Organic Broadcaster

Researchers test robots that locate, remove weeds without damaging crops

By Marianne Stein

Girish Chowdhary from the University of Illinois shows off the robot weeder he and his team are developing. Photo by Claire Benjamin

Weed management remains one of the most difficult challenges for organic growers. Even conventional growers are struggling as “superweeds” develop resistance to herbicides. Crop losses due to herbicide-resistant weeds are estimated at a half billion dollars per year in the U.S. and are expected to increase dramatically as chemical control is lost for many weeds.

Researchers in the College of Agricultural, Consumer and Environmental Sciences at the University of Illinois are working to provide alternative tools for weed control: small, autonomous robots that work together in teams, removing weeds at the ground level.

Standard mechanical control is impractical after planting because tractors and other machines are too large to work under the canopy. Hand-weeding by humans is more precise, but it is time-consuming and expensive, explained Girish Chowdhary, assistant professor of agricultural and biological engineering at U of I.

“Human laborers are really good at dexterity, but they can’t really scale up. Large equipment is really good at scaling up, but it’s not very dexterous,” Chowdhary said. “How do we fill this gap between high dexterity and high efficiency?”

To bridge that gap, Chowdhary’s team developed TerraSentia, a small robot that combines the speed and power of technology with the accuracy of a human’s close attention.

“These robots will be helpful for a broad swath of different grower types,” explained Adam Davis, professor and head of crop sciences at U of I. “Weed management for organic growers is a challenge as organic production systems do not have a lot of chemical options. And, with increasing herbicide-resistance, it has become necessary to find other ways of managing weeds for conventional growers as well.”

TerraSentia is a compact, rugged robot powered by a lithium battery. The robots are directed by a mobile app and are currently semi-automatic. They can run up and down rows of crops by themselves, but must change directions by remote control. They are steered by GPS when moving between rows, and LIDAR (light detection and ranging) technology under the canopy.

The robots are small enough to move under the canopy and between rows. They are nimble enough to work in dense fields, locating and pulling weeds without damaging the crop. They are also lightweight enough to run in wet soil without doing damage.

The robots are still being developed and researchers have several hurdles to overcome. One challenge is to teach the robots how to distinguish between weeds and crops.

“Farmers do this using knowledge; they know what a crop looks like,” Chowdhary said. “The robot needs to have the same skill. We use machine learning to do this, providing the robot with many examples of corn and not-corn so it learns to tell the difference.”

Through machine learning, or artificial intelligence, the robots also learn how to navigate in different field conditions and move past rocks or other obstacles.

Another major challenge has been to create the mechanical application that removes the weed. Researchers have tried different options, such as motorized arms that brushed off the weeds. The current iteration uses a small plow or hoe that pulls the weed out of the soil. Each robot carries its own hoe, removing the weed through soil disturbance as it moves down the rows.

“A mechanically pulled mechanism, like a hoe, reduces the complexity of the problem. As the complexity drops, so does the cost. The goal is to find the simplest solution, because otherwise it’s difficult to scale up,” Chowdhary explained.

Research also has shown it is most efficient to have the robots get the weeds in the white-thread stage as they are just germinating.

“There will be different worker classes of robots,” Davis added. “Some will have snippers, others will use soil disturbance. There will also be towing robots that can bring the weeding robots back to a solar power charging station when they run out of battery.”

The robots will work the fields as a coordinated team that provide information to each other and feed into a collective map of spatial weed population density.

“We’re developing a framework following a concept called ‘bandit’ after one-armed bandits, or slot machines,” Chowdhary said. “Imagine having a bunch of slot machines in front of you. How do you know which one to pull to get a payout? This framework can be applied to the problem of coordinating the robots so they know where to go. The question is which rows provide the greatest payout, in terms of the amount of weeds to be removed.”

This concept can be used to build a learning model of the field; first, the robots explore the conditions, then use that to build a heat map of hotspots that tells them where to go for maximum results.

The weeding robots are currently in the beta-testing stage; the researchers expect to have them ready in about three years. They have been tested in university fields, and will be tested with growers this fall.

PrairiErth Farm, a family-run 400-acre USDA certified organic operation in Atlanta, Illinois, will be one of the sites for field testing. Dave Bishop, who owns and operates PrairiErth Farm with his son, Hans, and daughter-in-law, Katie, believes that robot technology can be particularly useful for high-value crops.

“We grow 35 acres of vegetable crops, as well as corn, soybean, and small grains,” Bishop said. “The vegetable crops are high-value and typically more problematic from the standpoint of weed control.” Because of the cost of robots in the developmental phase, a higher-value crop would be the logical place to trial the initial models of these robots, he added.

A robotic weeder scouts a corn field. Photo by Claire Benjamin

“We hope that they’ll become one of the tools we can use in this larger toolbox of weed control techniques, along with increased diversity, increased rotations, and livestock inclusion,” Bishop explained. “To have a clean field, there will have to be multiple approaches and multiple tools used, and robots are being thought of as one of those tools.”

The robots will be manufactured and sold by EarthSense, a start-up company incubated at the University of Illinois research park. While the weeding robots are not available commercially yet, EarthSense currently provides a version of TerraSentia that scouts the field and collects crop data such as plant count, stem height, and leaf measurements. The scouting robots were first developed for phenotyping, a method of relating plant information to genetic traits that helps researchers develop better-performing varieties.

“There was a need for smaller vehicles that could move under the canopy and collect detailed plant data at large quantities,” explained Chowdhary, who is the Chief Technology Officer of EarthSense. He notes that while the robots were first intended for research, growers were also excited about the opportunity to get precise data at the ground level. For example, early indicators of nitrogen stress are found under the canopy and cannot be seen in data from aircrafts or satellites.

The cost of the robots is currently prohibitive for individual farmers, with a price of around $25,000 per robot, including the service expertise to run them in the field. The goal is to eventually get the cost down to around $5,000 per robot. However, it is still not likely that farmers will purchase their own robots. A more feasible model is a service contract: the farmer contracts with a company that will supply the equipment as well as people to manage and direct the robots in the field.

The researchers are also exploring other potential applications for TerraSentia, including fruit-harvesting robots. “The idea is to make octopus-like arms that can reach different parts of the plant,” Chowdhary said. “Hard arms do not work well, so we are working on designing soft arms.” He added that robots for fruit harvesting are still in the early stages of development and may take another 5 to 10 years to be ready for field testing.

Marianne Stein works in the Office of Marketing Communications in the College of Agricultural, Consumer and Environmental Science at the University of Illinois.

 

From the September | October 2019 Issue

 

Back to Current Issue

Comments are closed.