Advances in artificial intelligence are helping some startups develop another way to pollinate plants, which could increase yield compared with insects and human workers.
By guest author Mike Cherney from thew Wall Street Journal
The Future of Everything covers the innovation and technology transforming the way we live, work and play, with monthly issues on health, money, cities and more. This month is Artificial Intelligence, online starting July 2 and in the paper on July 9.
Farmers have long relied on insects, wind and even human workers to help pollinate their crops. Now, advances in artificial intelligence are helping some startups develop another way to pollinate plants: robots.
Across the globe, start-ups are testing robots to pollinate everything from blueberries to almonds. And in Australia, one company is so confident in robots’ abilities that it will soon deploy a fleet of them to pollinate tomatoes in its greenhouses.
Pollination robots could give future farmers a significant advantage, increasing yield compared with using insects, such as bees, and the human workers who are sometimes needed to help with certain crops. Scientists are also concerned that insect populations are declining because of habitat loss, pesticide use, climate change and other factors, which would make pollination robots even more important.
The robotic revolution is being sped up by so-called deep learning—a method that trains artificial neural networks that mimic the human brain. Advances in deep learning over the past decade have vastly improved AI’s ability to recognize images. That makes it easier for startups to develop robots that can quickly and accurately identify flowers for pollination.
“To see it actually happening, it was quite amazing,” says Tal Kanety, a senior grower manager in the vertical-farming unit at Costa Group Holdings Ltd. , the Australian grower which tested some robots before deciding to deploy more. “I thought it would work, I just didn’t think it would work that well on the first go.”
The robots were developed by Israel-based Arugga AI Farming, which used deep learning. The robots autonomously travel down a row in the greenhouse, use AI and cameras to quickly identify flowers ready for pollinating, and then blast air at the flowers to pollinate them. The cameras and air nozzles are on a mast and can reach as high as about 13 feet.
Unlike some other crops, tomato flowers pollinate themselves once vibrations shake pollen loose, so the pollen doesn’t need to be transferred between flowers. Costa will pay a monthly fee for Arugga to pollinate about 25 acres in its greenhouses, and roughly 30 robots will eventually be needed to cover the full acreage. The first seven are expected to arrive in the next few months.
In most countries, greenhouse tomato growers pay for bumblebees, which land on the flower and move their flight muscles to create the vibration needed for the flower to pollinate—unlike honeybees, which don’t perform this so-called buzz pollination. But in Australia, there are no bumblebees in most of the country, and strict biosecurity laws prevent them from being imported. So Australian growers rely on human workers, who use a vibrating wand to shake the plants.
Costa tested two robots before agreeing to more, and the results were on par with the labor-intensive manual pollination, says Mr. Kanety. He expects the robots to improve as Arugga makes refinements and hopes the robots will eventually achieve better results than manual pollination. The robots could also limit the possibility of disease, given that the air blast pollinates the flowers without touching them.
Deep learning involves feeding machines lots of data so the AI can learn patterns itself without requiring humans to program knowledge into the machine. Without deep learning, a similar robot would have taken much longer to build and the product would have been more expensive, says Iddo Geltner, Arugga’s chief executive and co-founder. Other startups could also take advantage of deep learning, which could make it easier to develop robots that help with a variety of farm tasks, such as picking or pruning.
“We have to have one algorithm to detect flowers anywhere in the world,” Mr. Geltner says. “We cannot teach the robot to find flowers every time in a different greenhouse.”
Each robot now costs about USD 10000 to manufacture, though costs are expected to come down as the company scales up production, says Eytan Heller, another co-founder. Arugga hopes to later add other capabilities to its system that will reduce labor costs for growers, such as pruning.
Other Israeli companies, including Edete Precision Technologies for Agriculture and Bumblebee AI, are also working on robotic pollination systems for other crops, while researchers at U.S. universities are developing their own prototypes. Investment in agricultural technology overall has grown in recent years, with startups doing everything from using drones to collect farm data to deploying other robots to help wrangle cattle.
A syndrome called colony collapse disorder has been a problem recently for honeybees, but that hasn’t had an impact on bumblebee production, says Karel Bolckmans, chief operating officer at Biobest, a Belgium-based company that produces bumblebees for farmers. Mr. Bolckmans says Biobest is investing in Arugga anyway.
“People ask me, ‘Why are you guys getting involved with robots? You guys are bumblebee people.’ No, we are pollination people,” he says. “If a robot is better than bumblebees, then robots it will be.”
Given recent advances in deep learning, the main challenge for AI researchers is getting enough training data to sufficiently teach an algorithm, says Ian Reid, a professor at the University of Adelaide and the director of robotic vision at the school’s Australian Institute for Machine Learning. Arugga’s computer-vision technique is fairly standard given what researchers know about deep learning, but it would have taken a lot of work to make it reliable enough for the robot to be commercially viable in a real-world setting, he says.
Prof. Reid, who isn’t involved with Arugga but reviewed its website, says the company did a good job building a robot that does a physical task as a result of what it recognizes in an image. In other applications, like the face-detection feature in digital cameras, simply detecting an object in the image is the end goal.
Growers in the U.S., where there are concerns that a new virus that affects tomato plants could be spread by bumblebees, have also tested Arugga’s robots. Josh Lessing, chief technology officer at AppHarvest, which grows greenhouse tomatoes in Kentucky and recently tried one of Arugga’s robots, says the robot beat manual pollination and was getting competitive with the bees. He says AppHarvest would continue to work with Arugga.
“Bees work pretty well, but in agriculture, every boost that you can get your hands on is a big deal,” Mr. Lessing says.