Automation and robotics in the cultivation of pome fruit: Where do we stand today?

The cultivation of apples and pears in orchards consists of several tasks that still demand much human labor. The cost of this skilled labor increases while the number of competent seasonal workers becomes insufficient. These facts are a threat to the fruit industry. To find a solution, this paper addresses current as well as future automation possibilities for the main orchard tasks as a profitable alternative to human labor. Besides an activity research in pome fruit orchards, this paper contains an overall review of the research and developments that have been performed to automate each major activity (e.g., pruning, thinning, spraying, harvesting and mobile navigating) in the cultivation of pome fruit. These tasks are individually evaluated on feasibility and profitability of the developed automations. Finally, this paper concludes that, despite the large amount of research, almost no fully automated and cost‐efficient solution has been developed. A possible option to increase the viability of the prototypes might be the simplification of the tree structures, and consequently the orchard architecture, to make it “robot‐ready.” Another option in this perspective is combining several techniques, for accomplishing individual tasks, in one multipurpose robot platform. As a result, the usability and efficiency of the robot increases.


| INTRODUCTION
In 2017 the production of apples in the European Union (EU) was valued at €3.8 billion. This accounts for 16.5% of EU-28's fruit production. Pome fruit in total (apple and pear) is the only type of fruit with a higher export than import value (De Cicco, 2019). It is fair to say that pome fruit is an important part of EU's fruit industry. The main share of the activities in this sector is still performed manually, often by hired seasonal workers. The related high labor costs, the low market prices and the lack of qualified workers are putting an ever-increasing pressure on the fruit sector today. Therefore, automation and robotics in orchards may provide a solution that additionally considers the increasing environmental challenges.
In the past five decades, there has already been performed a considerable amount of research on the automation of several tasks in an orchard, like harvesting and spraying. Most of these research projects concentrate on one specific task. There are for example several prototypes of automated apple harvesting robots, as will be discussed further in Section 6. This review paper summarizes the current status of automation for each major orchard management This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2020 The Authors. Journal of Field Robotics published by Wiley Periodicals LLC task that is needed to cultivate pome fruit. First, in Section 2 an activity research is discussed. The status for pruning, thinning, spraying, harvesting, and mobile navigating in orchard environments are described in Section 3, Section 4, Section 5, Section 6, and Section 7, respectively. Hereafter, Section 8 will discuss the progress on automated cultivation of other relevant fruits and vegetables.
Finally, the paper offers a conclusion for each task, which discusses the major difficulties and potential next steps for the future in Section 9. A graphical overview of the paper structure is given in Figure 1.
Several review papers about innovations in agriculture already exist. However, these are either very specific and detailed for one task in a certain cultivation, such as He and Schupp (2018) reviewing sensing methods for automated pruning for apple trees, or very broad covering a whole sector to give a general overview, such as Vougioukas (2019) reviewing automation in the whole agricultural sector. This review paper specifically focuses on the recent developments and innovations within the last few decades for the automation of cultivating pome fruit. Hence, the paper presents an overall view covering all the parts of a certain cultivation, but in a detailed way for each part. Specifically for pome fruit, no such review exists, as far as the knowledge of the authors reaches. Furthermore, this review paper is substantiated by an activity study, which exposes the actual needs of the sector.
In the majority of the following sections, a first step towards full automation is described as mechanization. This replaces the expensive and slow manual labor with a mechanical substitute, which is faster and cheaper. However, there is no sensing or controlling, which results in systems that are nonselective in their handlings. In the industrial sector, mechanization is very useful because the circumstances are controlled and continually steady. Fruit trees, and nature in general, however, are never totally controllable or steady.
There are no two trees that can be perfectly treated nonselectively in the same way and still have the most ideal outcome for both. Therefore, mechanization implies some constraints, but it has advantages and useful effects as well. Hence, it is relevant to discuss mechanization before reviewing automation.

| ACTIVITY RESEARCH
Besides reviewing the state of the art regarding automation and robotization of pome fruit cultivation, the current state of activities and related amount of labor for cultivating pome fruit need to be explored as well. This will give a clear vision on the actual needs of the cultivators. The study has been performed in cooperation with the Flemish institution for fruit cultivation Research Center for Fruit, (pcfruit) npo, with the overall goal to uncover the major issues in labor and costs of the sector, as well as uncovering for which cultivation tasks the largest progress can be made by automating a specific part of it. Based on the distance, weight, accuracy, and type of a manual action, this method estimates the duration of the action, by combining data with predefined time tables. This technique is commonly used in time and quality management in the manufacturing sector. Based on observations in the field, the average number of actions for each specific task was calculated. Combining this with the results of the MTM time tables, the total amount of labor time that an orchard task requires was estimated. These time estimates take into account a 13% loss due to organizational issues, such as unplanned timeouts and worker fatigue. Finally, this outcome was presented to five experienced fruit cultivators to confirm the results by comparing these to their know-how of field work. Despite this general verification, all these results have to be nuanced because they are theoretical estimates and thorough validation tests in the field were not carried out yet, but they are planned (ACROFRUIT-KU Leuven HBC2019.2051, 2020).

| General distribution of labor
A first result of the activity research is a general overview of the distribution of labor for each orchard task. For apples, the calculations have been done specifically for the cultivar Jonagold with a Tall Spindle tree architecture, an estimated production of 50 tonnes/ha and counted for three harvesting rounds. The latter has a large influence on the labor time, due to extra checking of ripeness during the picking and extra logistical efforts. The results for the distribution of labor for cultivating apples is shown in Figure 2a. In the study concerning pears (cultivar: Conference), the same tree architecture as for apples has been used (Tall Spindle), as well as the same estimated production of 50 tonnes/ha, but the harvesting of this type of fruit is done in one harvesting round. Therefore, the share of harvesting will F I G U R E 1 Graphical overview of the content of this review paper [Color figure can be viewed at wileyonlinelibrary.com] be smaller. The results of this study for pears are displayed in (1) Focussing on the total amount of labor, for both cultivations these numbers are excessive. To cultivate apple, a total amount of 466 h/ha is needed and for pear it even goes up to 482 h/ha. Spread out over an entire year it seems no problem, but many of these hours have to be performed within small time windows, which puts much stress on the cultivators, who have to organize the amount of seasonal workers based on the quantity of work and available time windows. (2) For both cultivations, the tasks of harvesting and pruning clearly take the largest shares of labor. Those two handlings are the most labor-intensive. For this reason, the highest need for labor reduction lies with these two tasks. The next two paragraphs will discuss the time study for harvesting and pruning more deeply. The small time windows mentioned above are especially problematic for harvesting, whereby this high share of labor has to be performed in only a few weeks. This is in contrast with pruning, for which a time window of multiple months is available. Besides harvesting and pruning, in future work the task of thinning will be investigated in the same way.

| Time study: Harvesting
In this activity research, two harvesting methods have been compared: a basic method using ladders and static bins (Method 1); and an advanced method using working platforms and moving bins (Method 2). Again, -this study has been done for cultivating apple and pear. As shown in Figure 3a, the second method applied to apple saves 30.6% of the needed time. For pear a reduction of 12.6% was obtained. This reduction is higher for apple because in every harvesting round there is some profit to be made. For apple three rounds were counted, unlike for pear with only one harvesting round.
Despite the reduction, still over 216 h/ha are needed for the harvesting task of both cultivations.
F I G U R E 2 Results of the activity research for the distribution of the amount of yearly labor for cultivating (a) apples (Jonagold-Tall Spindle) with a total average amount of yearly labor of 466 h/ha and (b) pears (Conference-Tall Spindle) with a total average amount of 482 h/ha [Color figure can be viewed at wileyonlinelibrary.com] F I G U R E 3 Comparison in consumed labor time between basic methods without the use of extra tooling (Method 1) and methods using aiding tools and platforms (Method 2), specifically for the orchard tasks of (a) harvesting and (b)  the basic method involves a manual shear and the use of ladders (Method 1); the second method uses electrical shears and a working platform (Method 2). The results in Figure 3b display a reduction in labor time of 33.1% and 30.8%, respectively for pruning apple trees and pear trees. However, these results still need to be validated with thorough field tests, which are planned next pruning season. As the results indicate, the use of mechanized aiding tools could reduce the working pressure. The next section discusses the current state of the art dealing with mechanization and automation of this part of cultivation.

| PRUNING
Pruning fruit trees has several purposes. On the one hand, the main purpose is to control the size and structure of the tree. On the other hand, it is possible to control the crop load at an early stage . By pruning, the tree structure can be manipulated to provide a balance between the energy for growing fruits and the energy for growing branches. Parts that would consume too much energy without future profit, such as old, unproductive or diseased branches, can be cut away. Moreover, making cuts at specific places could trigger the growing process, which can be useful in the next years (e.g., new twigs that could guarantee production 2 years later).
In the future, pruning will have another important purpose for the implementation of robotics in orchards. By pruning fruit trees into the right and simplified tree architectures, it is possible to make an orchard "robot-ready" (He & Schupp, 2018). Robinson and Hoying (2013) describe the different tree architectures and orchard systems, and which effect they have on yield and labor costs. They concluded that future orchards need a narrower canopy to decrease the complexity and to increase the visibility and graspability of the features of the tree. In literature, these "robot-ready" tree structures are also described as simpler, narrower, more accessible, and productive (SNAP) tree architectures (Karkee et al., 2014). Bloch et al. (2018) underlined the importance of adjusting the robot as well as the tree architecture in a way that both designs match together. Therefore, they demonstrated a methodology for simultaneously optimizing both the robot kinematics and the working environment. Besides the advantages of "robot-ready" orchards for robotics, these simplified structures will have the same advantages for manual labor, so the related costs will reduce as well.

| Mechanization
Mechanization of pruning is called hedging. As shown in Figure 4, a tractor is driven along the row of trees with a vertical trimming bar.
This results in a nonselective pruning system whereby every branch is cut off at the same distance from the trunk without taking into account the importance of the branch (floral buds, light coverage, age, diseased, etc.). Ferree and Rhodus (1993) concluded that replacing manual pruning in winter with this kind of mechanical pruning decreases the cumulative yield per tree with 35%. However, there are some advantages of hedging if it is applied in the right way. (1) Hedging can be used as an a priori tool to speed up the normal way of pruning. Hedging the outer branches and top of the trees can be useful to reduce an amount of manual pruning labor. (2) Using this principle in summer can be interesting as well. By hedging the outer layer of leaves during summer, the penetration of light through the canopy increases and the fruit matures better. In temperate climates, this can result in fruit of higher quality (Ferree & Rhodus, 1993). In warmer climates, this technique is less advisable or must be handled carefully, due to a greater risk of possible defects to the fruits like sunburn.

| Automation
For manual pruning a certain amount of knowledge and skills is needed to evaluate the tree structure and to decide where to prune, without damaging the fruit tree. The detection of those complex tree structures, pruning decisions, and collision-free robot planning make it even more challenging to automate this part of fruit cultivation.
The activity research reported that manual pruning for cultivating apples corresponds to 16.3% of the total amount of labor. For pears this number rises even up to 27.9%, due to the more labor-intensive F I G U R E 4 Example of a nonselective hedging system with a vertical hedging bar that trims the branches at the same distance from the trunk (He & Schupp, 2018) [Color figure can be viewed at wileyonlinelibrary.com] tree architectures of those orchards. Thereby, pruning can be ranked as the orchard task with the second largest share in the manual labor. He and Schupp (2018) confirm these numbers as they reported that pruning includes more than 20% of the costs for orchard management, although these numbers may vary depending on the practised orchard structure and tree architecture. Furthermore, the activity research reported specifically for pruning that the use of tooling, such as electrical shears and working platforms, reduces the labor time by more than 30%. Consequently, the need for and relevance of extra mechanical aiding tools and the next step of automating this manipulation in the orchard is clear. The maximum obtained accuracy for detecting branches was 92%.
Another conclusion was a 6% higher accuracy when adding the depth data compared to the system without including depth images. Chattopadhyay et al. (2016) and Elfiky et al. (2015) used the Kinect v2 camera as well to measure, reconstruct and model apple trees with the aim to automate pruning. Because existing reconstruction algorithms (e.g., Visual Hull Reconstruction) failed on thin textureless objects, such as fruit trees, Tabb (2013) developed a voxelbased formalism. Four years later, Tabb and Medeiros (2017) validated this system in field trails. Besides scanning and reconstructing the tree structure, they measured several characteristics of the tree as well (e.g., branch diameter, branch length and Although it is not developed for fruit orchards, a fully working prototype that can trim bushes and prune roses in regular gardens is the aim of the Trimbot 2020 project. The goal of the project is a commercial robot, which is similar to a lawn mower robot, that can be found in many gardens these days. In this project, much progress has been made in path planning for outdoor platforms, object detection in gardens and automated trimming (Kaljaca et al., 2019;Strisciuglio et al., 2018). More specifically for cherry orchards, You et al. (2020) recently developed a conceptual pruning robot. The detection of branches and the possible cutting points has been performed with a RealSense RGB-D camera in combination with an OctoMap model. They reported an average success rate of 92%, with an average throughput time of 5.71 s for each cut. These averages were based on the data of ten test runs done on a self-made indoor test setup.
Finally, Botterill et al. (2017) developed a pruning system for grapevines that uses three cameras to model the tree lay-out and an Artificial Intelligence (AI) system that decides where to prune. This system obtained a low error of 1% on the trajectory estimation and reached an acceptable working speed of 2 min per vine in field trails, which is comparable to human labor. Despite the above mentioned developments of robotic pruning prototypes, no developments have yet been made specifically for pome fruit orchards, which have a more complex branch structure than the cases of cherries and grapevines. Hence, still much progress can be made in this field of research.

| THINNING
The thinning principle practices the rule of quality over quantity.
Controlling the crop load is very important to indemnify the quality of the fruit. By removing a certain number of fruit, the remaining fruits will receive a higher share of necessary nutrients, producing more high-quality fruit instead of a high quantity of lower quality fruit. Furthermore, by selectively removing fruits with less potential (e.g., too small or with deformations), the fraction of high-quality fruit can be increased. As mentioned above, this can be done in an early stage by pruning in the correct way. However, this is not sufficient, so additional thinning is required. There are two types of thinning: blossom thinning and fruit thinning, which are compared in Table 1.
Both thinning types could be done with several methods, like mechanical thinning, chemical thinning, and thinning by shading. In this paper, only mechanical thinning will be discussed, because of its lower environmental impact than the chemical alternative. For other methods the reader may consult (Byers et al., 1986;Greene et al., 2013;Wouters, 2014).

| Mechanization
String thinners are a first kind of mechanical thinning whereby the most common type is called the Darwin machine (Miller et al., 2011). This nonselective mechanization of the thinning process VERBIEST ET AL.  Figure 5a and 5b, respectively. Besides string thinners, spiked drum shakers can be used to shake a number of fruitlets out of the tree as a manner of thinning. This method has the same disadvantages of nonselectivity, damage, and disease spreading. Moreover, it has the additional downside of shaking the largest fruitlets away, due to a higher inertia. However, these large fruitlets have the highest potential of reaching high quality and are preferably not removed (Wouters, 2014).
T A B L E 1 Advantages and disadvantages of both blossom and fruit thinning

Blossom thinning
Fruit thinning Advantages • The required nutrients to grow into fruitlets will be saved for other fruit.
• Small fruits are easier to handle than flowers.
• At the stage of blossom the leaf volume is not at its maximum, allowing an easier detection of the blossoms.
• There is more certainty about the expected yield.

Disadvantages
• A higher risk on lower yield because of late frost.
• The tree needs to deliver more energy to the starting fruitlets that eventually will be thinned.

| Automation
A fully operational and commercially available automated robot for this orchard activity has not yet been developed. However, Wouters (2014) engineered a working prototype, which solved some disadvantages of the mechanized solutions described above. By using pressurized air, multispectral computer vision, and precisely positionable nozzles, it is possible to selectively blow floral buds away.
This method does not touch any part of the tree nor does it cause any extra damage to it. Therefore, the system will not spread diseases.
However, the prototype (Figure 5c) is very slow and uses a large amount of pressurized air. Hence, the efficiency of this technique, in its current form, is too low to be used in an orchard in a profitable way.
Also Yang (2012) engineered a robot for automated thinning of fruit. However, this was a down-scaled prototype, tested in laboratory conditions. As a result out of these tests came a design of an end effector for selective thinning, which resembles a miniature version of a string thinner as shown in Figure 5d. Although the results were promising, many improvements need to be made towards a full-scale robotic fruit thinner. Future steps in this project could be the actual development of the end effector for outdoor field tests and validating the principle in orchard circumstances.

| SPRAYING
To protect an orchard against diseases, such as apple scab (Venturia The largest challenge is decreasing the amount of chemicals and the impact on the environment to a minimum. Therefore, drift reduction is very important. These days, drift reduction is mostly applied by using drift reducing spraying nozzles that produce bigger drops whose trajectory is less affected by wind. To validate spraying systems or to measure drift, water sensitive paper could be used, such as focused on the effects of canopy density on the spraying flow and the drift of the product in citrus orchards. This study reported that 28% of the sprayed volume is not deposited on any fruit tree. Because spraying in pome fruit orchards is always done with machines, the subdivision of mechanization and automation is not totally appropriate. For this task, another subdivision is preferable, namely the one used by Tona et al. (2018). The spraying equipment can be categorized into three technological levels. The first level L0 contains the conventional spraying techniques, level L1 contains the partly controlled spraying techniques and L2 is the level of precision spraying. Figure 6 shows the conceptual difference between the three levels.

| L0: Conventional spraying level
Conventional air-blast spraying is the most used type of spraying, but also the least automated one. An axial fan blows an air flow that

| L2: Precision spraying level
Precision spraying refers to canopy-optimized spraying systems that are based on 3D sensor data to record the full characteristics (volume, density, shape, etc.) of the trees in the orchard. Based on these 3D data the spraying can be controlled in amount and flow with controlled nozzles, in spraying distance and in spraying angle with certain actuators. Hočevar et al. (2010) developed and tested an automated system for precision spraying in orchards. They used RGB images as input to calculate the contours of the canopy. Out of the field test results could be concluded that a saving of chemicals of 23% in relation to traditional spraying systems (L0) is possible. However, they also noticed that these savings depend on the structure of the orchard. In high-density orchards the savings will be lower than in low-density orchards in relation to a conventional spraying system in such orchards. Osterman et al. (2013) modified this design using real-time laser scanner measurements as input data to record the canopy as a point cloud, whence they filter the contours. Based on the formation of this contour the optimal spraying flow, spraying distance, and spraying angle are calculated and executed with a controllable spraying arm consisting of three movable parts as shown in Figure 7.
In the Netherlands, Nieuwenhuizen and Stallinga (2013) used laser scan data for their fully autonomous spraying system, which could also navigate autonomously through the orchard. This system was tested and they reported a saving of product up to 53%. Berk et al. This was integrated on a robot platform which was part of the CROPS project. They reported that 85%-100% of the diseased canopy was treated with a reduction of 65%-85% in pesticide usage. Tona et al. (2018) analysed for the three technological levels whether it is profitable to implement them, depending on the size of the orchard (apple). They concluded that for orchards smaller than 17 ha level L0 is the most profitable, for orchards larger than 17 ha it is more profitable to use level L1. The level of precision spraying (L2) is currently not profitable, because the high investments could not be recovered by the additional saving of pesticides. The same analysis was done for vineyards (grapevines). In vineyards smaller than 10 ha the conventional level L0 is more economical, for vineyards of 10 ha up to 100 ha it is more profitable to use level L1 and for vineyards bigger than 100 ha it is more profitable to use a precision spraying system of level L2. However, these conclusions are based on a generic model. Depending on the circumstances and the used technologies, these numbers could be very different.

| HARVESTING
The goal of all previous tasks and labor is to harvest fruit of good quality in a profitable way. However, this harvesting has a high labor cost as well. The activity research discussed in Section 2 shows that for the manual picking of pears, the amount of labor could go up to 51.8% of the total labor load and for apple this amount is even 66.9% of the total labor hours for cultivating apples. Back in 1993, Sarig (1993) already reviewed the then actual possibilities of automating the task of picking apples. Although no cost-effective product was yet available at that time, they concluded that much research presumed that it would only be a matter of time and money before further robotization of fruit cultivation would replace manual laborers in orchards. The current status of mechanized and robotic harvesting is discussed below.

| Mechanization
The first type of mechanized harvesting is the nonselective harvesting machine such as a limb shaker, a trunk shaker, or rotating beater bars. As the names already indicate, these machines apply a brusque mechanical force whereby the fruit will fall off the trees.
However, these nonselective mechanical forces injure fragile highquality fruits like apples, causing many bruises, and decrease the quality and price of the fruit. In addition, the branches of the fruit F I G U R E 7 Precision spraying system with eight degrees of freedom. According to the canopy segments, the spraying distance, spraying angle and flow will be adapted (Osterman et al., 2013) trees will be damaged too. It can be concluded that this kind of mechanization is only applicable for industrial fruit (for jams, juice, etc.) and for less fragile classes of fruit (e.g., citrus and olives). More information about mechanized harvesting is reviewed in P. Li et al. (2011).
Another method for the mechanization of harvesting pome fruit is the use of a mechanical aiding platform. The picking will still be done by workers, but the actuation of the platform height, the outflow and the collection of fruit in bins will be done by a mechanical and partially automated platform. An example of a commercial harvest aiding platform is the Pluk-O-Trak, as shown in Figure 8 (Pluk-O-Trak; Munckhof, 2019). According to Baugher et al. (2009), mobile orchard platforms could increase the working efficiency with 19% and even up to 67%, depending on the platform type and the performed tasks. These numbers were confirmed by the performed activity study of Section 2.2, reporting a reduction in labor time of 13% up to 30%. Hence, these systems can reduce the labor cost, but they still include interaction with manual laborers.

| Automation
The automation of the orchard task of harvesting can be divided in two major automation challenges. On the one hand, the detection system for detecting the fruits. On the other hand, the robotic part of gripping and picking the apple. After describing these two parts, both detection and robotics, will be combined in the discussion of the currently developed robotic harvesting prototypes. Outdoor mobile navigation is an extensive field of research, which is not only applicable to the automated cultivation of pome fruit. Therefore, this paper does only contain a general overview of the challenges and developments of outdoor mobile navigation directly linked to orchards. It goes beyond the scope of this paper to give a detailed description of every technological realization in this broad field of research.

| Challenges
First of all, the changing weather and light conditions may considerably complicate outdoor navigation. On the one hand these conditions, like heavy rain, fog, sunny versus cloudy weather, angle of the sun, snow, and so on, could affect the measurements of sensors that are needed for localization and navigation. A way to take this into account is described in Bargoti and Underwood (2016) where these circumstances are added to the algorithms as metadata. On the other hand, the weather has consequences for the state of the terrain. Rain, freezing, or fallen leaves can cause a slippery underground. Morales et al. (2009) describe that piles of leaves or branches could be incorrectly recognized as obstacles, although it is possible to drive over these. Apart from the influence of weather, the state of the terrain is a big problem for outdoor navigation as well. Negative obstacles (holes and depressions) as well as sudden slopes can be unpredictable and due to this the vehicle could get stuck or tip over. Heidari (2014) shows an approach to detect and handle those negative obstacles. By pointing a 3D laser scanner at an angle to the ground, irregularities in the surface can be calculated as shown in providing enough metadata to the localization system as described in Bargoti and Underwood (2016). Strisciuglio et al. (2018) suggested segmentation as another possible solution for this issue. A segmentation algorithm distinguishes drivable from nondrivable areas. Thus by doing this, an apple tree will always be treated as a tree, with or without leaves. Although an orchard is an outdoor environment, global navigation satellite systems (GNSS) are not always reliable because of signal occlusions by the trees surrounding the vehicle. Underwood et al. (2015) and several others describe this as a large issue for mobile navigation in orchards. Therefore, they developed GNSS-free localization and navigation systems that will be summarized below.

| Developments
The navigation of mobile platforms through an orchard needs to be accurate. Besides the accuracy for navigating from point A to B within the orchard, the system has to take into account the manipulations towards the tree or fruits that will be executed simultaneously, whereby its trajectory has to be adapted. Consequently because radar has, next to its higher range, an improved penetration through canopies, but it has some disadvantages with relation to LiDAR as well. In most cases an EKF or particle filter algorithm is used for fusing information from different sensors. Hansen et al. in Crane et al. (2006), Ball et al. (2016), Paton et al. (2017), Gu et al. (2018), and Kragh and Underwood (2020 Vineyard Crawler-robotmakers GmbH, 2020) and (3) the Dutch company Precision Makers presents the Greenbot for ±€100 000 (Greenbot-Precision Makers, 2020). Figure 11 shows examples of the discussed autonomous vehicles. As mentioned above, the field of research is too broad to discuss it in detail in this paper; more detailed information about other autonomous navigating agricultural vehicles is reviewed in M. Li et al. (2009), Shalal et al. (2012, and Gao et al. (2018).

| ROBOTICS IN OTHER RELEVANT CULTIVATIONS
Besides for pome fruit, research has been performed for other cultivations as well. Although these research projects had another crop as purpose, the used technologies could also be useful for apples and pears. In this paper, the discussed crops are subdivided in cultivations in greenhouses, and outdoor cultivations.
For greenhouse environments, the following projects were con- In every part of the agricultural sector, precision farming, innovative technologies, and robotics are being investigated. In a similar way as this paper reviews the recent innovations for the cultivation of pome fruit, Vigneault (2016, 2017), as well as Fountas et al. (2020), both reviewed the developments in agricultural robots for field operations, and Vougioukas (2019) reviewed the recent innovations in the total agricultural sector.
However, the latter only provides a general overview without detailed descriptions.

| CONCLUSIONS AND FUTURE PERSPECTIVES
An overview of all mentioned developments in this review paper is displayed in Tables 1, 2, and 3 of Annex 1. The tables show that for the orchard management tasks harvesting, spraying and mobile navigation quite some progress has already been made. The systems in these areas claim high accuracy, but their efficiency, and consequently their profitability, is still too low to be directly applicable for average fruit cultivators. Furthermore, these systems are typically developed for specific and simplified circumstances, which are not generally present in standard orchards. For the orchard tasks thinning and pruning less progress has yet been made. Completely automated and selective pruning or thinning robots (or prototypes) for pome fruit trees have still not been developed. Combining this lack with the need for reducing the high amount of manual labor, indicates that these fields of research could have a high potential.
As this review paper covers all major parts of the cultivation of pome fruit, no general conclusions can be made that are applicable to every part. Concluding the review of thinning and spraying in one sentence is like comparing apples with oranges. Therefore, a proper and detailed conclusion will be made for each discussed part of cultivating pome fruit.
Activity research. The outcome of this study proves that harvesting takes the largest part of labor for cultivating apples (67%), as well as for pears (52%). Also manual pruning and thinning require a high amount of labor, even relatively higher for pears than Spraying. These days, the automation of spraying systems is an important topic, due to the related environmental concerns. Therefore, much R&D has already been done and will be done in the future, because the regulations continue to become stricter. Precision spraying is promising for the future, but it is still too expensive to be profitable. The profit of extra saved pesticides is not enough to counter the high development costs of a complex precision spraying system. Nevertheless, this level of sprayers probably will break through, not due to an economical motive, but due to the stern environmental regulations of the government.
Harvesting. Automated harvesting is probably the activity in an Out of these conclusions, multiple concerns are repeatedly highlighted. For future robotic manipulation in orchards it is important to focus on four prospects.
(1) It is necessary to make the orchards suitable for robotic automation, as previously called making them "robot-ready." This will simplify the complexity of the automation solution, whereby the automation cost could decrease. However, this will take time, because changing the orchard structure means growing trees in another way. To guarantee yield for the farmer, this changing of orchard system can be done gradually by replacing old trees with new trees, with the right structure, over the years. Even for manual cultivation, a simplified orchard will reduce labor. Hence, this is a critical point for the future of pome fruit cultivation. A point that should be taken, with or without future implementation of robotics.
(2) For each task, the existing automations need to be optimized. On the one hand, this means for well explored topics, such as harvesting, that less expensive techniques could be combined into a profitable and real-time system (e.g., RGB vision with CNNs and suction cup grippers). On the other hand, for the less explored topics, such as thinning, extra research is needed. However, several techniques used for other tasks could be transferred, as discussed above.
(3) All mentioned developments are dedicated to one specific task, making it not profitable at all. A dedicated harvesting robot, for instance, can only be used for four weeks each year and cannot be profitable in this short time of use. Therefore, future orchard robots need to be developed for more than one orchard task. If the platform could be used for several tasks, the profitability will increase. For example, the robotic manipulator used for picking apples could change its end effector into a selective thinning device. Consequently, by combining multiple modular units into one multipurpose robot platform, the profitability, the feasibility and the efficiency of the system will increase, so that regular cultivators can use it as a realistic solution for the challenges in their sector. 526 | (4) The majority of the discussed research projects tried to find a solution for either the detection of features, or the performance of an actuation in an orchard. Nevertheless, besides sensing and acting, robotics relies on decision-making as well. Where should be pruned? Which fruits should be harvested? What thinning rate is preferable according to the detected blossoms of that tree?
Only a few of the discussed research projects investigated this part of robotics, instead of generically choosing fixed parameters for these possible decisions of their system. The quality of robotic fruit cultivation will increase by making the right and selective decisions in pruning, thinning, harvesting, and so on. Besides this quality, it could influence the performance of the system as well.
Taking the example of picking, the decision of which apple should be picked first could also affect the difficulty of the picking task itself. By choosing the less complex and more effective picks, performance rates of the harvesting robot will increase. Although this means a second manual harvesting round will be necessary, for the current developments this is still necessary as well. So, in relation to the current developments, this kind of decisionmaking will not affect the amount of complementary manual labor. In conclusion, including more evolved decision-making in the think-part of future developments in the field of robotic fruit cultivation could have a positive effect on both quality and quantity of the cultivation.

ACKNOWLEDGMENTS
The authors are grateful for the opportunities and support of KU