[Interview] Korea Institute of Robot and Convergence Director Cho Yong-jun Begins Developing Full-Cycle Apple Robot to Prepare for “Appleflation”

In apple farming, the year’s crop depends on how well pollination goes during the brief spring blooming period. Pollination is the job of honeybees. But circumstances vary from year to year, with climate change and other factors even causing mass honeybee deaths. If frost damage coincides with the flowering season, fruit set becomes poor, apple prices rise, and this spreads into so-called “appleflation” (apple + inflation), pushing up overall prices. Conversely, in years when the crop is excessively good, prices fall and farmers suffer losses.

Farms that cannot rely solely on honeybees fill the gap with artificial pollination, in which pollen is applied directly to flowers. Most of that pollen, however, is imported. According to Cho Yong-jun, director of the Agricultural Food Research Division at the Korea Institute of Robot and Convergence (KIRO), 80–90% of the pollen used by domestic fruit farms comes from China. Concerns have consistently been raised that pathogens such as fire blight could enter the country along with imported pollen. But people in the field explain that imports cannot be blocked entirely because pollination work would immediately come to a halt, so management is limited to restricting affected regions.

The reason domestic pollen production is not increasing lies at the production site. According to Director Cho, Cheongsong, North Gyeongsang Province, has the country’s largest 2-hectare complex where only flowering trees are grown to obtain pollen. The period in which flowers can be picked is just over a week each year. Even if all the elderly residents of the village are gathered during that time to pick flowers by hand, the recovery rate remains around 30%, while the rest falls to the ground and is discarded. Labor costs for that one week reach 60 million won, but annual pollen production is only 5–10 kg. Since domestic pollen is four to five times more expensive than Chinese pollen, farms do not seek it out; and because they do not seek it out, production does not increase—a vicious cycle. This is the background behind KIRO’s move to develop fruit-tree robots covering everything from pollination to pollen production.

In an interview held on Oct. 30 at KIRO’s Agricultural Robot Demonstration Center in Andong, Director Cho said, “The biggest cause of poor harvests is the environment, followed by honeybees, whose situation changes every year. When people hear that robots will pollinate, they may misunderstand and think honeybees are no longer needed, but pollination is best done by honeybees. However, since honeybees are unstable, the idea is for robots to assist and stabilize production.”
 

Integrated RDA–MAFRA project…KIRO handles the “lower body” and integration

This project is an integrated agricultural AX (AI transformation) project jointly promoted by the Rural Development Administration and the Ministry of Agriculture, Food and Rural Affairs. It consists of three subprojects, with overall coordination handled by R&D Project 1 (Metafarmers Inc.), led by the RDA. If the robot is compared to a human body, the upper-body hardware corresponding to the arms and torso, along with task AI, will be handled by Metafarmers Inc., which is in charge of R&D Project 1, “Development of Control Technology Based on Imitation Learning of Behavioral Characteristics in Fruit-Tree Farming Operations and Research on Standardization of Agricultural Robot Technology.” Visual intelligence—the robot’s “eyes” for recognizing orchards and fruit—is being developed by Chungnam National University Industry-Academic Cooperation Foundation (Professor Lee Dae-hyun) under “Development of Core Visual Intelligence Technologies for Analysis of Required Performance in Farm Work Responding to Orchard Environments and Decision-Making.” KIRO’s R&D Project 3, selected through the Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET), “Development of a Cultivation Management Robot Platform and Intelligent Technologies for Stabilizing Fruit Production,” will develop the lower-body autonomous driving platform and driving AI. The research period runs for four years and nine months, from April 2026 to December 2030.

KIRO is also responsible for system integration—combining the upper body and visual intelligence produced in each project into one system—as well as field demonstration. Director Cho explained, “Overall coordination is handled by the RDA, but because MAFRA projects have to consider mass production, it was agreed that we would handle system integration.” About 20 institutions participated in each project (consortium), and the competition ratio was about 4 to 1. Director Cho said, “You can consider that virtually all domestic institutions, companies, and universities working in robotics took part this time.”

In KIRO’s project, local funding from North Gyeongsang Province and Andong City was matched in addition to national funding. The ratio of national to local funding is about 6 to 4. Although it was not a mandatory requirement, local governments in major apple-producing areas requested that flower harvesting and pollen production—pain points in the field—be included in the tasks in exchange for making additional investments. Director Cho explained that MAFRA’s Fruit and Floriculture Division also has strong interest in securing domestic pollen.
 

Why apples? Focus narrowed to one multi-axis planar tree form

The target crop has been narrowed to apples, and among them to the multi-axis planar tree form. The multi-axis planar form, in which trees are grown flat like a wall, makes it easy for machines to enter, and if it is machine-friendly, it is also robot-friendly. Director Cho said that during farmer interviews in the planning stage, he heard comments such as, “Older farms are reluctant to change, asking how long they can keep farming after pulling out all the trees and replanting, but farmers in their 40s and 50s are switching everything because the differences in production and labor savings are significant.” According to Director Cho, the multi-axis planar form is most widely adopted in Geochang, South Gyeongsang Province, and major producing areas such as Andong and Cheongsong are also rapidly transitioning from the existing slender spindle form.

Fixing the tree form to one type is a lesson learned from past failures. To win competitive projects, proposals had to claim they could handle spindle, planar, and straight-line forms alike, but after winning the contracts, that often became a burden and results failed to materialize. Director Cho said, “Saying robots can do everything in all conditions is pie in the sky. The right approach is to change cultivation methods to be robot-friendly and improve efficiency with robots suited to that. This time, the tree form has been clearly designated, so I personally have high expectations.”

Even so, orchards are not easy sites. In open fields, direct sunlight can wash out camera images, making it impossible for robots to see anything. Many existing orchards are on hilly land with slopes of 5–10 degrees, and if a robot damages a tree that has been grown for 10–15 years after planting, the loss is not limited to one year’s crop. Difficulty in finding farms for demonstration is also an obstacle. Director Cho said, “Even if we pay rent, it has to make sense in terms of apple production profitability, so there are not many farms willing to accommodate us.”
 

One platform with interchangeable work tools…flowers are “beaten and sucked in”

A year of apple farming is broadly divided into nine operations, from winter pruning to harvesting. Of these, the national-funded project covers four—pruning, pollination, flower thinning, and harvesting—while flower harvesting and pollen production have been added with local funding. The upper platform is used in common, while the end hand tools and auxiliary devices are changed according to the task. For pruning, dual arms equipped with grippers to hold branches and cutters are attached; for pollination, a tip that precisely sprays pollen close to flowers is mounted. During flower thinning, a tank for thinning agents is loaded, and during harvesting, a conveyor that pushes out boxes is installed. Pruned branches are immediately shredded by a front-mounted crusher and spread through the orchard. Director Cho said, “If branches are cut and left there, they have to be collected again, but if they are shredded, they also become nutrients, so there were many positive reactions.”

For harvesting, work speed is the key. The work efficiency required by the project request for proposals (RFP) is 1,000 square meters per hour (about 300 pyeong), but that speed could not be met with a method of picking apples, turning around, and placing them in boxes. Therefore, the team chose a structure in which apples are dropped immediately into a corrugated hose and sent to boxes. The remaining task is to prevent fruit damage that may occur during descent.

The flower harvester being developed with local funding does not pick flowers one by one like a person. When a rotating striking unit shakes the flowering tree, a case blocks the falling flowers and a negative-pressure suction device draws them in and collects them in a tank. The target is to recover 70% in a single pass. Related patents have already been secured. Some loss remains because flowers scatter outward, but Director Cho said, “Even recovering only 70% is far better than people picking them by hand. Even if the equipment costs tens of millions of won, if it succeeds for just one year, it will be groundbreaking for pollen production complexes.” He said farmers in the field also saw the demonstration concept and responded that it had potential. As flower harvest volume increases, KIRO plans to create a pilot model with expanded pollen manufacturing capacity and hand it over to major producing areas.
 

Routes from a single satellite image…thousands of robots learn in simulation

The technological differentiation KIRO emphasizes lies in driving. According to Director Cho, existing autonomous agricultural machinery requires a person to drive it once to create a map. On large farms, that task alone is work. KIRO extracts tree rows from satellite imagery using AI and automatically plans the entire driving route (global path) by reflecting the robot’s specifications. In the field, it responds only to obstacles and detailed routes (local paths) using lidar, cameras, and GPS. Director Cho explained, “If you compare it to navigation, it is like deciding which highway to take with just one satellite photo, and responding only to alleys and crosswalks on site.”

For the technology that keeps the upper body level on uneven open-field terrain, “Sim-to-Real” AI is applied. The method creates millions of pyeong of irregular terrain in a virtual environment and runs thousands of robots, allowing them to learn on their own the control conditions that maintain posture, prevent slipping, and use the least power. Each of the four wheels functions as a suspension, absorbing ground undulations while the upper body remains steady. Director Cho said, “In the past, algorithms were written in the form of setting angular velocity according to steering angle, but the conditions changed with even a single rock. AI’s ability to adapt to the environment on its own is the advantage of physical AI, and that is why we are trying to cover the full cycle rather than just one task.”
 

Selling “work” instead of selling robots…about half the labor cost

The commercialization model has two tracks. The core is robots as a service (RaaS). When a farm books work by mobile phone, engineers from a service company load robots, bring them to the site, deploy several at once to do the work, and then retrieve them. The plan is to set fees at about 50% of existing labor costs and vary them according to the type of work, area, slope, and location. For very large farms or institutions that employ their own engineers, finished products will be sold turnkey. The sale price has been set at around 80 million won based on a 2.5-year payback period, and below 100 million won. Current discussions point toward Metafarmers of Project A handling services and Daedong handling sales, but specific roles will be discussed again at the commercialization stage.

Director Cho said, “Even if you bring a robot to a farm, the farm cannot use it. Agricultural machinery has dealerships, so after-sales service is available, but robots do not yet have that infrastructure. That is why we designed a structure in which farms outsource cultivation work and pay service fees.” The choice of a full-cycle robot was made for the same reason. He emphasized, “If it only harvests, it is used for one week a year; if it only pollinates, it is used for three days. Who would buy equipment like that for tens of millions of won? You have to attach all the operations to make it a robot that is used year-round.”

Field deployment assumes swarm work in which at least three units cooperate. With three units, the target is to process 1,000 square meters per hour (about 300 pyeong) for pruning and harvesting, and several times that for pollination and flower thinning, which involve spraying while passing through. Daedong and Daedong Robotics (jointly established by Daedong Mobility and KIRO) are participating in the consortium, and integration is also being reviewed in which harvested boxes are transferred by conveyor and then received and moved by Daedong’s transport robot. Dexweaver, which has experience controlling industrial cleaning robots, will handle central control of multiple robots and service systems, while the Korea Construction Equipment Technology Institute will handle auxiliary systems such as crushers and conveyors that require high load and high torque. Sunchon National University will conduct control and safety research to supplement robots that easily break down in open fields from an agricultural machinery perspective, and the Korea Future Agriculture Research Institute in Andong will build the demonstration site.
 

Flower harvester to be deployed for the first time in spring 2028…re-pollination after frost damage is the goal

The schedule is to release an initial driving platform in the second half of 2027, combine it with the upper body in 2028 to demonstrate harvesting, pollination, and flower harvesting, and then sequentially demonstrate pruning and flower thinning. The final two years will be used for demonstration and improvement. The goal is to create a prototype flower harvester after mid-2027 and deploy it in the field for the first time when flowers bloom in spring 2028. Demonstration will take place in Andong. KIRO will convert an existing slender spindle orchard next to its demonstration center into a multi-axis planar form in winter 2026 to create its own test site, and will also use leased farm plots, public plots managed by the Korea Future Agriculture Research Institute, and a robot-dedicated demonstration site where flowering trees will also be planted. The Korea Future Agriculture Research Institute’s plots are equipped with sensors and 3D modeling from a previous MAFRA open-field smart agriculture project.

What Director Cho expects is not just labor savings but reducing the range of harvest fluctuations. When spring frost damage occurs, essentially the only response is to pollinate the remaining flowers once more, but at present, that requires touching each flower with a brush, making it hard to even attempt. Some farms spray pollen with speed sprayers (SS machines), but Director Cho said research findings have shown that this is ineffective.

Director Cho said, “When frost damage occurs, if pollination robots can be deployed in time to pollinate the surviving flowers once more, the damage can be reduced. If we have work machines that can enter at any time, we can come much closer to the ultimate goal of stabilizing fruit production.”