Why the winning AI greenhouse grew its tomatoes closer together

Team IDEAS won Wageningen’s fourth Autonomous Greenhouse Challenge by keeping dwarf tomato plants unusually close together. The competition, whose results were announced in January 2025, tested whether algorithms could direct growing decisions all the way through to a real harvest.[1]

Six 96-square-metre compartments were planted on 3 September 2024. Teams could revise their software until 15 September; thereafter, algorithms directed operations until harvest. Climate equipment responded automatically. Greenhouse staff physically spaced and harvested plants when instructed by the algorithms. Five competing teams and a WUR reference algorithm took part.[2]

The scoring rewarded ripe fruit and charged for resources, equipment and time. IDEAS’ denser planting spread those costs across more saleable pots, alongside efficient use of energy screens. Giving individual plants more space can improve their shape; the trial showed that a different arrangement could earn more across the greenhouse.[1]

The greenhouse had to report back to its controller

The public dataset records the information surrounding those decisions: greenhouse climate, weather and control states at five-minute intervals, plus plant measurements and canopy images. Each compartment had a camera capturing colour and depth four times daily. Researchers also recorded the final harvest manually. The evidence therefore connects the changing environment to both the actions taken and the crop eventually produced.[3]

That feedback matters because a growing plan is exposed to changes throughout the season. A controller needs to decide whether the conditions it creates are producing the intended plant response. A record of warm air and bright lamps alone cannot establish that the crop is developing as expected.

MuGrow illustrates one approach. According to team member Gardin, the controller combined measurements of chlorophyll fluorescence, which reveal aspects of photosynthetic performance, with models that planned climate adjustments and reinforcement learning that supplied longer-term targets. Gardin participated alongside TU Delft and Rijk Zwaan. Its account describes software inspired by a grower’s daily inspection: observe the crop, reassess its development and adjust the strategy. This is a participant’s description of its own system.[4]

The scientific paper assesses the resulting strategies; teams did not disclose their complete algorithms.[2]

Knowing when to finish was part of the job

Dwarf tomatoes have a single harvest. Waiting allows ripening but adds costs. Organisers reported one irrigation intervention and two interventions over harvest timing. Pest-management choices remained a separate human responsibility. Substantial cultivation decisions had been delegated, with people available to make exceptional corrections.[1]

Gardin’s account says MuGrow used computer vision to recognise growth stages and inform spacing and harvest. It attributes a late harvest to conservative rules around those predictions. Its estimate of better returns from earlier picking is retrospective, so it should be read as an opportunity identified after the trial. The crop itself cannot be rerun on a different date.[4]

A useful experiment beyond the winning score

The released data include raw readings and processed records, with calculated energy use and costs. Researchers can inspect the sequence of actions, compare the compartments and test ways to recognise crop development. The files provide observations from a real growing cycle; reproducing a particular team’s controller would require additional information about its software.[3]

The strongest lesson is that autonomy can involve sustained responsibility across many ordinary decisions. Heating, watering, spacing and harvest timing have to work together over a crop’s lifetime. This experiment makes that responsibility visible: there is a physical outcome to weigh, a history to inspect and a clear account of where people stepped in. Better autonomous growing will depend on improving that entire sequence, including the decision that the job is finished.

Report an error or suggest a correction