Robotic Button Mushroom Harvesting Systems: A Review of Design, Mechanism, and Future Directions

Abstract

The global demand for button mushrooms has surged in recent years, driven by their health benefits, creating a significant challenge for the mushroom industry in meeting this increasing demand. The increasing reliance on human labor, which is becoming unsustainable due to labor shortages and rising wage costs, highlights the urgent need for automated harvesting solutions. This review examines the integration of automated systems in button mushroom harvesting, delving into the key components such as robots, mechanisms, machine elements, programming, and algorithms. It offers a thorough analysis of the performance, design, operational mechanisms, and advantages and limitations of robotic systems, comparing the different methods employed in automated harvesting. This paper compares the performance of all the mushroom harvesters, including the commercially available ones with manual harvesting, and identifies their potential and limitations. The commercial harvesters are shown to pick 2000 mushrooms per hour on average, which is similar to how much a skilled worker picks at the same time. However, commercial automation harvesting has a relatively low success rate, high initial cost, high operating cost, and energy consumption, identifying areas for future research and challenges. This paper serves as a valuable resource for researchers and industry professionals striving to advance automated harvesting technology and improve its efficiency in meeting the rising demand for button mushrooms.

1. Introduction

The global demand for agricultural products has rapidly increased due to a significant growth in population. By 2050, the world population will rise by 20%, posing a critical challenge to meeting agricultural and food demands [1,2]. Compounding this issue is the declining human involvement in agriculture, driven by factors such as rising labor costs, global pandemics like COVID-19, an aging workforce, and high-paying alternative careers that make agriculture less attractive [3,4]. In the US, the percentage of people undertaking farming as a main profession has decreased by three times in the last 60 years and has been declining by 10% every year while the population has more than doubled. Moreover, the increase in labor expenses has caused a decrease in the profitability of agricultural products [5,6,7]. These trends have left the agriculture industry in a vulnerable state, making it clear that reliance on human labor alone is no longer feasible [8]. To address these pressing challenges, there is a crucial need for the widespread adaptation of automation, which must be both efficient and sustainable [5,9]. Automation systems in agriculture involve the use of control systems, computers, and information technology to operate machinery with minimal human intervention. This can take the form of a fully automated system where machines perform all assigned tasks, either independently or in collaboration with humans and other machines.

The use of machines and automation in agriculture began after the Industrial Revolution in the 19th century. With the advent of computers and robotics in the latter half of the 20th century, automation in agriculture rapidly expanded. In the present time, the use of machine learning and artificial intelligence (AI) has improved the performance of machines by increasing the accuracy of object detection, monitoring, quality control, and much more [10,11]. This growth is driven by the reliability, efficiency, and long-term sustainability that automation offers. Today, various automation systems are employed in the agricultural industry, with some mechanical harvesters significantly boosting harvesting efficiency compared to manual labor [12].

Mushrooms are low in calories, fat, cholesterol, and sodium levels, yet they provide rich nutritional benefits, including high levels of dietary fiber, protein, vitamins, and essential minerals like copper, manganese, selenium, and zinc [13,14,15]. Both the mushroom caps and stems offer these nutritional advantages, with dried mushroom stem powder also being used as animal feed [16]. Button mushrooms, the most popular species, are known for their exceptional nutritional value, texture, and medicinal benefits, such as anti-aging effects [17], immune support [18], and potential anticancer properties [19,20]. In 2022, button mushrooms accounted for 61.8% of the global edible mushroom production [21,22]. The global market for button mushrooms was valued at USD 25 billion in 2023, with a 6.6% CAGR (compound annual growth rate) [23,24]. Traditionally, the mushroom industry relies heavily on human labor for tasks from pre-treatment to harvesting. However, since the early 1990s, researchers have explored automation in mushroom harvesting [25]. Pioneers like Reed, Noble, and Tillett developed early robotic systems for button mushrooms [26,27,28,29], and by 2024, advancements have significantly improved the design, operation, and performance of these automated harvesters.

Automated harvesting of button mushrooms involves several steps, such as identifying the target mushroom’s location, guiding the manipulator and end-effector to the mushroom, picking it, and transporting it to the post-harvest site. These tasks are performed by various robotic components and mechanisms aided by computer vision, sensors, and control systems (Figure 1). Due to recent technological advancements and the complexity of harvesters, many studies have focused on optimizing individual components to develop the most efficient system. This review analyzes the latest developments in automated mushroom harvesting, examining the properties and applications of different components while also discussing their potential and associated challenges.

2. Methodology

This review conducted a thorough analysis of the existing harvesting technologies for various fruits and vegetables, with a particular focus on the automated systems relevant to button mushroom harvesting. The keywords used to find the relevant works included automation in agriculture, mushroom harvesting, computer vision in agriculture, AI in agriculture, Agaricus bisporus, button mushroom production and life cycle, automated harvesting of button mushroom, etc. To focus our work on button mushroom harvesting, only the works that were directly related to button mushroom harvesting were considered for Section 3 and Section 4. The study covered work from 1990 to 2024, drawing data and information from a variety of sources, including Web of Science, academic theses, scientific reports, online articles, and videos, with most referenced from Google Scholar. Out of 95 cited sources, 77 were recent works published between 2018–2024. The number of references used for mechanical systems and computer vision systems in Section 3 and Section 4 are 42 and 33, respectively, with some common sources. Additionally, the analyses in the Discussion and Potential and Challenges sections are informed by our visits to existing mushroom farms. The information from the mushroom farm visit includes the performance of human picking, the mushroom growing environment, the mushroom life cycle, and the sizes of Dutch shelves. These data directly influence the performance of the harvesters; therefore, they have been used in the Discussion, Potential and Challenges, and Conclusion sections.

The review is divided into two main sections: the Mechanical System and the Computer Vision System. The Mechanical System section explores the physical components of automated harvesters, such as robotic arms, end-effectors, and conveyors. The Computer Vision System section delves into the visual sensing technologies used in these systems, including cameras and the programming that enables precise mushroom identification and selection. This approach provides a thorough understanding of both the mechanical and sensory aspects of automated mushroom harvesting systems, offering a comprehensive analysis of the current technologies and their future potential.

3. Mechanical System

3.1. Robot Manipulators

Robot manipulators play a crucial role in the harvesting system by facilitating the movement of the end-effectors during the picking and transportation of mushrooms. These manipulators are equipped with end-effectors and, in many cases, with the control system and sensors for the measurements of data and motion control. While various types of manipulators are used in harvesting fruits and vegetables, the unique morphology of mushrooms and their growing environment often necessitates the use of Cartesian and articulated arms.

Cartesian robotic manipulators operate with linear motion along the x-, y-, and z-axes. For mushroom harvesting, these manipulators are frequently enhanced with additional mechanisms to increase their degrees of freedom (DOFs) and better support the end-effectors. A fundamental example is a three-DOF Cartesian manipulator, which typically features an end-effector attached to the z-axis, with two sliders facilitating movement along the x- and y-axes. This straightforward mechanism is especially effective for use with Dutch shelves, making it a popular choice for both commercial harvesting and research applications [30,31,32,33,34].

The Champi-ON project, for instance, developed an automated mushroom harvesting system tailored for Dutch shelving farms but was adaptable to other cultivation methods. The system integrates a gripper, vision system, suction mechanism, stem cutting, and container shaking system into a Cartesian trolley, ensuring precise handling and minimal damage to the mushrooms [35]. In experiments, researchers like Jia et al. [36] and Hu et al. [37] utilized Cartesian manipulators with multiple sliders to either test the end-effectors or optimize the picking path, adding more end-effectors and improving their performance through proper path planning. These multi-linked Cartesian manipulators were also equipped to carry additional accessories such as cameras and sensors.

Yang et al. modified a Cartesian manipulator to include additional vertical and rotational DOFs, creating a mechanism suitable for Dutch shelves [38]. Similarly, Huang et al. employed a four-DOF positioning end-effector, combining a three-DOF Cartesian system with an additional rotational mechanism, which enhances picking performances through precise positioning [39]. Mycionics also utilized a Cartesian manipulator with a revolute arm for commercial harvesting, enabling efficient picking and greater accessibility to other harvesting components. These manipulators are typically mounted on shelves, with wheels supporting their locomotion along specially designed slots, allowing them to cover the entire shelf workspace [40]. Dorna Robotics has used a combination of Cartesian and multi-revolute arms (a six-DOF robot) to design a commercial harvester that can also be transported between the shelves and have a picking performance close to that of skilled human pickers. Despite the claim that the harvester can be transported easily, the harvester has to be moved manually at the current stage [41].

A three-DOF articulated arm for mushroom picking was developed by the University of Warwick, while the Vineland Research and Innovation Center in Ontario created a multi-DOF articulated arm based on portable beds [30,42]. In commercial and high-volume mushroom harvesting, SCARA arms are frequently used as manipulators. A SCARA robot employed in mushroom harvesting mostly moves on the XY (parallel to the ground) plane and translates into the z-axis. TechBrew Robotics, for instance, has integrated the SCARA arm to speed up mushroom harvesting on Dutch shelves [43]. Similarly, 4AG Robotics recently developed an automated harvesting system with a SCARA arm [44]. These SCARA arms have been shown to be faster at picking and have easy access to the collector units installed outside. At Nanjing Agriculture University, two SCARA manipulators were employed to pick mushrooms on both sides of the shelves [45]. The independent operation of these two arms significantly increased the picking speed, a key factor in large-scale mushroom industries.

Although articulated arms have not been widely tested in commercial settings, they have been used for various mushroom harvesting analyses and laboratory tests. Recchia et al. utilized the Omron TM5-700 Cobot to evaluate the picking performance of a developed gripper [46]. Additionally, Recchia explored the potential of using similar articulated arms in fully automated harvesting [47]. Mohanan et al. designed a SCARA manipulator as a harvesting robot using a probabilistic road map and an inverse kinematics approach [48]. Rowley developed a laboratory-based model with a six-DOF anthropomorphic flexible robot to test harvesting in a lab environment [49]. However, these manipulators have yet to be tested on multi-storied shelves, and the practical effectiveness of some—such as those developed by Mohanan and Rowley—remains unproven [48,49]. Various manipulators used in automated mushroom harvesting are illustrated in Figure 2. Cartesian manipulators generally offer a large workspace within Dutch shelves compared to articulated manipulators like SCARA. However, when dealing with multiple shelves, the SCARA arm provides greater flexibility due to its prismatic joint, which allows vertical movement between shelves. Additionally, articulated arms can easily extend beyond the shelves to interact with other components as needed, whereas Cartesian manipulators require an additional transport mechanism to perform tasks outside the shelves. That said, articulated arms themselves often rely on a Cartesian cart or mobile robot to move along the shelves.

3.2. End-Effectors

An end-effector is a tool attached to the wrist of a manipulator, enabling direct interaction between the manipulator and the mushroom [50]. In the case of button mushrooms, the picking action is typically carried out by either grasping the mushroom cap or suctioning it at a contact point on the cap. As a result, most of the end-effectors are either graspers, grippers, or vacuum suction cups. To enhance their functionality, end-effectors are often equipped with additional mechanisms that allow them to perform actions like bending, twisting, or lifting, which are essential for picking mushrooms. A critical consideration in the end-effector design is ensuring that they avoid causing any injury or damage to the mushroom cap [3].

Table 1 provides a list of various end-effectors used for picking button mushrooms, along with their respective actuation mechanisms. Vacuum suction cups are a commonly used type of end-effector, which creates negative pressure at the contact point with the mushroom cap to grip the mushroom from the top. Reed et al. were pioneers in automated mushroom harvesting and employed vacuum suction cup end-effectors in their work [26,27]. In the early stages of developing mushroom-picking end-effectors, this type was widely used [31,37,49,51].

In 2021, Huang et al. studied hand-picking mechanisms and incorporated these insights into their design of the positioning end-effector [52]. Yang et al. developed an end-effector that combines a motor and gear box combination with flexible suction cups [38]. These vacuum suction end-effectors are also used in commercial mushroom harvesters, such as those by TechBrew Robotics [43]. Zhao et al. designed pressure-stabilized flexible suction cups that significantly improved the picking success rate to 100% compared to other suction cups [31]. These end-effectors, with only a single point of contact, simplify picking by allowing them to grasp the mushroom cap from any exposed part, even on inclined mushrooms. The actuation mechanism for these end-effectors was pneumatic, with the pressure required to pick mushrooms ranging from 9 KPa [31] to 172 KPa [52]. The picking success rates typically ranged from 90% to 100%, with minor damages like bruises and stains appearing a few days after harvesting [52].

Graspers, in contrast to suction cups, can be either open grippers with finger-like structures or closed structures resembling tubes with internal fingers [53]. These are the most common types of end-effectors used for picking soft fruits and vegetables, as they can exert greater force when needed and securely hold objects, unlike suction cups. Many graspers used in mushroom harvesting share similarities with those designed for fruit picking in terms of their design, actuation, and working principles [54]. Unlike suction cups, which grip from the top, graspers typically grab mushroom caps from the side and can have multiple contact points.

One popular design is the three-finger gripper, inspired by the human hand’s thumb, index, and middle fingers. These bio-inspired grippers can either be soft and flexible, allowing the fingers to conform to the mushroom cap’s shape [30,55], or a hybrid gripper combining rigid and soft components [3,40,46]. Soft grippers offer advantages such as being lightweight, easy to control, adaptable to different conditions, and gentle in handling [56,57]. Mavridis et al. developed a soft gripper with a ribbed structure on the tension side and a plain structure on the compression side, which enhanced flexibility and allowed the gripper to adjust to different mushroom shapes and sizes [33]. Recchia et al. designed a hybrid three-finger gripper, demonstrating improved picking performance by incorporating various shapes and thicknesses of compression slots on the fingertips [46].

Table 1. Comparison of end-effectors used in mushroom picking.

End-Effector Type	Actuation	Picking Time [s]	Success Rate [%]	Application	References
Suction cup	Pneumatic	1.7	70–100	Commercial harvesting	[26,27,31,38,52]
Two-finger gripper	-	-	-	Commercial harvesting	[45]
Three-finger gripper	-	-	-	Commercial harvesting	[58]
Hybrid three-finger gripper	Electric motor	1.2	64–100	-	[3,46]
Closed graspers	Hyper-elastic	-	100	Organic objects	[59]
Soft three-finger gripper	Pneumatic	3–4	-	-	[55]
Conventional hand picking	Manual	3	~100	Commercial harvesting	[52]

Table 1. Comparison of end-effectors used in mushroom picking.

End-Effector Type	Actuation	Picking Time [s]	Success Rate [%]	Application	References
Suction cup	Pneumatic	1.7	70–100	Commercial harvesting	[26,27,31,38,52]
Two-finger gripper	-	-	-	Commercial harvesting	[45]
Three-finger gripper	-	-	-	Commercial harvesting	[58]
Hybrid three-finger gripper	Electric motor	1.2	64–100	-	[3,46]
Closed graspers	Hyper-elastic	-	100	Organic objects	[59]
Soft three-finger gripper	Pneumatic	3–4	-	-	[55]
Conventional hand picking	Manual	3	~100	Commercial harvesting	[52]

In commercial harvesting, Mycionics in Canada employs a three-finger gripper with asymmetric finger placement, resembling human fingers, where the index and middle fingers are closer together and the thumb is farther apart [40,58]. Koirala et al. analyzed mushrooms of different shapes and sizes to design a compact hybrid gripper suitable for a wide range of mushroom sizes [3]. The hybrid grippers mentioned in this review featured soft pads on the tips made from materials like PDMS [59], Ecoflex silicone rubber [3], or thermoplastic polyurethane (TPU) [46,47]. These soft pads cushion the gripping force, reducing damage to the mushroom caps while increasing the picking success rate. Galley et al. designed a hyper-elastic actuating closed grasper resembling a tube with three PDMS fingertip-like structures inside, specifically for picking organic objects like button mushrooms [59].

Nanjing Agricultural University developed a two-finger gripper for commercial harvesting [45]. The two fingers, positioned opposite each other, close linearly to grasp the mushrooms from both sides and can be made from soft materials to prevent injuries. Tao et al. designed a specialized end-effector with a drill mechanism for picking smaller mushrooms during bud thinning [60]. The grippers are actuated using either electric motors [3,60] or pneumatic systems [55,59]. Dorna Robotics has incorporated both suction cups (two) and two-finger grippers in the development of commercial harvesters. The gripper is motor-controlled and adjusts according to the size of the mushroom. The suction cups can approach mushrooms from different angles, which makes the harvesting of inclined mushrooms more efficient [41]. Soft fingers require a relatively low gripping force, as little as 1 N, whereas other grippers may need up to 18 N. The success rate for picking fully grown mushrooms with these graspers typically ranges between 90–100%. Soft grippers cause minimal to no injuries, while other grippers may leave small bruises and dents when applying higher gripping forces. Figure 3 illustrates the end-effectors used to pick button mushrooms.

Most end-effectors have been designed and tested primarily for picking individual, full-grown mushrooms. However, button mushrooms often grow in clusters, and there are no end-effectors reported that match the success rate and efficiency seen with single mushroom picking when applied to the clusters [3,52]. A few end-effectors have attempted cluster picking, but the current progress indicates that regardless of the type, additional mechanisms like bending and twisting are necessary to improve the success rate for cluster harvesting.

3.3. Collection System

After the mushrooms are detached by the end-effectors, they are typically moved to a cutter area before being dropped at the collection site. In fruit and vegetable harvesting, cutters are often integrated with the end-effectors as a single unit since the end-effector alone might not fully detach the produce [54]. However, for button mushrooms, end-effectors are usually designed to pick either just the mushroom cap or the cap along with part of the stem. Since the stem can be contaminated with soil from the bed, it often needs to be trimmed. This trimmed stem can still be utilized for various applications [15,16].

Although there is limited literature on the detailed analysis of cutters or collection systems specifically for button mushrooms, some designs incorporate these mechanisms (Figure 4). For example, Reed et al. experimented with different blade shapes and cutting mechanisms, including horizontal blades and circular rotating blades, though the cut profiles were not always clean [26]. Huang et al. integrated a planar knife blade with a pneumatic cylinder and photoelectric sensor for mushroom detection, positioning the cutter on a Cartesian frame with the end-effector [39]. However, the performance of the cutter is not discussed in the published work.

In commercial harvesting, integrated collection units are more common than separate cutting units. On-shelf collection mechanisms, like the rotating platform used by Mycionics or conveyor belts on shelves, allow manipulators easier access and reduce travel time [38,58]. The rotating disc and some vertical collectors had some slots where the picked mushroom was dropped. These slots ensure the secure holding of mushrooms during transportation to the other units. The rotating platform also requires additional mechanisms to transfer the mushrooms from place to place. Off-the-shelf collection units—such as vertical collectors like baskets or integrated systems that handle cutting and post-harvesting—offer more flexibility by being accessible to all layers of multi-storied shelves [40,43,44,45,51]. Some systems, like the one developed by Wang et al., even include sorting mechanisms with integrated cameras for size-based sorting and collection [61]. These off-the-shelf mechanisms can be moved between locations, often requiring mobile platforms or vehicles for transport.

3.4. Mobile Platform

A mobile platform is used to move the robot between locations. In fruit and vegetable harvesting, platforms like four-wheel drives and rail carriages often transport the harvester or collected produce. However, mobile platform options for mushroom harvesting are more limited. This is mainly due to the design of Dutch shelves, which have a rectangular workspace, meaning that the target mushrooms are always on a specific horizontal plane [61,62,63,64]. Consequently, the manipulator must move along this plane, and most mobile platforms are trolley- or cart-like mechanisms made of Cartesian manipulators installed on the shelves. There are a few exceptions where mobile vehicles carry articulated manipulators. The different mechanisms used as a mobile platform in automated mushroom harvesting are shown in Figure 5.

For example, Reed et al. employed a stepper motor-driven frame to move the system [26,27], while Ghahraei et al. designed a horizontal rail system for the vertical manipulator to slide along [32]. Other systems use carts or trolleys with wheels that roll on slots in the Dutch shelves, such as round groove pulley wheels [37,40,58] or train wheels [51] that roll on I-channel profiled shelves. Mavridis et al. used V-groove wheels to move the system on T-slot profiled shelves [33]. Some systems employ metal wheels on a specific shelf, and others use moving carts with round profile wheels [36,38,60]. These cart-like mobile platforms are usually part of the Cartesian manipulator with wheels, where the platform moves in one direction [30,34,40]. These types of on-the-shelf mobile platforms are very efficient on the shelf where they are installed. However, they may need additional transport mechanisms to switch the shelves.

Off-the-shelf mobile platforms include robot chassis or vehicle-like mechanisms that move on the ground and carry manipulators or accessories like cameras for scanning [48]. For instance, TechBrew’s commercial mushroom harvester has a mobile platform that moves the entire system around the shelves while remaining outside them [43]. Ground-based vehicles are commonly used to transport the articulated arms or harvesting systems, offering flexibility to reach multiple layers and navigate between shelves [44,45]. These outside vehicles allow a single manipulator to access all areas, including inside multi-layered shelves.

3.5. Sensors

Sensors in mushroom harvesting are used to measure various parameters like force, pressure, displacement, and object detection. In addition to the robots’ built-in sensors, various other sensors are integrated with components like end-effectors, manipulators, and mobile platforms.

Table 2. Sensors that are used for different applications in mushroom harvesting.

Application (Measurement)	Sensors	References
Gripping force, load, picking parameters	FSRs, load	[3,30,38,39,46,51,52,55,60,65]
Suction pressure	Pressure	[38,39,51,52,64]
Distance, displacement	Distance, electromagnetic	[51,59]
Bending	Bending	[55]
Strain changes	Strain-resistance gauge	[66]
Angle	Inclination, attitude angle	[39,52,60]
Temperature, humidity	DHT	[67]
Object detection	LIDAR, camera	[11,66]
Grading and quality	camera	[11]

lists the different sensors used in mushroom harvesters.

Force-resistive sensors (FSRs) are commonly employed to measure the gripping force, typically installed on the fingertips of end-effectors. These sensors are calibrated with reference to the standard picking parameters, like safe gripping force, and are programmed to provide accurate force measurements [3]. These readings are often used as feedback to control the applied force, especially in open finger-like grippers where precise force control is crucial [3,30,46,55,65]. Force sensors are also integrated into end-effectors and human fingers to analyze the picking forces [39,52,55,60]. FSRs or load sensors are used to measure the picking parameters, while air pressure sensors are commonly utilized with vacuum suction end-effectors to monitor and control pressure during picking [31,38,39,51,52]. The use of FSRs or load sensor readings as feedback significantly reduces the chances of injury to the mushroom, resulting in a higher success rate.

Other sensors measure displacement and strain. For example, Galley et al. used an electromagnetic sensor to measure the displacement of the actuating surface [59]. This allowed the precise control of the end-effector during gripping and minimized the damage done to the mushroom. Bending sensors can be used, especially with the shape of adaptive fingers. Mbakop et al. incorporated bending sensors with soft grippers to assess the actuator shape in real time [55]. Zhu et al. employed strain gauge sensors to measure the strain changes during gripping [66]. Huang et al. used inclination angle sensors to measure the angle of the end-effector during bending [39,52]. Tao et al. incorporated six-axis attitude angle sensors to measure the acceleration and velocity, providing the attitude angle [60]. These angle sensors improved the picking ability of the end-effector by either aligning the end-effector with the mushroom direction or providing feedback. Zhong et al. utilized distance sensors to measure the gap between the end-effector and the mushroom cap [51], and Rowley used laser sensors to identify the 3D position of the mushroom [49]. Additionally, IoT-based monitoring systems have been developed for mushrooms using temperature, humidity, and light sensors, and LIDAR has been used for object detection and path planning in mobile robots [66,67]. For mushroom identification, various camera-based sensors are used, which are discussed in the Computer Vision section.

4. Computer Vision System

In automated mushroom harvesting, computer vision systems are crucial for mushroom detection and localization, growth monitoring, disease detection, and quality assurance. A typical computer vision system includes cameras, lighting, a communication unit, and a processing device (usually a computer). As illustrated in Figure 6, various algorithms and methods are employed to effectively utilize the vision system for accurately identifying mushroom characteristics. This section reviews the applications of vision systems in button mushroom harvesting, as explored in various studies.

Table 3. Summary of vision system used in button mushroom harvesting and growth.

Application	Camera/Sensor	Algorithm/Method	Tasks	Results	References
Identification	Video camera	Morphological target detection and image processing	Identifying and locating mushrooms inside harvester rig	86% identified	[29]
	Monochromatic camera (MC)	Image processing and morphological target detection	Detecting and locating mushrooms inside the harvester	84% located, 67% picked	[26]
				76% picked	[28]
	MC with a zoom lens			81.6% picked	[26]
	Monochromatic camera	Image processing algorithm with centroid method	Locate mushrooms in a growing field	Effective	[68,69]
		Harris corner detection with an iterative algorithm combined with a watershed algorithm	Background suppression, center detection for identification	86.3% located	[70]
		Image processing using edge detection, convex hull extraction, and Harris corner detection	Segmenting and identifying overlapping mushrooms	96% recognition accuracy	[71]
		Improved YOLOv5s model with CBAM module and Mosaic image augmentation	Detection of Agaricus bisporus in complex environment	98.8% detection accuracy	[72]
	RGB-D	Template-based approach with 3D meshes or point clouds, density clustering, and modified ICP algorithm	Detection and 3D pose estimation	Very effective	[73]
	RGB-D	Greyscale conversion, active contour, and circular Hough transform	Detection, localization, and 3D pose estimation	Effective in lab and farm settings	[74]
		Improved YOLOv2 algorithm with ResNet50	Detecting and positioning mushrooms	33.9 frames/s detection rate; high accuracy	[75]
		Recursive-YOLOv5, ASPP, improved IOU metrics	Identifying edible mushrooms	98% accuracy	[76]
		Region-based convolutional networks with LoG algorithm	Detection	92.142% detection rate	[77]
Classification and growth	Monochromatic	Image analysis based on color, shape, stem cut, and cap veil opening	Mushroom classification and quality		[78]
		Image processing algorithm	Classification based on pileus diameter		[79]
	IP67 Network camera	Novel image detection algorithm using YOLOv3 and SP algorithm	Measuring mushroom cap size and growth rate	Better accuracy than CHT	[80,81]
	Kinect RGB-D	Image processing and machine learning	Monitoring growth stages	70.93% accuracy	[82]
Quality Assessment	Color video camera	Vectorial normalization method	Disease detection based on discoloration	81% of diseases classified	[83]
	Digital webcam	Image processing with artificial neural network and fuzzy logic	Assess mushroom quality based on color, area, weight, and volume	Detection rate 95.6%	[84]
	Optical zooming digital camera	L-a-b color model and hyperspectral imaging	Distinguish damaged and undamaged mushroom based on color and browning	Highly accurate	[85]
	Android camera	MobileNetv2	Distinguish poisonous and edible mushrooms	72% confidence rate	[86]

summarizes the different vision components, along with their associated algorithms and applications.

4.1. Two-Dimensional (2D) and Three-Dimensional (3D) Vision

In mushroom harvesting, 2D vision is primarily used to identify mushrooms by assessing their size, shape, color, and texture. This information helps the harvester determine which mushrooms are ready to be picked. In contrast, 3D vision provides spatial orientation and position, allowing for precise depth perception. This is essential for robotic systems to accurately grasp and pick mushrooms without causing damage. Additionally, 3D vision aids in estimating the mushroom’s volume and ensuring the picking mechanism interacts with mushrooms at the correct angle and depth.

Both 2D and 3D camera systems are widely used in mushroom identification, monitoring, and quality assurance. Commercially available RGB cameras, including those in smartphones and digital devices [11], capture 2D images. In the past, video recorders, monochromatic cameras, and webcams were commonly used for these tasks [26,27,28,29,78,83,85]. Today, more advanced cameras like the IP67 network cameras and smartphone cameras are popular in mushroom harvesting applications [80,81,86]. For 3D vision, RGB-D cameras, which combine RGB imaging with depth measurement, are commonly used. Popular 3D cameras include Intel Realsense and Microsoft Kinect [73,82]. The quality of images captured by these cameras is significantly influenced by the lighting conditions, impacting the accuracy of image processing [27]. To address this, various lighting setups, including linear and circular sources, controlled illumination, and arrays of lights, are used. While fluorescent lights were once common, LEDs have become the preferred choice in recent years.

The selection between 2D and 3D vision systems greatly influences the accuracy of mushroom detection and localization. Studies have demonstrated that 3D vision systems generally provide higher accuracy due to their ability to capture depth information, which is crucial for precise spatial orientation. For example, Baisa et al. [74] effectively utilized low-cost RGB-D sensors for detection and 3D pose estimation, achieving high accuracy in both laboratory and farm environments. Similarly, Retsinas et al. [73] reported significant effectiveness in mushroom detection and pose estimation using a template-based approach with 3D point clouds.

Conversely, 2D vision systems, while useful for identifying surface features like color and shape, often face challenges with depth perception and occlusion, impacting their overall accuracy. Traditional 2D methods like those employed by Reed [26,27] and Tillett et al. [29] achieved detection accuracies of 84% and 86%, respectively. However, the lack of in-depth information can limit precise localization, which is essential for automated harvesting.

4.2. Mushroom Detection and Localization

In Tillett et al.’s study, an algorithm was developed to identify the size and position of mushrooms by starting from the center point and tracing the outline in a gray-level image, using a low threshold based on the expected shape. The algorithm successfully located 76% of the mushrooms, with 86% of those having fairly accurate outlines [29]. Noble et al. used a vision system to analyze the impact of factors like mushroom strain, distribution, size, growing angle, and flush number, concluding that these factors significantly affect harvester performance [28]. In 1994, Reed et al. used a monochromatic vision system with an image analysis algorithm to locate mushrooms, achieving a 57% picking success rate and identifying 84% of the mushrooms [26]. They later improved the system to 90% accuracy using a zoom lens and an algorithm based on Lambertian reflector principles [27].

Yu et al. applied an image processing algorithm using the centroid method and Fourier descriptors to locate and contour mushrooms in a growing bed [68,69]. Yang et al. introduced Harris corner detection for background suppression combined with watershed segmentation and elliptical fitting, achieving an 86.3% success rate in automated picking [70]. Anibal et al. employed a machine vision system to measure the mushroom volume [87], while another Yang et al. study used advanced image processing techniques like edge detection and convex hull extraction to achieve over 96% accuracy in identifying overlapping mushrooms, although the method was computationally intensive [71]. Chen et al. introduced an improved YOLOv5s model, enhanced with CBAM modules and Mosaic image augmentation (Figure 7), achieving a 98.8% detection accuracy and rapid processing times, outperforming other models, and showing potential for integration into harvesting robots [72]. Retsinas focused on 3D pose estimation using a template-based approach with RGB-D sensor data, proving effective despite its simplicity [73]. Baisa et al. developed an algorithm for detecting, localizing, and estimating the 3D pose of the mushrooms using low-cost RGB-D sensors, which was effective in both lab and farm settings [74].

Lin et al., in 2021, developed an automated system using an improved YOLOv2 algorithm for real-time mushroom sorting, achieving high accuracy with minimal error [75]. Wei et al. employed Recursive-YOLOv5, enhancing YOLOv5 with recursion and other improvements, achieving a 98% accuracy in identifying edible mushrooms [76]. Olpin, in 2018, compared region-based convolutional networks with the Laplacian of Gaussian algorithm, finding that RCNN outperformed the LoG algorithm and RFCN in accuracy, though with slightly longer training times [77].

4.3. Classification and Growth Monitoring

In 1994, Heinemann et al. developed a mushroom classification system using image analysis to assess the parameters, such as color, shape, stem cut, and cap veil opening, with the goal of creating an algorithm to evaluate mushroom quality [78]. Wang et al. later designed an image processing algorithm based on the watershed method, Canny operator, OR operation, and closed operation to classify mushrooms based on their pileus diameter according to industrial standards, categorizing them as small (less than 2.5 cm), medium (2.5–4.5 cm), or large (more than 4.5 cm) [79]. Their work demonstrated significant potential in determining when mushrooms are ready for harvest. Lu et al. proposed a smart mushroom measurement system using image processing and deep learning to automatically measure the mushroom’s cap size and growth rate, aiding in harvest time estimation and production management. With strong performance in mushroom localization and positioning correction, the system was suggested as having the potential for integration with greenhouse control systems to optimize growth conditions and could contribute to the development of AI in agricultural systems [80]. Lu et al. also used a novel image detection algorithm to calculate the circle diameter of mushroom caps using YOLOv3 for positioning and an SP (Score Punishment) algorithm for circle estimation. The algorithm proved robust across various conditions without needing specific parameters, outperforming the CHT (circle Hough transform) in accuracy [81]. They highlighted its potential applications in optimizing greenhouse climate control, tracking mushroom growth, and improving harvest management.

Lee et al. conducted a study monitoring the growth stages of 86 mushrooms using time-lapse imaging, identifying maturity through image processing and machine learning. By utilizing color, NIR, and depth images captured by a Kinect camera, the faster RCNN model was able to localize mushrooms, and their morphological traits were quantified. An SVM classifier achieved a 70.93% accuracy in determining crop maturity, with the potential for extension to commercial-scale mushroom farming [82,88]. Figure 8 shows the RGB image and the image with RGB depth.

4.4. Quality and Disease Detection

Vízhányó et al. developed a vectorial normalization method to improve the detection of bacterial disease-related discoloration in mushroom images. This method distinguished disease-specific color points from healthy ones within the RGB color space, effectively identifying diseased spots without mistakenly detecting healthy or naturally senescent areas and achieved acceptable results in recognizing the two diseases tested [83]. In a study by Nadim et al., image processing systems were used to assess the mushroom quality based on color, area, weight, and volume, utilizing artificial neural networks and fuzzy logic. The system evaluated 250 images across three quality categories, achieving a correct detection rate of 95.6 [84].

Arjun et al. explored the use of the L-a-b color model and hyperspectral imaging to differentiate between undamaged (UD) and mechanically damaged (D) white button mushrooms (WBM). Their key findings highlighted the ability to distinguish mushrooms based on the a-values (redness) and b-values (yellowness), with browning and firmness loss more pronounced in D mushrooms. Hyperspectral imaging, particularly at 600 nm, paired with k-nearest neighbors, proved highly effective in accurately classifying mushrooms as UD or D, even on the day of damage induction, indicating potential for online mushroom classification systems [85]. Jacob et al. developed an Android app designed to classify mushrooms as poisonous or edible alongside a smart sensor-integrated mushroom chamber that monitors and controls temperature, automating the cultivation process. The system achieved a 72% confidence rate, with plans to enhance its accuracy by incorporating a more diverse dataset in the future [86].

4.5. Mushroom Image Datasets

Most of the studies leveraged solutions that require extensive amounts of training data, including mushroom images and their corresponding annotations. Although some studies have used proprietary datasets for the training of ML models [82,84,90,91,92], the lack of utilization of common datasets renders it infeasible to provide a fair comparison of models with identical test sets and metrics. To address this challenge, several publicly available mushroom datasets have been introduced, offering standardized resources for model development and evaluation. Zakeri et al. [89] introduced the M18K dataset, a comprehensive resource with over 18,000 annotated mushroom instances across 423 RGB-D image pairs captured using an Intel RealSense D405 camera (Intel Corporation, Santa Clara, CA, USA). This dataset fills a crucial gap by providing realistic growth environment scenarios, serving as a benchmark for mushroom detection and segmentation algorithms in smart agriculture. Similarly, Anagnostopoulou et al. [93] presented the realistic synthetic mushroom scenes dataset, comprising 15,000 high-quality synthetic images with detailed annotations. This dataset not only supports detection and segmentation tasks but also addresses the challenges of 3D pose estimation in cluttered farm environments, extending its applicability to various crops beyond mushrooms. In addition, Cao [94] contributed a dataset specifically focusing on white button mushrooms, consisting of 1150 pairs of aligned RGB-D images with bounding box annotations. This dataset further supports the development of automated systems for mushroom harvesting and monitoring.

5. Discussion

Research on automated mushroom harvesters has primarily focused on designing, developing, and testing robot manipulators, end-effectors, and computer vision systems. Although collector and mobile platform components are used, they have not been extensively studied. Camera-based vision systems are common, but other detection technologies like LIDAR have yet to be fully explored. Mushroom harvesting robots typically use Cartesian or SCARA manipulators, chosen for their suitability with the standard mushroom shelf design, which features a horizontal workspace. These manipulators excel in broad planar coverage, crucial due to the limited space between shelves (11–14 inches for standard aluminum Dutch shelves) [3,46], which restricts the use of cylindrical or spherical manipulators. As a result, other manipulators must operate outside the shelves, increasing the travel time. For mushrooms grown in open spaces or 3D structures like wooden logs, universal or spherical manipulators would be more appropriate. Commercial systems occasionally employ multiple manipulators to enhance the picking speed, though this approach is used by only a few companies. Figure 9 illustrates the workspace comparisons for different manipulators.

Graspers and vacuum suction cups are the main end-effectors used for mushroom harvesting. Vacuum suction cups, which use pneumatic actuation to create negative pressure, perform well with clustered mushrooms by grabbing the exposed cap without penetrating the cluster. However, they require a higher gripping pressure, which can cause mushroom damage, as listed in Table 1. In contrast, graspers surround the mushroom cap from multiple sides, allowing for a lower gripping force and reducing the risk of damage. However, they face challenges when picking clustered mushrooms because they must navigate within the cluster to grasp individual mushrooms. Improving grasper designs could address these issues. Commercially, two-finger graspers are preferred due to their compact size compared to the three-finger or closed graspers. Graspers designed for delicate fruits like tomatoes and strawberries could offer insights for mushroom harvesting. Comparing the end-effector performance, hand picking remains superior in speed and for minimizing mushroom injuries. Yet, with advancements in design, optimal force application, and integration of sensors and actuators, end-effector performance could improve. Additionally, using multiple end-effectors on manipulators could enhance overall harvesting efficiency.

Collection units for mushroom harvesting can be either on-shelf or off-the-shelf. On-shelf collectors, positioned close to the manipulators, reduce the pick-and-drop time but are generally installed on horizontal planes, which makes reaching other processing units slower. In contrast, off-the-shelf collectors, typically vertical, shorten the mushroom travel time and can be relocated to different shelves. However, moving these collectors requires more time and energy. The choice between on-shelf and off-the-shelf collectors depends on the shelf design. On-shelf collectors are ideal for scenarios with sufficient space between shelf layers, allowing direct connection to post-harvesting mechanisms. When shelf gaps are narrow, off-the-shelf collectors are more practical. These vertical collectors can also incorporate post-harvesting mechanisms to handle tasks while transporting mushrooms. The collector’s design and installation are influenced by the type of manipulators and mobile systems used. Off-the-shelf collectors require a mobile mechanism for relocation. Detailed study of collector designs and their integration with other systems is essential for optimizing their performance.

Mobile platforms for mushroom harvesting can be on-shelf or off-the-shelf. On-shelf platforms, such as Cartesian carts with manipulators, require specially designed shelves to support their movement. Off-the-shelf platforms, like ground-based vehicles, are often used for outdoor fruit and vegetable harvesting. For mushrooms, the space between shelf arrays must accommodate these vehicles, which could reduce the growing space and require the platforms to be robust enough to carry the entire harvester unit. Despite these challenges, ground-based vehicles can be flexible and useful in open 3D growing spaces and for transporting mushrooms from collectors to processing units. The feasibility of such mobile carriers in mushroom harvesting needs further exploration. Sensors in mushroom harvesting are primarily used for data measurement, but their potential as feedback systems remains underutilized. Sensors like LIDAR, ultrasonic, radar, and infrared can aid in obstacle detection and avoidance for mobile platforms and can be combined with vision sensors for mushroom identification. Additionally, sensors such as thermal, acoustic, and vision could be studied to assess mushroom health, detect diseases, and ensure quality.

Another critical aspect of automation is the efficiency of the machines in terms of productivity and profitability. Efficiency is a broad topic and can represent different performance metrics. Nevertheless, we can directly compare some of the picking performance metrics of commercial harvesters with those of manual picking. The picking performance metrics of the modern commercial harvester are compared with manual picking in

Table 4. Comparison of commercial harvester picking performance with manual picking.

Harvester	Picking Speed [Mushrooms/h]	Success Rate [%]	References
Manual	2000	~100	[41,52]
Mycionics	2700	-	[58]
Dorna Robotics	2200	80	[41]
AG Robotics	1600–2000	95	[7]

. Automated harvesting has caught up with manual harvesting in picking speed; however, accounting for the success rate, automated harvesting still has a long way to go. Not to mention, manual picking can be equally effective in picking mushrooms from any growing conditions like cluster mushrooms; however, we do not have enough information to claim how automated harvesters can adapt to different growing conditions. Having said that, automated machines can work 24 h a day [7], increasing the overall productivity.

Although automation has shown its competitiveness in terms of picking speed, energy consumption and economic considerations present significant challenges. Many mushroom farms operate with traditional setups, making the necessary modifications to integrate automation both complex and costly. Additionally, implementing automated systems often requires hiring skilled labor, which comes at a higher wage. Operating expenses, maintenance costs, and potential losses from downtime further contribute to the financial burden, making the transition to automation a substantial investment for many industries. In addition to these challenges, automation systems require a reliable and substantial energy source. The costs associated with energy management, consumption, and potential losses due to power outages must be carefully considered. Therefore, a comprehensive assessment of automation with respect to costs, profitability, and energy consumption must be performed for commercial harvesters.

6. Potential and Challenges

With the global mushroom market continually expanding, automated harvesters present significant opportunities. However, since automated harvesting technology is still developing, there are many challenges and unknowns for both researchers and the mushroom industry.

Potentials:

• Increased Productivity: Semi-automated or fully automated harvesting systems greatly reduce reliance on human labor. Reliable machines can operate for longer hours, enhancing productivity. These systems can be optimized to use the most effective picking methods based on the mushroom’s growth stage, reducing both the picking time and energy expenditure.
• Enhanced Mushroom Monitoring and Quality Control: Automated harvesters can track the entire lifecycle of mushrooms, identifying their growth stages, detecting damage and diseases, and monitoring distribution. With the help of various sensors and feedback systems, these harvesters can detect and address diseases early, preserving the quality and extending the life of the remaining mushrooms, thus improving overall production.
• Seamless Integration with Automated Production Systems: Automated harvesters can be integrated with other automation stages in the mushroom industry, including pre-harvesting and post-harvesting processes. This integration minimizes time delays between stages and enhances communication, leading to greater overall efficiency.
• Advanced Computer Vision: Enhancing computer vision through faster controllers and advanced computers can improve speed and accuracy. Utilizing 3D cameras and depth sensors can help the harvester approach mushrooms from the optimal angle, reducing injuries. Better model training and 3D mapping can also enhance detection accuracy, especially for clustered mushrooms, and shorten detection times.

Challenges:

• Working Environment and Space Availability: Most of the work that has been conducted so far in button mushroom harvesting is for the aluminum Dutch shelves with a specific design and arrangement. There are extant mushroom farms that use wooden shelves and other growing environments, which offer significant challenges to designing the harvesters that can be used in these environments. Modifications to the existing designs may not be feasible. Furthermore, some of the harvester components require significant space inside the shelves to be installed. Therefore, the harvesting system, which is flexible in terms of the working environment, is a critical challenge.
• Adaptability: Button mushrooms grow in different sizes and shapes. The distribution of mushrooms on the growth bed is very irregular: from sparsely distributed single-grown mushrooms to densely packed cluster mushrooms. Thus far, the works have mostly focused on single-grown mushrooms of a limited size range. The damage associated with end-effectors is higher than that of human pickers. The development of an end-effector that can adapt to any mushroom regardless of its size, shape, or distribution is very challenging at the present time.
• Maintenance and Downtime: This is one of the fields that has not been explored for mushroom harvesters. In the long run, the machine will need regular maintenance and may experience downtime due to technical issues. The reliability of the machines overtime and the loss associated with downtime are not yet studied, which could be a critical issue in the future.
• Technical complexity and efficiency: Automated harvesters include different physical and non-physical components that must be meticulously designed and operated. The machine must be able to replicate human sensitivity to avoid any damage and operate at human-level efficiency. This requires significant research progress in the design and programming of the system, which demands skilled manpower and investment. Furthermore, automated machines require high initial investments, operating costs, and energy consumption. The design of energy-efficient machines, as well as a low-cost automation solution for the parts and maintenance, great product service, etc., are some critical areas that cannot be overlooked.