty trays can boost harvesting efficiency by minimizing pickers’ unproductive walking times. The study of robotic postharvest handling can benefit specialty fruits such as blueberries, lychees, grapes, and cherries. Harvesting-aid robots have been used in conjunction with pickers, following the departure times, harvesting sequence, computed vehicle routes, and the number of trailers as guidelines. Harvesting-aid robots can handle the transport of empty and full bins and trays between the collection stations and the pickers, thus improving harvesting efficiency by decreasing the unproductive walking times of the pickers in the field. A good maneuvering strategy for a four-wheel independently steered robotic bin-handling machine can effectively accomplish the navigation of bin transport on the desired trajectories without damaging fruit trees or hitting bins. Hanging force analysis for grape clusters during on-farm transportation is helpful to realize the low vibration of fruit clusters, which provides the theoretical basis for reducing berry drop during on-farm transportation (Faheem et al., 2021).