How Root AI's Agricultural Robots Are Powering the Farmtech Revolution
“Robots don’t get biological viruses.”
Rob Leclerc makes the point not simply as a timely spin on the familiar robots-don’t-take-sick-days case for automation. Right now, in agriculture, the threat feels existential.
“We’re seeing farmers right now with this acute sense that their entire business is at risk,” said Leclerc, founding partner of agtech venture capital firm AgFunder. “One person could get everyone sick.”
Like many in or adjacent to agriculture, Leclerc believed the industry needed further automation even before COVID-19 cast the issue of labor tenuousness into stark relief. Growers have long struggled to plug labor shortages, lamenting both the low appetite among Americans to do the hard work of harvesting and the high cost of bringing in foreign seasonal laborers, on H2A visas.
“For most kinds of crops, it’s still a very expensive proposition.”
The problem isn’t unique to America, either. Even after the U.K. government issued a call to the nation’s unemployed to help pick fruits and vegetables during the busy summer season, it still had to fly in hundreds of Romanian migrant workers after few Brits who applied actually could or would do the work.
Farmworkers who work in the States on H2As will still be allowed to do so during the pandemic, but fears of visa delays nonetheless rattled the industry, and protections for workers remain inconsistent.
“The importance of automation, even partial automation, is becoming pretty significant,” Leclerc said.
Agriculture is already heavily automated and mechanized. In fact, the sector has shrank to less than 2 percent of the labor force in the U.S., surely due, in part, to the rise of the machines. And that includes harvesting. Field and row crops such as potatoes, wheat and corn are often gathered mechanically. But many delicate leafy greens and soft fruits and vegetables have remained largely resistant to robotic picking. Like the Pointer Sisters, they require an easy touch.
Not bruising bruisable crops is hard enough, but the challenge doesn’t stop there. Any scalable solution has to be cost-efficient to implement. The technology requires “quite complex motion planning, computer vision and object manipulation in the field,” said Vikram Adve, co-director of the University of Illinois–Urbana-Champaign’s Center for Digital Agriculture, which includes agricultural automation among its research focuses. Some commercial solutions exist, he said, “but only in very high-end, expensive equipment.”
“It’s far from being economical for most farmers,” he added. “For most kinds of crops, it’s still a very expensive proposition.”
Ready for a Commercial Close-Up
Josh Lessing has been working to crack the challenge since he co-founded agriculture-robotics startup Root AI, in 2018, and he believes his company is on the precipice of a big step forward. Root AI has developed a robot, dubbed Virgo, that can pick at least one of those high-value, delicate fruits — tomatoes — and potentially more.
Most tomatoes grown to be processed into concentrates and condiments, like tomato paste and ketchup, have long been harvested mechanically, but fresh market tomatoes are often handpicked. Virgo aims to help automate that, too, through a sophisticated combination of AI-enabled ripeness detection and robotic dexterity.
Root AI has already put Virgo through R&D trial runs in greenhouses throughout North America, but it’s now about to hit prime time. The robot is being deployed commercially for tomato harvesting to a series of growers this summer, and it will be more broadly available in North America and Europe beginning in 2021, Lessing told Built In.
The company hit the accelerator in part due to the pandemic. “We were asked by a commercial grower to speed up our timelines because of COVID, because of the need to have security around the work now,” said Lessing, who, along with his co-founder Ryan Knopf, previously worked at robotics trailblazer Soft Robotics.
“We were asked by a commercial grower to speed up our timelines because of COVID.”
Here’s how it works: Cameras help the robot maneuver itself along greenhouse rows. (Most market tomatoes are grown indoors, but Virgo can also be deployed outside, according to Lessing.) Cameras also scan the vineyards, and computer vision processes the quality and degree of ripeness of each tomato. (For viewers of a certain age, a video demonstration of Virgo’s “point of view,” with isolating squares around each fruit and corresponding ripeness percentages, plays a bit like an agriculture version of Terminator vision.) Sensors and a whole lot of data input help make sure Virgo can distinguish between fruits and leaves and vines.
The picker itself is plastic, pliable and three-fingered — imagine a streamlined, horizontal version of the old claw-crane arcade games, except made to actually grip. A robotic arm extends the gripper to the tomato, it gently clasps then spins the tomato free with a quick, flick-of-the-wrist-style motion. The gripper deposits the tomato into one of several storage shelves built into the mobile robotic unit. Then, it repeats as necessary.
Going ‘Cross-Crop’ by Mimicking Human Help
Virgo is being deployed commercially to harvest tomatoes, but its big selling point might be that it’s designed to one day pick much more. The goal is what Leclerc, whose AgFunder is among those who invest in Root AI, calls “universal dexterity” — one robot, many different crop applications, including more of those soft-and-delicate crops that have long flummoxed robot pickers. The hope is that it would scale better than single-crop solutions, such as the strawberry-picking robots being developed at Harvest CROO, another noteworthy agbot startup. (Harvest CROO did not respond to requests for an interview.)
Root AI is able to attempt multiple crops because Virgo was designed to mimic human harvesters, according to Lessing. “We can go cross-crop, in a world where no one else ever has, because we’re architecting our entire system to mimic a human body,” he said.
“We can go cross-crop, in a world where no one else ever has, because we’re architecting our entire system to mimic a human body.”
“If a human being, with how their arm is arranged, is able to pick multiple different kinds of crops, then designing your robot based on that biomimicry is the most sensible way to make machines that can do multiple crops themselves,” he added
The proportions of the robotic arm mimics those of a six-foot-person’s arms — shoulder-to-elbow, elbow-to-wrist, the cross-section of the hand.
“I could stand next to it, lay my arm on top of its arm and you’d see: this is a human’s arm,” Lessing said. That biomimicry has allowed the team to tackle what might otherwise seem “a boundless technical challenge.”
It looks and sounds quite impressive, to be sure, but is it truly ready for the majors? Abundant Robotics, a Google Ventures-backed agbot startup that focuses on apple-picking, for instance, has reportedly had to delay its commercialization timeline more than once. (Abundant Robotics did not return a Built In request for an interview.) And is Root certain it’s not just trading old labor costs for new technical ones? Robotic greenhouses and vertical farms, for instance, often have several roboticists and crop experts overseeing the mechanical help.
Lessing didn’t elaborate on prices, but he believes the answer is yes, thanks to a confluence of factors, namely reduced costs, improved dexterity and strong demand. “The underlying costs of building a robot have massively dropped,” he said. “We’re now able to deploy these systems for growers in a way that their cost structure doesn’t go up.”
Toward the Robot Revolution
Root AI isn’t alone in the mission to further automate harvesting. Another notable agbot company, Israel-based FFRobotics is developing a robotic apple picker that it claims can be adapted to pick other fruits as well.
Picking tree fruits sold for fresh consumption, rather than processing, has also proven tricky to automate. Growers around the world could help the transition by adapting their planting practices to better suit robots, said Manoj Karkee, who leads agricultural automation engineering research at Washington State University and advises FFRobotics.
“No robotic solution would be practically adoptable for conventional, dimensional trees,” he said. “Getting to fruit that’s hidden in the canopy behind branches, leaves and other fruit is not easy. It’s possible, but not practical, because agriculture runs on thin margins.” Planting higher-density, so-called “fruiting wall”-style orchards would make the job easier for robots.
“Getting to fruit that’s hidden in the canopy behind branches, leaves and other fruit is not easy. It’s possible, but not practical.”
Like Lessing, he’s optimistic about timelines. “I’m pretty confident that we’ll start to have commercially adopted machines in some scale in the next few years,” he said.
Even if that means a hybrid system, rather than total automation. “Even in the modern orchards, I think our philosophy has to be a collaborative model,” he said. Up to 85 percent of the job might be done by fruit-harvesting bots, and a few workers would follow to finish up — a potentially more difficult task than first-pass picking.
According to Leclerc, the central challenge has shifted from dexterity to data, and how quickly machines can train. “You need to bump into every possible edge case of limitation, so you get the data network effects,” he said. “The more data and experience you get, the better the system gets.”
Another roadblock, according to Karkee, could be connectivity in rural areas. Machines that require even heavy processing lifts can mostly run on the cloud, as long as it can be reached. AI at the edge is improving, and cloud resources “for creating new models are just getting less expensive and more capable,” Lessing noted.
Karkee is also developing handheld image-processing devices, currently in the prototype stage. Workers would scan fields, and the system would determine crop viability, volume and location, for automated picking — all of which is handled by the cloud, no edge computing necessary.
There are other aspects of agriculture aside from harvesting that could use robotic help. Weeds in the Midwest are getting so pesticide-resistant that improvements will be needed in mechanized weeding, for example, according to Adve. Breakthroughs in automated planting and tilling could aid cover cropping, which farmers often eschew due time and cost demands, Adve added.
But harvesting has long been the key ask. Karkee recalls meeting years ago with stakeholders — researchers, farmers, engineers, horticulturists — and being told that picking was the biggest challenge they wanted to solve. That drove his research focus.
“It’s not just because I wanted to work on a cool project,” he said. “It came from the industry.”
It still does, especially now. The pandemic has been “a very stark reminder that we need to build a durable food supply chain that heavily leverages automation,” Lessing said. “Crops are going unplanted and fruits are going unharvested.”
There are, of course, legitimate concerns about the effects automation will have on workers, and whether or not the need for certain automation is indeed urgent. In the trucking industry, for example, some voices have urged the hastening of self-driving trucks to fill a labor shortage in that sector, even though some statistics cast doubt on just how severe said shortage really is.
No such debate seems to exist in agriculture harvesting, where the consensus firmly points to a need for a mechanical helping hand, or gripper. Perhaps that’s why Lessing is unafraid to broach the question head-on.
“There’s a big public debate around what robotics and artificial intelligence mean for the future of work, and it’s a conversation we all need to have in earnest right now, [but] agriculture is one of these rare examples where it’s kind of a separate conversation,” he said. “The societal benefit from a hyper-efficient, durable, sustainable food supply fundamentally enables the resolution of one of mankind’s biggest needs.”