Wageningen’s Autonomous Greenhouse: What the Cucumber Trial Proved

The project: remote decisions for a cucumber crop

In 2018, five AI teams and a grower reference operated six greenhouse compartments in Bleiswijk, the Netherlands. Each compartment covered 96 m² and grew the cucumber cultivar Hi-Power during an approximately four-month period. Wageningen’s dataset record identifies the teams and describes the recorded climate, irrigation, crop and resource information. WUR experiment and dataset record.

This was a physical growing experiment with bounded conditions. For a facility owner, that makes it more informative than a simulation alone, while still leaving a substantial gap between a research compartment and a different commercial operation. The useful reading question is which parts of the result depend on the trial’s setting.

What the algorithms controlled—and what people still did

According to the researchers’ 2020 account, teams used sensor and crop observations to set climate and fertigation controls remotely. Weekly instructions covered leaf and fruit pruning. Pest and disease management remained WUR’s responsibility and was outside the competition. Elings and colleagues’ trial account.

The 2019 paper describes approaches combining crop knowledge, models and data-driven control. It discusses simulation models used to develop strategies and the different choices teams made around lighting and crop management. Hemming and colleagues’ technical paper. The case should therefore be read as coordinated control in an established growing system.

For an indoor project, Hamfy’s interpretation is to draw the control boundary explicitly. Mark which system measures each variable, which software proposes or changes a setpoint, which controller enforces the limits and which person performs crop work. If a task remains manual, include its timing and cost in the operating plan.

The result supports a capability claim with limits

The researchers report that the winning AI approach outperformed the grower reference. Their account also notes that the winning team’s higher production involved greater investment in lighting and trade-offs in some resource-use efficiencies. Published results. This is evidence of successful control under those experimental conditions, not a universal promise that AI reduces every input.

The technical paper discusses how strategy and economic conditions interact. Hemming et al., 2019. Hamfy’s practical reading is to ask what was optimised, what was counted and how the result changes when energy prices, crop prices or operating constraints change.

Before adopting a headline result in a business plan, separate saleable output, resource use and financial performance. A control strategy may improve one measure while worsening another. Ask for the definition and denominator behind each number, and compare it with the objective for the proposed site.

Transfer the evaluation method before the setpoints

A greenhouse with sunlight and a fully enclosed indoor room have different operating constraints. This trial does not supply a ready-made indoor climate recipe. Its more transferable contribution is the idea of evaluating a coordinated strategy against a named reference over a complete crop period.

Hamfy proposes beginning with a narrow operating objective and an observation phase. Gather reliable records of climate, irrigation, crop work, energy and saleable harvest. Then run proposed control decisions alongside the existing process without applying them, and review the disagreements with the grower.

Only after that review should a supervised trial change a limited set of controls. Define allowed ranges, the manual override and the conditions that return control to the established system. Record software versions and all interventions. This sequence is our proposed engineering approach; it was not a separate treatment measured in the Wageningen trial.

Build a comparison that can answer the business question

Write the baseline before running the trial. Decide how the comparison accounts for crop stage, planting density, usable area, light exposure and any difference in labour. A previous crop cycle can provide context, but changing weather or production conditions may prevent a clean causal comparison.

Agree on primary outcomes and secondary constraints. For example, an objective around stable crop quality can coexist with limits on energy and operator workload. Record failed runs, downtime and interventions as part of the result. Excluding the difficult days would make a pilot less useful for the team that has to operate the facility.

The WUR dataset record links to experimental data that make further investigation possible. This article has not independently recalculated the trial’s financial results. Our camera-pilot guide develops a smaller first step for facilities that are still establishing their observation and data systems.