Design and Preliminary Evaluation of Automated Sweetpotato Sorting Mechanisms

Abstract

Automated sorting of sweetpotatoes is necessary to reduce labor dependence and costs that are significant at today’s sweetpotato packing sheds. Although optical sorters have been widely adopted in commercial packing lines for many horticultural commodities, there remains an unmet need to develop dedicated technology for the automated grading and sorting of sweetpotatoes. Sorting mechanisms are the critical component that physically segregates products according to quality grades determined by a machine vision or imaging system. This study presents the new engineering prototypes and evaluation of three different pneumatically powered mechanisms for sorting sweetpotatoes online. Among the three sorters, the sorting mechanism, which employs a linear air cylinder to drive a paddle directly striking products, achieved the best overall accuracy and repeatability of 98% and 96.8%, respectively, at conveyor speeds of 4–12 cm/s. The sorter based on a rotary actuator also delivered decent accuracy and repeatability of 97.9% and 95.6%, respectively. The best-performing sorting mechanism was integrated with a machine vision system that graded sweetpotatoes based on size and surface defect conditions to separate graded sweetpotatoes into three quality categories. The errors of 0–1% due to the sorting process were obtained at conveyor speeds of 4–12 cm/s, confirming the efficacy of the manufactured sorting mechanisms. There was a declining trend with the conveyor speed in the performance of the sorting mechanisms when evaluated either in a standalone or integrated configuration. The proposed sorting mechanisms that are simple in construction and operation and of low cost are useful for developing a more full-fledged sorting system. More research is needed to enhance sorting performance and conduct extensive tests at higher conveyor speeds for practical application.

1. Introduction

Sweetpotato (Ipomoea batatas) is an economically important root vegetable crop. In 2023, the United States (U.S.) harvested approximately 126,300 acres of sweetpotatoes, yielding a farmgate value of around $676 million [1]. The sweetpotato crop in the U.S. is primarily produced for the fresh market, representing about 90% of the total production value, with the remaining for processing. Ensuring the delivery of high-quality fresh sweetpotato roots to the marketplace necessitates a meticulous sorting process at packing facilities, which is currently done manually [2]. Manual sorting is, however, labor-intensive, accounting for up to 50% of packing line labor costs, and susceptible to human evaluation error, negatively affecting consumer satisfaction [3]. Therefore, there is a crucial need to develop labor-saving, automated sorting technologies for sweetpotato packers to remain profitable and competitive.

However, automated sorting of sweetpotatoes is challenging due to the irregular shape of the commodity [4]. Sorting sweetpotatoes for size has been fairly mechanized at packinghouses using sizers that consist of expanding-pitch parallel rollers at specified spacings, although such equipment does not directly measure product size. The mechanical sizer leads smaller roots to drop through the gaps between adjacent rollers early on, while larger roots travel further before being deposited through wider spacings into a separate area [2]. Optical sizers are also commercially available for online size measurement of sweetpotatoes, which utilize imaging technology to measure the size of sweetpotatoes traveling on a conveyor belt. These sizers more accurately gauge root size and mechanically eject them into corresponding channels, allowing for finer size grading and higher efficiency compared to traditional mechanical sizers [5]. However, belt conveyors do not rotate sweetpotatoes to image their full-surface, and sorting sweetpotatoes for other important quality characteristics, such as defects, is still performed manually.

Machine vision technology offers a promising solution for automated fruit sorting and has been commercially deployed across a range of horticultural commodities, such as apples, citrus fruits, nectarines, and blueberries [6,7,8]. Sofu et al. [9] described a two-lane automatic apple sorting system equipped with color cameras that sorts fruit based on size, color, weight, and surface defects. This system employs a roller conveyor to rotate the fruit for imaging and a transporter conveyor with bowls to sort graded fruit, which achieves a sorting accuracy of 79–89% at speeds ranging from 0.05 to 0.2 m/s. Deep learning-based techniques have also been employed to detect and track defective citrus fruits on a roller conveyor system for online sorting [10]. Fan et al. [11] proposed a real-time apple defect inspection method based on the YOLOv4 algorithm, which was applied to a two-lane fruit sorting machine and achieved an average detection accuracy of 93.9% at the online test assessing five fruits per second. However, the development of vision-based automated sorting technologies for sweetpotatoes is still relatively unexplored, and current research mainly focuses on classifying and measuring their physical attributes [12]. In our previous work, a machine vision system was developed for online sweetpotato grading based on quality information accumulated from multi-view images, achieving an overall grading accuracy of 91.7% for four classes of sweetpotatoes [13]. However, the vision system and grading pipeline remain to be integrated with appropriate sorting mechanisms for further evaluation.

Sorting mechanisms are essential hardware components of an automated fruit sorting system to physically segregate products according to quality grading decisions. The separation process typically takes place after the products exit from the imaging area for grading. Depending on quality grades, the products may be diverted into multiple outlets corresponding to different grades. This may involve configuring single or multiple sorting units at different positions, depending on the capacity of sorting mechanisms. Ensuring successful sorting requires products having relatively constant velocity as they approach the sorting unit, and moreover, it is desirable to position sorting units close to the vision area to minimize the unpredictable variations of the trajectory of products to be sorted. Effective sorting mechanisms must have rapid action, reliability, long lifetime, and adequate mechanical strength. A variety of sorting mechanisms have been designed and implemented for fruit sorting at packing facilities [14,15], which tend to vary with horticultural commodities.

Several studies utilized the air jet principle for sorting fruits, where an air ejector fires a pulse of air to propel the item toward its designated destination [16,17,18]. This type of sorter may not be efficient for sorting large or heavy fruit such as sweetpotatoes. Moreover, the irregular shapes of sweetpotatoes can lead to inconsistent responses to air jets, causing unpredictable movements. A cup-based sorting system is commonly used for fruit sorting, where individual cups that are mounted on a conveyor belt or chain are designed to securely hold a single piece of fruit, and as the fruit approaches its designated separation point, the associated cup will be tilted or opened, effectively releasing the fruit into the intended area [9,19]. However, the cup-based sorting system, if applied to sweetpotatoes, would face challenges. The diversity of sizes and irregular shapes of sweetpotatoes necessitates significant customization of the cups and sorting mechanisms, which may be infeasible given the system complexity and the cost of specialized hardware. Recently, a compact paddle sorter that employs an electric rotary solenoid with its shaft attached to the paddle operating in open and closed positions was proposed for in-orchard sorting of apples [20,21]. However, the paddle sorter was only applied to two-grade sorting, and without major adaptations, it would not be effective for sweetpotatoes, which require a higher torque for deflection. Hence, there is a need for new efforts in the design and development of dedicated sorting mechanisms that have the adaptability to handle the shape variability of sweetpotatoes without compromising sorting speed or accuracy.

This study, therefore, aimed to manufacture and evaluate sorting mechanisms in conjunction with a machine vision unit for automated sweetpotato sorting. Specific objectives were three-fold, i.e., to (1) present the manufacturing features of three different pneumatically powered sorting mechanisms, (2) conduct a preliminary evaluation of their performance in terms of sorting accuracy and repeatability at varied conveyor speeds, and (3) integrate the best-performing sorter with a machine vision-based grading unit and assess the performance of the integrated system in three-grade sweetpotato sorting.

2. Materials and Methods

2.1. Sorting System

2.1.1. Sorter Manufacture

Three different simple mechanisms, as schematically shown in Figure 1, were manufactured and prototyped for sweetpotato sorting. Each of the prototypes comprised a pneumatically powered actuator and a customized, 3D-printed pushing paddle attached to the rod or shaft of the actuator.

Table 1. Specifications of sorting mechanisms. L and W denote length and width, respectively.

	Actuator Type	3D-Printed Paddle Size	Stroke or Swivel Angle
Sorter 1	Linear air cylinder	220 × 80 mm (L × H)	80 mm
Sorter 2	Linear air cylinder	180 × 110 mm (L × H)	200 mm
Sorter 3	Rotary actuator	170 × 100 mm (L × H)	90°

summarizes the basic specifications of the three sorting mechanisms, where the stroke length indicates the maximum linear movement of the paddle, and the swivel angle represents the range of rotary motion in degrees that the paddle can move to deflect the sweetpotatoes.

Sorter 1 employed a linear air cylinder with its driving piston pivotally connecting to a base-fixed linkage to produce rotary action, acting as an ejection finger [20], which enables the paddle to swiftly enter and exit the product stream, deflecting targeted items. This mechanism was originally used in a commercial potato sorter but was adapted for sweetpotato sorting in this study. Such finger-type sorters have been widely used in commercial fruit packing lines for removing inferior products or foreign materials. With better structure simplicity, Sorter 2 consists of a linear air cylinder with its rod directly and fixedly attached to the paddle. Such manufacturing produces a straightforward linear trajectory for the paddle, requiring minimal mechanical complexity. The direct attachment ensures robust and consistent performance with minimal maintenance requirements. Sorter 3 incorporates a rotary pneumatic actuator with the paddle fixed to the shaft, enabling rotational movements. The rotary actuator provides large angular displacement, allowing the paddle to effectively divert sweetpotatoes to their respective destinations.

Each single mechanism, upon being actuated, was responsible for diverting the sweetpotato to one grade-specific destination. To produce three-grade sorting in this study, two identical units for each mechanism were thus constructed for evaluation, which were positioned adjacently on the same side of a belt conveyor belt (Figure 1) with 20 cm spacing to facilitate accurate coordination. The actuator in the sorting mechanisms was controlled individually via a 12 V pneumatic solenoid valve (with a maximum response of 4 cycles per second). An air compressor with a maximum pressure of 150 PSI (pounds per square inch) was used for powering the actuators.

The three mechanisms are simple to construct, cost-effective, and easy to operate. The compactness of the overall setup allows for easy integration into existing belt conveyor lines, even in limited space. The pneumatic actuators provide adequate force, which can be adjusted by controlling the operating air pressure, to separate sweetpotatoes out of the product flow. The rapid response of the pneumatic mechanisms permits high-speed online sorting of sweetpotatoes to meet practical needs. It is apparent that the 3D-printed paddle used in the sorters is not limited to the size used (Table 1) and can vary as needed. The 3D-printed parts, made of polylactic acid (PLA), which were strong enough for experimentation and offered the advantage of easy prototyping, were utilized in this study. However, for enhanced strength and durability, particularly for long-term use, more experimentation would be needed to optimize the material through methods such as increasing infill density and structural reinforcement. Machining appropriate metal plates for the sorters could also achieve better strength and durability.

2.1.2. Sorter Activation

The state of the sorting mechanism was controlled via an Arduino board with a relay connecting to the pneumatic valve for automated sorting. Prior to sorting, grading was achieved by a roller conveyor-based machine vision unit [13]. As conceptually illustrated in Figure 2, as the sweetpotato exits from the vision area and leaves the roller conveyor, it flows down to a belt conveyor via a transitional roll-down ramp that connects the roller and belt conveyors, and a deceleration curtain made of vertically hanging belting stabilizes the sweetpotato’s movement, minimizing sample rebound. Upon reaching the belt conveyor with an initial speed of nearly zero, the sweetpotato is carried out to flow forward at the speed of the belt conveyor and pending for sorting. To time sorting mechanisms for actuation, an infrared (IR) distance sensor was installed at the beginning of the belt conveyor to detect the incoming sample, based on which empirically determined delays were implemented for triggering corresponding sorting units at the right time. To minimize timing uncertainties, the sorter mechanisms were located close to the IR sensor.

In the present design scheme for three-grade sorting, the sorting units, regardless of specific mechanisms, are normally in the retracted position to allow high-quality sweetpotatoes to flow down undisturbed on the belt conveyor (Figure 2 left). When low-quality products are detected, the first sorting unit (closer to the roller conveyor) is activated to extend its paddle to enter the product stream and divert the detected item (Figure 2 right); the second sorting unit is actuated for sorting intermediate-quality sweetpotatoes.

2.2. Experimentation

Evaluation of Sorting Mechanisms

Figure 3 shows the pipeline for evaluating the manufactured sorting mechanisms. A collection of 120 sweetpotato storage roots of varied sizes and shapes purchased from local grocery stores were used for the experiment. Each sample underwent thorough cleaning and measurements for its weight (in grams) and size (including length and width), as in our prior work [13]. Figure 4 shows the distribution of sweetpotato samples used for sorting mechanism evaluation. The sample weight ranged from 191.5 g to 612.3 g, with an average weight of 286.7 g. Their width ranged from 49.8 mm to 88.7 mm, averaging 64.6 mm, while their length ranged from 130.1 mm to 246.2 mm, with an average of 162.8 mm. According to the U.S. sweetpotato grade standards [21], the length of marketable sweetpotatoes is in the range of 3 to 9 inches (76.2–228.6 mm), and the width ranges from 1–3/4 to 3–1/4 inches (44.45–82.55 mm). The selected sweetpotatoes fully covered these size ranges, as illustrated in Figure 4. Such diversity of samples is important for developing a robust sorting system for processing sweetpotatoes across a range of physical characteristics. Photographic records were made for each sample as references.

Sweetpotatoes were randomized into six groups of 20 samples each, representing diverse sizes and shapes, and placed individually on the roller conveyor (Figure 2) to prevent sample interference. The speed of the roller and belt conveyors was monitored using a Hall Effect magnetic sensor, and their speed was maintained to be the same during experiments. Three conveyor speeds, i.e., 4 cm/s, 8 cm/s, and 12 cm/s, were experimented with to examine the effect of speeds on sorting performance. The pressure of the air compressor was set to 25 PSI, which was found adequate to deliver the force needed for deflecting sweetpotatoes. The time delay, following the receipt of the IR sensor signal, was empirically set to 1200 ms, 600 ms, and 300 ms for actuating the first sorting unit on the belt conveyor, and it was set to 1600 ms, 1200 ms, and 800 ms for actuating the second sorting unit at the conveyor speed of 4 cm/s, 8 cm/s, and 12 cm/s, respectively, based on preliminary testing. Moreover, to examine the sorting repeatability, each of the 120 samples was subjected to three replicated tests at each speed.

This experiment focused on evaluating the sorting mechanisms, regardless of grading decisions. The grading results by the machine vision unit [13] were not utilized for actuating sorters in the experiment. Instead, at a given conveyor speed, the first sorting unit was programmed purposely to be active, while the second one was kept inactive to simulate the sorting of low-quality sweetpotatoes. After running one batch of samples, the first sorting actuator was kept inactive, while the second one became active to simulate the separation of the intermediate-quality sweetpotatoes. For performance evaluation, the following criterion was applied to define the success or failure of a given sorting event: a sweetpotato successfully ejected by the pushing paddle was marked as “1” (success), while any sweetpotato that was not touched or deflected by the paddle was marked as “0” (failure). The sorting experiment with three replications was video-recorded to facilitate subsequent performance analysis. Figure 5 shows the sorting processes of the three mechanisms.

Two performance metrics, including sorting accuracy and repeatability, were employed for evaluating the sorting mechanisms at each conveyor speed. Sorting accuracy, indicating the sorter’s ability to successfully eject the detected sweetpotato out of the product stream, was calculated as the ratio of the number of successfully sorted sweetpotatoes to the number of testing samples (i.e., 360 for three replications here) at a specific conveyor speed. Sorting repeatability reflects the sorter’s ability to produce consistent results over repeated testing for the same sweetpotato. Given three replications for each sample in this study, if the three sorting results were the same, the sorting was considered repeated; otherwise, it would be unrepeated. Hence, the repeatability metric was calculated to be the ratio of the number of repeated sporting events over three replications to the number (i.e., 120) of the sweetpotato samples at a specific conveyor speed. In addition to the metrics at each conveyor speed, the overall accuracy was calculated by aggregating the successes across the three conveyor speeds, and likewise, the overall repeatability was calculated by summing the consistent results across all three speeds, with the formulae given as follows:

$$\mathrm{SorterTotalAccuracy}=(N_{successful}/1080)\times 100\%$$

(1)

$$\mathrm{SorterTotalRepeatabilit}y=(N_{repeated}/360)\times 100\%$$

(2)

where N_successful is the count of successful sorting events at all speeds, and 1080 is the number of sorting attempts (120 samples × 3 replications × 3 conveyor speeds); N_repeated is the total number of repeated or consistent sorting events summed across all speeds, and 360 is the number of repeated tests (120 samples × 3 conveyor speeds).

2.3. Integrated System Evaluation

After the sorting mechanism evaluation, the best-performing sorter identified was integrated with a roller conveyor-based machine vision system developed in our prior work [13] to further assess the sorter’s performance in response to vision-based automated grading. A detailed description of the development and evaluation of the machine vision system and the associated computer algorithm of automated sweetpotato grading is described in our prior work [13,22]. Briefly, the machine vision system acquires and analyzes multi-view images of sweetpotatoes using a deep learning model-based algorithm pipeline while moving and rotating on a roller conveyor, tracks individual samples, determines their quality conditions (size and surface defects) for each view, and eventually grades each sample into three grades (i.e., “Premium”, “Good”, and “Fair”) by accumulating quality information from multiple views of the sample. Then, the graded sweetpotatoes, after moving onto the belt conveyor (Figure 2), were separated into the corresponding catching bins by the sorting mechanisms.

A separate set of 96 sweetpotato samples of varied quality conditions were collected for testing the sorter integration. These samples were first manually graded based on visual inspection of surface defect conditions and size (width and length) measurements [13], complying with the U.S. quality standards of sweetpotatoes [21] and resulting in 23 “Premium” (high-quality), 22 “Good” (intermediate-quality), and 51 “Fair” (low-quality) samples. Then, an online experiment was conducted at the same three conveyor speeds as in the sorting mechanism evaluation, where each sweetpotato was individually loaded onto a roller conveyor for automated grading and sorting. The grading accuracy was obtained by comparing the vision-based grading result with the ground-truth grade of each sweetpotato [13]. The sweetpotatoes sorted into the correct bins were counted to calculate the sorting accuracy. The impact of the sorter was assessed by examining the difference between the grading and sorting accuracies at each conveyor speed. The overall grading and sorting accuracies were determined as follows:

$$\mathrm{OverallGradingAccuracy}=(N_{graded}/96)\times 100\%$$

(3)

$$\mathrm{OverallSortingAccuracy}=(N_{sorted}/96)\times 100\%$$

(4)

where N_graded and N_sorted represent the total numbers of accurately graded and sorted samples, respectively.

3. Results and Discussion

3.1. Sorter Evaluation

The accuracy of the three sorting mechanisms at different conveyor speeds was obtained in three repeated experiments. The mean accuracy with standard error (SE) at each conveyor speed is depicted in Figure 6a. The sorting repeatability, which was calculated based on the consistency of the sorting results for the same samples over three replications, is shown in Figure 6b.

For Sorter 1, the mean accuracy values at the conveyor speeds of 4, 8, and 12 cm/s were 95.8%, 94.4%, and 93.3%, with the corresponding SE of 2.23%, 4.18%, and 3.66%, respectively. The large error bars suggest that Sorter 1’s performance is highly variable, and the relatively high SEs at 8–12 cm/s indicated reduced stability at higher speeds. Importantly, the range of the error bars for Sorter 1 does not overlap with those of Sorter 2 and Sorter 3, indicating a statistically significant difference in performance. Sorter 2 shows more stable performance with improved mean accuracies of 98.9%, 98.0%, and 96.9% at three conveyor speeds and lower standard deviations of 1.28%, 0.46%, and 1.27%, respectively. The relatively small error bars indicate consistent performance and lower variability in the accuracy measurements. However, the overlap between the error bars of Sorter 2 and Sorter 3 suggests that their performance differences might not be statistically significant. Sorter 3 also delivered comparably decent performance with mean accuracy of 98.3%, 97.5%, and 97.2%, and the corresponding SE of 0.85%, 1.44%, and 1.47%, respectively. In general, Sorter 2 consistently showed the highest performance, followed by Sorter 1 and Sorter 3. However, at the highest speed of 12 cm/s, Sorter 3 slightly outperformed Sorter 1 by 0.2%, but the changes in accuracy between the sorters at the same speed were not statistically significant. Furthermore, the repeatability trends reveal that the performance of Sorter 1 declined at higher speeds, from 90.8% at 4 cm/s to 85.8% at 12 cm/s. In contrast, Sorter 2 exhibited excellent repeatability, ranging from 97.8% to 96.5%. Sorter 3 also showed strong repeatability from 97.8% to 93.3%, despite a declining trend with the conveyor speed. These results demonstrate that Sorter 2 is the most reliable across different speeds, followed by Sorter 3, while Sorter 1 still requires further optimization to ensure consistent performance at higher speeds.

The average accuracy and repeatability across three speeds were 94.9% and 88% for Sorter 1 (finger type), 98% and 96.8% for Sorter 2 (linear air cylinder), and 97.9% and 95.6% for Sorter 3 (rotary actuator), respectively. Sorter 2 and Sorter 3 consistently outperformed Sorter 1 (finger type) regarding both metrics. The two sorting mechanisms demonstrated good potential to reliably sort sweetpotatoes. They appeared to be on par with each other in terms of sorting accuracy, but Sorter 2 showed noticeably better repeatability than Sorter 3 at the fastest testing speed of 12 cm/s. Hence, Sorter 2 would be integrated with the machine vision system for further testing as described below.

As noted above, the three sorting mechanisms tended to show declining performance with the increase in the conveyor speed (Figure 6), except that the decrease in accuracy between different levels of speeds (4 cm/s vs. 8 cm/s and 8 cm/s vs. 12 cm/s) was not statistically significant for specific sorters. This downward trend is expected for general automated sorting systems since increased speeds add to the challenges with the precision of mechanical movements and sensor responsiveness. Here, in sorting sweetpotatoes, the inertia of the products with zero speed relative to the belt conveyor could have pronounced resistance to changes in motion at higher speeds, potentially causing unpredictable responses to the sorting mechanisms. More importantly, the great shape variability and irregularity of sweetpotatoes could have complicated the sorting process at higher conveyor speeds. Further refinements of sorter configurations and settings, such as adjusting air pressure and the size of the pushing paddle, could help enhance sorting performance in different operational conditions.

3.2. Integration System Performance

Table 2. The overall grading and sorting accuracy of the integrated system with Sorter 2.

Conveyor Speed	Overall Grading Accuracy	Overall Sorting Accuracy	Sorter Error
4 cm/s	97.9%	97.9%	0%
8 cm/s	95.8%	94.8%	1%
12 cm/s	94.8%	93.8%	1%

summarizes the overall performance of the integrated system with Sorter 2 in grading and sorting three grades of sweetpotatoes. Detailed descriptions of the machine vision system are given in our previous work [13]. At a conveyor speed of 4 cm/s, the same accuracy of 97.9% was obtained for both grading and sorting, indicating that the sorting process introduced no errors. As the conveyor speed increased to 8 cm/s, the grading and sorting accuracies went slightly down to 95.8% and 94.8%, respectively. Their differences indicate an error of 1% due to the sorting process. The same sorter error was observed at the highest speed of 12 cm/s, with the grading and sorting accuracies of 94.8% and 93.8%, respectively. Both grading and sorting performance declined with the conveyor speed, which was because of the increased difficulty in capturing and analyzing images for sweetpotato quality detection at higher speeds [13,22], but the low sorter error of 0–1% confirmed that Sorter 2 was effectively integrated with the machine vision grading system, maintaining high accuracy at different conveyor speeds. The reported errors of up to 1% were primarily attributed to timing inaccuracies in the sorter actuation and misalignment caused by variations in sweetpotato shapes. To minimize these errors in the future, efforts are still needed to implement more precise timing mechanisms to improve synchronization between the vision system and sorter actuation, develop adaptive algorithms to account for variations in sweetpotato shapes and sizes, and fine-tune the mechanical settings of the sorters, such as air pressure and paddle size, to ensure more consistent deflection of sweetpotatoes. Overall, these results agreed well with the accuracy metrics of Sorter 2 when tested without system integration.

4. Discussion

Sorting mechanisms are important to implementing the automated grading and sorting of sweetpotatoes at packing facilities. Despite automated sorters available for a diversity of horticultural products, there have been few automated sorting technologies specifically developed for sweetpotatoes. Three experimental sorter prototypes were designed and evaluated in this study, which are simple in structure, cost-effective, and easy to operate and integrate with vision systems. The use of pneumatic actuators, which are common in commercial sorting systems, offers rapid responses for high-speed sorting as well as the flexibility of adjusting ejection force. The paddle in the sorters can be readily modified via 3D printing to adapt to different shapes and sizes of sweetpotatoes, although it is susceptible to wear depending on material strength. Both Sorter 2 and Sorter 3 showed decent accuracy and repeatability. Sorter 2, constructed simply with a linear air cylinder attached to a paddle, has minimal mechanical complexity and a line trajectory that allows for precise deflection of sweetpotatoes. The rotary actuator of Sorter 3, with a relatively large swivel angle (90°), also enables the paddle to reliably divert sweetpotatoes to their respective destinations.

The performance of the sorting mechanisms described in this study is encouraging compared with the sorters reported in previous studies, although direct comparisons may not be made given different applications and methodologies. Heinemann et al. [23] presented a machine vision-based automated inspection station for potatoes, in which the sorting unit consisted of a sheet metal incline with three trap doors controlled by electrical solenoids, achieving sorting accuracies of 77–88% for three grades of potatoes according to size and shape at speeds of 3 items/min. Jarimopas et al. [24] employed a pneumatic sorter similar to Sorter 2 of this study in conjunction with a machine vision system for sorting sweet tamarind, achieving a shape sorting efficiency of 96.0% at a belt speed of 13.34 m/min. For sorting small items, Blasco et al. [25] utilized air nozzles for sorting satsuma segments based on computer vision and morphological features, which achieved a classification accuracy of 93.2% for sound segments and operated at a conveyor belt speed of 600 mm/s. Sofu et al. [9] reported a magnet trigger sorter for sorting apples and obtained accuracies between 79% and 89% at speeds ranging from 0.05 to 0.2 m/s. Zhang et al. [26] developed a sorting system for Panax notoginseng taproots, which used a jet mouth sorter activated pneumatically and achieved an average sorting accuracy of 77% at a conveyor belt speed of 1.55 m/s.

There are still limitations in this study, which remain to be addressed in future work. The constructed sorting mechanisms are considered an experimental prototype for proof-of-concept rather than for practical usage. The primary concerns include the wearing of pneumatic actuators and 3D-printed paddles, potential air leakage, and the need for regular calibration to maintain accuracy, which all require regular maintenance. To further enhance reliability, it is worth exploring the use of more durable materials for critical components and implementing redundant systems to minimize downtime during maintenance. Moreover, low conveyor speeds (4–12 cm/s) were implemented in sorter testing experiments to provide an initial assessment of the sorting mechanisms’ performance due to the speed constraint of the gearmotors used in the roller and belt conveyors. An upgraded conveyor system will be used to perform extensive sorting tests at more practically meaningful speeds (e.g., 30–50 cm/s). Limited by the low conveyor speeds, relatively large time delays (0.8–1.6 s) were used for actuating the second sorting units (Figure 2); shorter time delays would be necessary at higher conveyor speeds. Although sorting mechanisms were integrated with the vision system, sorted sweetpotatoes were manually caught from the belt conveyor and dropped into the receiving bins. Dedicated catching channels have yet to be designed and integrated to facilitate receiving segregated products while inflicting no damage, which are an important component for a full-fledged sorting system for industrial applications. It is also important to point out the relatively small number of sweetpotatoes used in this study, which may not fully represent the variability of sweetpotatoes encountered at real-world packing facilities. A substantially expanded set of diverse, freshly harvested samples before packing line operations will be collected and used in future performance validation studies. Our experiments have shown that the shape and size variability of sweetpotatoes impact the sorting process. Irregular shapes and varying sizes can cause inconsistencies in how sweetpotatoes are transported and rotated on the conveyor belt, potentially leading to misalignment for target deflection by the sorting mechanisms.

In this study, sorter configurations were adjusted to handle the diversity, including fine-tuning the air pressure for pneumatic sorters and the size of the pushing paddles to ensure effective sorting regardless of the sweetpotato’s dimensions. Future efforts to improve the sorters in this study will focus on optimizing the operational settings (e.g., air pressure) of the pneumatic actuators to enhance sorting reliability and efficiency. The scheme of using two sorting units to realize three-grade sorting is still suboptimal in terms of space requirements and time delays for the second unit. A more elegant solution is to have a single unit that is capable of sorting sweetpotatoes into three or more grades. Multi-position actuators can be potentially employed for this purpose and are, hence, worthy of further investigation. More dedicated efforts are needed to develop a full-fledged sweetpotato sorting system that integrates functional modules for sample feeding, grading, sorting, catching, and bin filling before performing system-level performance studies in commercial settings. In recent years, soft robotics guided by 3D computer vision has gained attention for food sorting [27]. Soft grippers mounted on a delta robot manipulator are commercially available for handling irregular agricultural commodities such as sweetpotatoes [28], offering a potential solution to dexterously sort the products into multiple grades if costs are no longer a concern. Compared to fixed sorting mechanisms, soft grippers promise to emulate hand sorting by conforming to the shape of targeted items and applying adequate forces for handling.

5. Conclusions

This study presents new engineering and performance evaluations of sorting mechanisms for sweetpotatoes. Three different pneumatically powered mechanisms were prototyped and tested for online sorting at varied conveyor speeds of 4–12 cm/s. Sorter 2, which employed a linear air cylinder for product separation, achieved the overall best performance, with average sorting accuracy and repeatability of 98% and 96.8%, respectively. The rotary actuator-based Sorter 3 also delivered comparable performance, with average accuracy and repeatability of 97.9% and 95.6%, respectively. Sorter 2 was selected for further integration with a machine vision system for sorting three grades of sweetpotatoes based on automated quality detection and grading. In the integrated system, the sorting process yielded errors of up to 1% at conveyor speeds of 8–12 cm/s, confirming the efficacy of the manufactured sorting mechanism. The performance of sorting mechanisms, when evaluated either in a standalone or integrated configuration, tended to decline as the conveyor speed increased. Despite the simplicity and high performance of the sorting mechanisms of this work, future research is needed to enhance sorter performance and conduct extensive tests at higher conveyor speeds toward industrial application.