Sign up for our daily Newsletter and stay up to date with all the latest news!

Subscribe I am already a subscriber

You are using software which is blocking our advertisements (adblocker).

As we provide the news for free, we are relying on revenues from our banners. So please disable your adblocker and reload the page to continue using this site.
Thanks!

Click here for a guide on disabling your adblocker.

Sign up for our daily Newsletter and stay up to date with all the latest news!

Subscribe I am already a subscriber

All-or-nothing strategy pays of in Autonomous Greenhouse Challenge

The crop did not look very pretty after three months of autonomous growing, but with converted 45 kilograms of tomatoes per square meter, the winning team can go home. Team IDEAS, with Chinese bright minds won the fourth edition of the Autonomous Greenhouse Challenge, in which dwarf tomatoes were grown. Stef Maree and Silke Hemming from Wageningen University & Research (WUR) explain the choices that made the difference and the valuable insights this edition has provided for autonomous greenhouse horticulture.


Team IDEAS: Judge Leo Marcelis, Tao Lin, Wei Liu and Jury Kathy Steppe. (Foto: Sarah Vlekke)

The winning team, IDEAS, representing Zhejiang University in China, dared to place the highest number of pots per square meter in their greenhouse compartment. By using 40 pots per square meter, the team achieved the highest net revenue. Their all-or-nothing strategy paid off.

All five teams, primarily consisting of Asian participants but also experts from commercial companies, including those in the Netherlands, managed to achieve positive net revenue. However, it became evident during Thursday's presentation that there's still much to learn.

Team Tomatonuts finished fifth. The team included participants from Wageningen University, China Agricultural University, Jingwa Agricultural Science and Technology Innovation Centre and Golden Scorpion.

Simulation vs reality
For example, Team Tomatonuts, which included participants from Wageningen University, China Agricultural University, Jingwa Agricultural Science and Technology Innovation Centre, and Golden Scorpion, placed fifth. They acknowledged the significant difference between simulations and real-world greenhouse conditions. Similarly, Team Trigger discovered during the challenge that they lacked sufficient knowledge and experience with all the technical systems in the greenhouse. Good artificial intelligence alone wasn't enough.

In the summer, the teams worked hard to develop an algorithm for autonomously managing their greenhouses. Cultivation began in early September, with the algorithms taking full control of lighting, heating, CO₂, water levels, plant density, and harvest timing.

Vreugdenhil Breeding & Seeds was one of the sponsors. The company develops dwarf tomato plants.

Impact of plant density
Fear of pest outbreaks due to high plant density led Team AgriFusion to limit their density to 25 pots per square meter. Their team captain admitted they were surprised by how effective pest management techniques, such as sticky traps, could be. "If we had known that, we would have chosen a higher plant density."

Despite this, Team AgriFusion finished second. The team included participants from Croft, IMEC, GreenBites, Harvard University, Korea University of Technology and Education, and Seoul National University. The team was supported by Xing Zhao of COSMOS as the domain expert / grower, who co-developed the algorithm with the team from CROFT-AI.


Team AgriFusion finished second. The team included participants from Croft, IMEC, GreenBites, Harvard University, Korea University of Technology and Education and Seoul National University

Stef Maree, who is involved in the Challenge as a data scientist on behalf of the WUR, says about the plant density. "It is common to place the plants further apart during cultivation, so that all the leaves get enough light and the plant keeps a nice shape. But for a good yield, that does not appear to be necessary."

Interestingly, the reference growers, who for the first time used autonomous insights from the AGROS project on cucumber cultivation, also opted for 25 pots per square meter. This yielded a year-round equivalent of just under 35 kilograms of tomatoes per square meter, earning them fourth place among six participants.

Team Trigger came in fourth. The team included participants from Grit, Ridder, Daeyoung, Bigwave, Seoul National University and Keimyung University


When to pick?
All teams struggled to determine the optimal harvest time. "What makes autonomous cultivation of dwarf tomatoes challenging is deciding when to harvest," said project leader Silke Hemming of Wageningen University & Research (WUR). "The plant must have enough ripe fruits, which cannot simply be inferred from the plant's weight, unlike crops like lettuce, which was the focus of the previous edition."

On average, the teams harvested a week too late, incurring unnecessary extra costs. The dense foliage and occasional unintended shoots made it difficult to estimate fruit ripeness from a distance. Hemming added, "Dwarf tomatoes are suitable for autonomous cultivation because they require only one harvest moment, can be harvested by robots, and grow at the same height."

Monique Bijlaard of the WUR repotting the tomato plants, in the run-up to the cultivation competition. On Thursday afternoon, Monique said on behalf of the reference growers that she had a lot of fun seeing the plants from the greenhouse. They were not all equally beautiful, and sometimes they fell over, being top-heavy as they were. "You could clearly see from the plants what was happening in the greenhouse."

AI and sensor use
Although the teams relied on AI, they didn't fully trust it. Several participants imposed limitations on their algorithms to prevent errors. For instance, Team MuGrow restricted their lighting algorithm, which resulted in excess light usage, even when other signals from the greenhouse indicated it wasn't necessary.

Team MuGrow had knowledge on board from TU Delft, Gardin, Birds.ai, Rijk Zwaan, Wageningen University

Sensoren
For greenhouse feedback, the teams relied on sensors. Researcher Kathy Steppe from Ghent University highlighted the importance of sensors, mentioning their role in the hybrid vertical farming setup in Agrotopia, Roeselare. This tower system, equipped with numerous sensors, revealed temperature differences of up to four degrees Celsius between the top and bottom layers on a summer day.


Kathy Steppe

Models
In addition to sensors that generate data, the teams also work with models. That is nothing new, said Leo Marcelis, professor at WUR. He showed models from the 1990s that, for example, could predict harvests fairly well in cucumbers based on climate. If it were up to the WUR researcher, greenhouse horticulture would develop from cultivation based on greenhouse and climate control to cultivation based on plant control. That requires much more targeted sensor data, which can then be converted into models to control entire greenhouses.

Leo Marcelis

The fourth Autonomous Greenhouse Challenge showed that it is still too early for fully autonomous cultivation, but project leader Silke Hemming of the WUR is very satisfied with the level. "The teams have achieved an incredibly high level. With the exception of a few minor interventions, each team has autonomously achieved a successful harvest. And all that well before the deadline of 15 December. This shows that they have developed very good algorithms. And that in a relatively short time. This year you saw - just like in previous years - that the teams initially underestimated how much work it takes and how complex it is. Especially during testing before the challenge starts, things often go differently than planned. But in the end, it turned out well for each team."

"Releasing an algorithm on a greenhouse, where you have a full harvest after a few months, does not yet exist in practice. There is not a single grower who does this completely autonomously. However, individual aspects, such as autonomous temperature regulation, are already being applied. We have shown that you can control a crop – with the exception of components such as IPM – completely autonomously. Of course, there are still plenty of challenges and improvements possible, but the proof that you can reach the finish line with an algorithm is now there." The full results of the Autonomous Greenhouse Challenge 2024. At least 150 grams of tomatoes, of which 1/3 ripe, a fruit weight of at least 6 grams and as a bonus point more than 7% dry matter were requirements imposed on the teams. More data can be found on the dashboard challenge.


The cultivated crops

By the time the harvest was ready, it was noticeable that the crop no longer looked attractive, but that was not the goal.


In-Bok Lee, a jury member from Seoul National University, talked about the development of horticulture in South Korea

Sponsors
Tencent, Biobest, Quantified Sensor Technology, Vreugdenhil Breeding & Seeds, Certhon, Fluence, Lensli, Pöppelmann, Gebr. Geers and LetsGrow.com supported the Autonomous Greenhouse Challenge.