Industries / Food, farming and forestry / Agriculture and farming / AgTech
Indoor farming
Indoor farming technology is the equipment and software sold to growers who raise crops in greenhouses or closed rooms: soil-free growing systems, grow lights, and the computers that run heat, light, water and carbon dioxide.
4
Milestones achieved, from 1977 to 2022.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1977
A computer runs the greenhouse
Priva, a Dutch firm founded in 1959 by Jan Prins and Cor Valk to import greenhouse heaters, introduced its first climate or process computer for horticulture in 1977. A case study held in Erasmus University Rotterdam's repository says it was "the first computer in horticulture that controlled climate, energy and water". Such climate computers still move the equipment: in Ontario trials from 2022 to 2025, AI software sent set points every five minutes to a greenhouse's Hoogendoorn climate computer, which worked the heaters and vents.
Credit
Priva. )
2018
Grow lights get outside checks
The DesignLights Consortium (DLC), an independent US nonprofit, started its horticultural lighting program in 2018. Its Qualified Products List admits only grow lights whose efficacy, the light plants can use per watt, it has verified against its threshold, and by January 2025 it listed more than 1,200 products from more than 130 makers. Version 4.0 of its rules, announced in March 2025, raised the threshold 8.7%, so listed fixtures must be at least 45% more efficient than the best non-LED option, a 1,000-watt double-ended high-pressure sodium lamp.
Credit
DesignLights Consortium
2020
Food computers shut down
MIT Media Lab's Open Agriculture Initiative built "personal food computers": boxes the size of a small refrigerator with LEDs, sensors, pumps, fans, control electronics and a hydroponic tray, run by plant "recipes" that machine learning was meant to improve. After allegations that its leader, Caleb Harper, had told staff to show the boxes with plants not grown in them, and that nutrient solution had been discharged into a well above state limits, MIT investigated and closed the initiative on 30 April 2020, IEEE Spectrum reported.
LessonClaims for growing hardware hold only for plants actually grown in it, shown to someone outside.
Credit
MIT Media Lab (initiative); IEEE Spectrum (reporting)
2022
Gas prices darken greenhouses
Statistics Netherlands (CBS) reports that Dutch gas and electricity prices rose by more than 200% in 2022, driven by the war in Ukraine and the ban on Russian gas imports, which "mainly affected horticulture under glass". By CBS figures, Dutch greenhouse vegetable output fell 7.9% to 1.8 million t, with tomato yields per hectare down 12.5% and cucumber yields down 9.1% as growers delayed planting, and the volume of plants and flowers fell 6.7%.
LessonGrowing that relies on lamps and heating is only as cheap as the energy behind it, so control software is now judged on energy as well as yield.
Credit
Statistics Netherlands (figures)
Watch for
AI climate control in Ontario
Koidra, a software company, reports from trials at Great Lakes Greenhouses in Ontario, part-funded by the province's Independent Electricity System Operator (IESO), that its AI raised eggplant yield by 28.5% over a grower-run zone in 2022, raised cucumber yield by 16.1% on the same electricity, and in a third trial used 24.7% less heating energy. A separate measurement and verification (M&V) report looked only at that third trial, in which the AI zone was planted later, received about a third less light and harvested 13% less (7.55 against 8.69 kg/m²); the final report, written by Koidra, rebuts its method.
Achieved whenan outside reviewer confirms a yield or energy gain in a trial where both zones are planted on the same day and judged by measures agreed before planting.
Claimed by
Koidra and Great Lakes Greenhouses, July 2025
Open
01
An AI grower in simulation
A software agent built on a large language model that runs a simulated tomato greenhouse for a full season, setting heating, vents, carbon dioxide, screens and lamps every hour, and a public scoreboard comparing it with the simulator's rule-based controller and with reinforcement-learning controllers.
The briefCredit · unclaimed02
Fair tests for climate software
A written trial method and open analysis code for side-by-side tests of greenhouse control software: how to pair compartments, what to fix before planting, and which results to report (yield per square meter, yield per mole of light received, heat and power per kilogram), each with its uncertainty, so a seller and a buyer can agree in advance what counts as a gain.
The briefCredit · unclaimed
Open 01 · the brief
An AI grower in simulation
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A software agent built on a large language model that runs a simulated tomato greenhouse for a full season, setting heating, vents, carbon dioxide, screens and lamps every hour, and a public scoreboard comparing it with the simulator's rule-based controller and with reinforcement-learning controllers.
Why it is still open
GreenLight-Gym (van Laatum, van Henten and Boersma, 2024, revised 2025) is an open simulator for testing greenhouse controllers, built for reinforcement learning, in which software learns by trial and error over many simulated seasons. Language-model agents have run simulated crops before: AgriAgent (Qiu and colleagues, Smart Agriculture, 2026) set irrigation, lighting, heating and carbon dioxide in the DSSAT crop model for five crops, tomato among them, but was compared only with leaving the crop unmanaged and scored on yield alone, with no energy cost. Rasheed, Ravik and San (2025) showed that language models given only a prompt can produce workable control actions. Web and GitHub searches found no language-model agent scored on GreenLight-Gym against its rule-based and reinforcement-learning controllers.
Built with what exists
GreenLight-Gym (open-source Python); an AI model that can call tools, such as Meta's Muse Ember, which an outside firm's tests put level with Claude Haiku 5.5 at a third of the cost per task (see the Note "The jobs agents repeat"), or a small model fine-tuned on one graphics chip, such as Liquid AI's LFM3 (see the Note "One GPU to fine-tune").
First step
Install GreenLight-Gym, run its default controller for one season, and write the hourly loop that reads the simulator's state and returns set points.
Done looks like
Ten simulated seasons with different weather years, with the agent's net profit (crop sales minus the cost of heat, power and carbon dioxide) and its count of climate-limit breaches published beside the rule-based controller's and a reinforcement-learning baseline's, plus the model-call cost per season.
Open 02 · the brief
Fair tests for climate software
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A written trial method and open analysis code for side-by-side tests of greenhouse control software: how to pair compartments, what to fix before planting, and which results to report (yield per square meter, yield per mole of light received, heat and power per kilogram), each with its uncertainty, so a seller and a buyer can agree in advance what counts as a gain.
Why it is still open
In the Ontario trials the software company and the M&V reviewer read the same trial in opposite ways, and the company's own report says metrics "should be discussed and finalized before the conclusion of the trial" and that M&V planning should run "in parallel with trial design". Web searches found trial designs from single projects, such as Wageningen University & Research's Autonomous Greenhouse Challenge, but no published standard method for such comparisons. The Challenge's own climate and crop data are public on 4TU.ResearchData.
Built with what exists
Python with pandas and statsmodels; the open climate, harvest and image data from four rounds of the Autonomous Greenhouse Challenge on 4TU.ResearchData, to test the method on real results; the International Performance Measurement and Verification Protocol (IPMVP) for energy savings, as a model to adapt.
First step
Download the third Challenge's lettuce climate data and recompute each team's published result.
Done looks like
The method and code published and applied to all four Challenge rounds and the Ontario trials, reproducing each published headline result or showing in numbers why it differs, and at least two growers or software companies registering a trial under it before planting.