Is it possible to grow a quality lettuce crop without entering the greenhouse?

GLASE project members contributed to the winning team of the third edition of the Autonomous Greenhouse Challenge using artificial intelligence to produce lettuce crops without entering the greenhouse.

by David Kuack

Someday in the future robots will move through greenhouses growing, monitoring and harvesting food crops without any humans touching the plants. Research continues to develop the technology, equipment and sensors to produce these crops while minimizing inputs including energy, water, fertilizer, supplemental light, carbon dioxide and manual labor.

The Autonomous Greenhouse Challenge hosted by Wageningen University & Research (WUR) in Bleiswijk, Netherlands, began with the first edition in 2018-2019 to grow cucumbers using artificial intelligence (AI) control algorithms to minimize production inputs. The goal of the challenge was to produce a commercial quality, salable crop. A second edition of the challenge was completed in 2020 focused on greenhouse tomato production.

The third edition of the challenge concluded this past June and focused on the production of greenhouse lettuce. As in the previous challenges, five teams were selected for the final crop production phase of the competition. Each team was tasked with using its own AI algorithm to produce the best quality lettuce crop while minimizing inputs and making as few changes to their control algorithm as possible.

The crop production phase consisted of two rounds. The first round allowed the teams to test their AI algorithms and gauge plant performance. During the second round the goal was to allow the AI algorithms to operate as autonomously as possible. The teams were penalized for any local interventions to either the control system or crop production system.

A.J. Both at Rutgers University said short-term greenhouse crops like lettuce and leafy greens can be as difficult to produce as long-term crops when they are grown on a commercial scale and growers are trying to optimize parameters. Photo courtesy of Neil Mattson, Cornell Univ.

GLASE researchers on winning team

Team Koala, which won this year’s challenge, included team captain Ken Tran, founder of Koidra Inc., Neil Mattson, greenhouse horticulture professor at Cornell University, and A.J. Both, greenhouse engineering professor at Rutgers University. Mattson and Both are researchers with the Greenhouse Lighting and Systems Engineering (GLASE) consortium. The Koala team name is a combination of the letters in Koidra and GLASE.

Tran was also captain of the winning team Sonoma during the first Autonomous Greenhouse Challenge that focused on the production of greenhouse cucumbers.
“Ken and his Koidra team did the bulk of the work for the third edition challenge,” Both said. “Neil and I were brought in as advisors and sounding boards for some of the decisions the Koidra team wanted to make.

“The fact that Ken and his team had participated and won the first challenge was very instrumental in us being successful during the third challenge. Ken was very familiar with the whole challenge process. The main difference between the two challenges was the crop.”

Both and Mattson contributed their expertise on hydroponic lettuce production to the team.

“The artificial intelligence component, the nuts and bolts of the control system, came from Koidra,” Both said. “Neil and I were not involved in any of the programming or any of the thinking that went into the entire AI approach. We provided Ken and his team with the environmental parameters such as temperature, humidity and light levels for lettuce production.

“The challenge was set up so that the control system should work as autonomously as possible. The decisions that the control system was going to make were programmed ahead of time. There were some opportunities for us to intervene as a team, but during the final production cycle, any interventions came at a cost.”

Mattson made recommendations on crop responses to the environment feeding into the environmental set points for the algorithm. There were two crop cycles during the challenge.

“The first crop cycle, which occurred in February, was a preliminary round where we tested the AI algorithm that Koidra developed,” Mattson said. “The lettuce cultivar ‘Lugano’ that we grew was completely new to us. ‘Lugano’ was not very responsive to light. It got tipburn or a related disorder under high light conditions. During our weekly meetings we discussed what we were observing with the plants and we would modify the algorithm. For the second crop cycle we decided to take a more conservative approach as to how much light the plants should receive. That turned out to be a good decision.”

Koidra’s expertise comes in mathematical modeling, artificial intelligence, software development and whole system optimization.

“Growing lettuce or any other crop in a highly controlled environment with a lot of automation and variables to control is a complex system to optimize,” Tran said. “It requires the know-how of many experts from multiple disciplines with a system optimization mindset to do well.”

The Koidra artificial intelligence algorithm was able to show a top-down version of lettuce plants. Blue circles indicate plants that appear different from other plants. Red circles highlight areas on some of plants with discolored tissue. Photo courtesy of Koidra LLC.

No easy crop

Both said growers often have the misconception that short-term crop cycles are easier than long-term crop cycles.
“Some people might think growing lettuce and other leafy greens is easy,” Both said. “However, once growers start producing them commercially and try to make a living producing them, it’s not as easy as it appears. It’s probably a misnomer in terms of talking about leafy greens as an easy crop.
“Producing leafy greens is still a biological system that requires a lot of attention. All the parameters need to be optimized to get a crop to harvestable size as quickly and as economically possible.”

Need for more data

Mattson said one of the things he learned from participating in the challenge is the greenhouse climate is very complex and there are many interacting factors, including light, humidity, carbon dioxide and light levels, involved with controlling HVAC and lighting equipment.

“All of these factors have different interacting effects on each other and I am impressed with growers’ ability to do what they do without AI,” Mattson said. “I have often heard this from growers and from our predecessor Lou Albright, who helped found the Cornell CEA Program, about the need for crop production experience throughout the whole year. This experience is critical to being able to really understand how to grow a crop because the growing conditions are so different in February than in June.”

One area where growers are going to need more information if they want to adopt AI is on specific plant cultivars.

“To really try to optimize productivity, growers are going to need information on a cultivar by cultivar basis,” Mattson said. “Using AI, growers are going to need training data sets for each cultivar along with production location data. Ideally breeders should be able to provide growers with training data sets. This is necessary even when growing in a controlled environment because production environments can vary greatly among growers.

Members of Team Koala said side images of ‘Lugano’ lettuce plants would have been helpful to see how they were developing and whether that development was on track with what they were expecting. Photo courtesy of Rijk Zwaan.

“One thing the Autonomous Greenhouse Challenge organizers provided us with was images of lettuce heads from the top down and their dimensions–width, length and fresh weight. This kind of data can be used to train an AI control algorithm on monitoring crop growth. Having some of this same information available from breeders, including images, plant age and dimensions, could be very useful for growers and for implementing AI.”

Both said working with AI during the challenge showed him the more data the better.

“Trying to observe a crop from a distance on a computer screen and trying to decide whether the AI control was working properly is not easy,” Both said. “We had top-down images of the crop, but no side images. We could sort of see the plant’s expansion from the top, but could not see the plant growth from the side. These side images would have been very helpful to see how the plants were developing and whether that development was on track with what we were expecting.”

Another area that Both said would have been helpful was detailed information about development of the root system.

“During the two crop cycles we really wanted to see what was going on related to the roots,” he said. “The more data that can be collected, the more it can be used to train the control algorithm. The more precise the control system can be, the more successful growers will be.”


Net Profit: The five teams that participated in the third edition of the Autonomous Greenhouse Challenge competed to autonomously grow a lettuce crop using an artificial intelligence algorithm. The team scores were calculated based on profitability– sellable heads of lettuce minus fixed costs and operational costs (including heating, electricity, lighting, carbon dioxide and intervention to the algorithm) per crop cycle. Team Koala won the competition with a net profit of 5.93 Euros per square meter.


For more: Neil Mattson, Cornell University, School of Integrative Plant Science, Horticulture Section,;

A.J. Both, Rutgers University, School of Environmental and Biological Sciences, Department of Environmental Sciences,;

Koidra Inc., (425) 276-8400;;


David Kuack is a freelance technical writer in Fort Worth, Texas;