[Jeong Sun-woo Column] Robots That Prune Well and Robots That Grow Well

Robots are entering orchards. Cameras scan the trees, reading the position and thickness of branches, and robotic arms extend to cut the branches with a set amount of force. The cut surfaces are smooth. The sight of machines replacing pruning work that people used to do over several days is impressive in itself. It is also true that rural communities short of labor have long awaited such technology.

 But when lifelong fruit growers watch this scene together, the same question usually comes back. “But why did it cut that branch?”

That is because the essential question in pruning is not 'how to cut' but 'which branches to leave'. Depending on how the branches to leave are chosen, it changes where the nutrients stored in the tree will go, whether sunlight will reach the inside, and how many fruits will set and how large they will grow. Furthermore, it extends to how many flower buds will form the following year, and even whether alternate bearing, in which yields fluctuate every other year, will occur. Pruning is not a cutting task but a physiological judgment that carries over beyond a single year.

This example clearly shows where the center of gravity currently lies in agricultural artificial intelligence research. Technologies that recognize and move are rapidly becoming more sophisticated, but the question of what that work leaves behind in the crop has been placed relatively in the background.

Looking back, over the past 15 years, the subject of agricultural technology discourse has continued to change. Plant factories that grow crops under artificial light regardless of the season emerged as a symbol of future agriculture, and soon afterward smart farms took their place. The stage advanced from remote control through big data and artificial intelligence to unmanned automation, and now physical AI is becoming the new subject. It is clearly an achievement that technology has taken the lead in attracting attention and investment. However, in the meantime, crop physiology research has remained in a supporting role.

 I recall one record from those days. There is a survey from the late 2000s that examined the technological level of plant factories among domestic horticulture experts. When the world’s highest level was set at 100, transport devices and mechanical equipment were around 72, and seeding·seedling systems were about 66 , while items related to factory facilities and environmental control were generally distributed between 50 and 60 . But the field of breeding fruit and vegetable varieties dedicated to plant factories remained at 20.

There is probably no disagreement that the biggest reasons plant factories did not spread as much as expected were the burden of initial investment costs and electricity bills. Still, it is also worth remembering which way the center of gravity of technology was leaning.

 Similar structures are repeated in many places even now. Harvesting robots judge maturity by color and size, but multiple studies point out that the sugar content of fruit has no direct causal relationship with color. If fruit is harvested based only on its appearance, fruit whose sugar content has not yet risen enough is picked along with it. Irrigation·nutrient solution control artificial intelligence responds to sensor values, but even the same moisture condition has different meanings depending on the growth stage. 

The periods most sensitive to water shortage are usually concentrated around flowering and fruit set, and the baseline to be maintained differs from stage to stage. Image-based disease diagnosis reads visible symptoms. However, while the potato late blight pathogen reveals symptoms only after several days of incubation following infection, Alternaria leaf spot pathogens in the same field show symptoms almost immediately when conditions are right. The time it takes for symptoms to appear on the screen differs by pathogen.

 I believe this is why on-site adoption remains slow despite the steady development of technology. Research sites describe performance in terms of recognition accuracy, work success rates, and work time. Farmers judge by yield and quality, and by the stability of next year’s farming. The two languages are different. What farmers want is not a robot that cuts well, but a robot that does not throw off next year’s farming. Trust is created not by a demonstration video, but by the results after a season has passed.

 Then what should be supplemented. I would like to consider three things together.

 First, I hope data reaches all the way to outcomes. Today’s training data generally stops immediately after the task is completed. But the results of farm work always return with a time lag. For leafy vegetables, they appear in growth a few days later, for fruiting vegetables or food crops, in the yield and marketable rate of that cropping season, and for fruit trees, in the flower buds and yield of the following year. If artificial intelligence is to learn whether its judgment was correct, there must be records that continue to the end of this time lag. As the time it takes for answers to return differs by crop, the period for collecting data and the rhythm of research also need to be aligned with that time.

 Next, it would be good to use together what we already have. We have crop growth models and cultivation physiology research accumulated over a long period. In agricultural fields where data is not abundant, it is more advantageous to provide existing knowledge together rather than making systems learn from data alone. Physiology research is not a competitor to artificial intelligence, but an asset that can help it learn.

 Finally, I hope we establish the criteria for success together. When indicators such as yield and quality, and the stability of production in the following year, are placed alongside recognition accuracy, the languages of research and the field grow closer. And such standards are created not at the stage of verifying results, but at the very first stage of deciding what problem to solve. Perhaps the most practical starting point is for engineering researchers and horticulture·crop physiology researchers to sit together at that table.

 In fact, this demand is not new. When the stages of smart farm technology development were divided, the second stage was, from the definition stage, a level at which artificial intelligence diagnosed growth and helped farmers make decisions. Crops were already in the blueprint. It is just that the center of gravity in actual implementation remained on environmental control.

 For physical AI to take root in agriculture, precise hands alone are not enough. It also needs eyes that can read what plants are experiencing now. Crop physiology is not a field that has fallen behind in the age of artificial intelligence, but a pathway that brings artificial intelligence into real fields and orchards.

 I believe that 'agriculture for all', in which technology does not remain only with a few advanced farms but reaches the fields of every farmer, also begins here. So that we do not have to confirm once again the lesson from 15 years ago, I hope technology and physiology research will stand shoulder to shoulder.