Could Robots Feed Each Crop Plant Only What It Needs?

The problem

Camera-based targeted spraying already makes a decision for every patch of ground: weed or no weed, spray or skip. The AI in Practice review of that technology rated it "In early use". One vendor reports more than 5 million US acres in 2025 (vendor-reported), roughly 3% of US corn and soybean acres, but measured herbicide savings shrink sharply under heavy weed pressure and independent trials remain short and small¹. This paper looks one step further, at a different input and a different problem: whether the same plant-by-plant precision could be turned from killing weeds to feeding the crop, so that each plant receives the nitrogen it needs rather than a field average.

The scale of the waste is well documented. Nitrogen recovery efficiency is the share of applied fertiliser nitrogen that ends up in the crop rather than in water, air or soil. A 2023 meta-regression in Nature Communications, drawing on 2,436 paired field observations from 407 studies, put the current global mean at 48% (measured in controlled field trials, synthesised, 2023)². The rest is the problem. The European Commission's Farm to Fork Strategy identifies excess nitrogen and phosphorus as a major source of air, soil and water pollution and of climate impacts. It commits to cutting nutrient losses by at least 50% and fertiliser use by at least 20% by 2030 (official target, set May 2020)³, and names the wider use of "precise fertilisation techniques" as one route³.

Why a field average fails. Soils and plants vary across a field, so a single rate overfeeds some places and underfeeds others. The same meta-regression found that site conditions explained 57–65% of the variation in how much a management change improved nitrogen recovery (measured, meta-regression, 2023)². In plain terms, what works depends heavily on where it is done. The evidence is uneven by region: 77% of the underlying studies were in Asia, 14% in North America and 4% in Europe².

What already works

The research found three building blocks, each doing part of the job.

Dosing each seed at planting. John Deere's ExactShot senses each seed as it enters the soil and sprays about 0.2 ml of starter fertiliser onto it instead of a continuous stream down the row. Deere says this cuts starter fertiliser by more than 60% (vendor-reported, January 2023)⁴. Independent field work is more sobering. Beck's Hybrids, a seed company that runs its own Practical Farm Research trials, tested ExactShot over three years at several locations. Dosing onto the seed was yield-neutral compared with a continuous stream and let growers use 2–4 gallons per acre instead of 6. But the three-year return on investment was negative for every starter treatment, and the best ExactShot rate averaged −$0.82 per acre in 2025 after a $5 per acre technology fee (measured in a controlled field trial, seed-company research, not peer-reviewed, 2023–2025)⁵. Precision cut the input; it did not, on its own, make the input pay. Every seed also receives the same dose, so this is precision of placement, not a judgement about what each plant needs.

Robots that return to each row all season. Upside Robotics, an Ontario company, runs small autonomous robots between 30-inch corn rows, applying liquid nitrogen in small weekly doses from a solar-powered base station in the field. About 20 robots ran autonomously on Ontario farms in 2025, against one remotely operated robot in 2024⁶. The company reports that its 2024 trials cut fertiliser by 30–70% without affecting yield (operator-reported, 2024)⁶. The system depends on soil testing and soil mapping on a 3.5-metre grid, from which software estimates how much nitrogen the soil itself is releasing, so that the robot applies only the difference⁶. Participating farmers grew from 7 in 2024 to 11 in 2025 and 31 in 2026, and the company says its robots make up to 20 applications from emergence to near harvest (operator-reported, 2026)⁷.

Sensing and dosing a single plant. At the Universidad Politécnica de Madrid, researchers built a proof-of-concept platform that uses three lidar scanners (laser distance sensors), a multispectral camera and a robotic arm to find each cabbage in a strip-cropped organic field and aim fertiliser at its base. In field tests it located itself to within about 11 mm over the first 120 cm of a row (measured in a field test without a control group, published December 2020)⁸. The set-up was manually operated and moved at 1 km/h, and the paper reports positioning accuracy, not fertiliser saved or yield⁸.

What does not exist yet. The research found no peer-reviewed field trial in which a robot set each plant's nitrogen dose from that plant's own measured need and then measured nitrogen saved, yield and nitrogen lost against a uniform rate.

The possibility

One way this could work combines the three blocks. At planting, a per-seed doser would give every plant a small, identical start. Through the season, light robots would pass along each row repeatedly, reading each plant's size and colour with cameras, combining that with a soil map, and topping up only the plants falling behind. Late doses could reach the crop after it has grown too tall for a conventional sprayer; Upside says corn still needs about 40% of its nitrogen after tasselling (operator-stated)⁷.

The farmer would not necessarily own the machines. A contractor or the robot maker could run them and charge per acre, a service model that robotics entrepreneurs have discussed because it spares farmers the risk of owning technology that may quickly become obsolete⁹. The first users would plausibly be corn growers, for whom nitrogen is one of the largest costs⁶, and later high-value vegetables, where single-plant robots have already been tested⁸.

If such a system worked, the change would be less about the robot than about the rhythm of fertilising: from two or three large passes based on a field average to many small ones based on each plant's response. The 2023 meta-regression models a rise in global mean nitrogen recovery from 48% to 78% if the best combinations of nutrient, crop and soil practices were adopted (projected by the source)². Per-plant robots would be one possible tool for the right-rate, right-time, right-place part of that gain, not the whole of it.

What would have to be true

1. The savings would have to be real, not a measure of past over-fertilising. In Ontario, some areas where equipment downtime meant no nitrogen was applied still yielded above 200 bushels per acre (operator-reported, 2025)⁷. In a separate reduced-nitrogen programme in 2025, four of seven farmers cut nitrogen on test plots, lost one or two bushels per acre and gained $13–23 per acre in profit, with no robots involved (self-reported, programme data)⁷. If the soil was already supplying the nitrogen, a robot that applies less of it has saved money but has not yet shown it is smarter than a cheaper test strip.
2. Cameras would have to read nitrogen need, not just plant size. A plant can look small or pale because of drought, disease or compaction as well as hunger. The commercial system leans on detailed soil maps alongside its cameras⁶, and the research platform measured position, not need⁸. Because site conditions explain most of the variation in what improves nitrogen recovery², the decision rule would have to be calibrated field by field.
3. The robots would have to keep working through a whole season. Upside's own account lists corroded electronics from leaking fertiliser tanks, causing significant downtime in 2024, antennas snapping off in tall corn and recurring connectivity problems; some acres did not receive the intended nitrogen (operator-reported)⁷. A missed weed pass is visible within days; a missed nitrogen dose would show up only at harvest.
4. The numbers would have to work on ordinary farms. The independent ExactShot trials found lower input use but a negative three-year return once the technology fee was counted⁵. A systematic review of 18 economic studies of crop robots found that all identified some profitable scenario, but most used partial budgets built on design specifications or prototype data rather than farm experience (peer-reviewed review, 2019)⁹. Upside says its system is most efficient on fields of 50–100 acres (operator-stated, 2025)⁶.
5. The rules would have to let robots work without a person beside them. The same review noted that some EU countries required a human supervisor to be physically present while a crop robot works, which would remove much of the labour saving (2019)⁹. The EU Machinery Regulation, which is designed to cover autonomous mobile machinery, applies from 20 January 2027¹⁰.
6. The suppliers would have to survive long enough to learn. Funding for agricultural robot hardware has proved fragile, as the next section shows¹¹.

What could stop it

What the evidence shows

A pioneer failed. The UK's Small Robot Company, which developed small robots to map, weed and plant arable fields, entered liquidation in February 2024, with Kroll appointed as administrator on 1 February. It had raised a reported $13.18 million over six funding rounds and grown to more than 50 staff, but a signed investment did not arrive before its money ran out; the company said agriculture is perceived by investors as very risky¹¹.

Tuning for savings has a cost elsewhere. In the anchor technology, a three-year Arkansas trial found that the least sensitive sprayer setting let far more weeds survive: harvest escapes averaged 1,380 plants per hectare against 268 for broadcast spraying (measured in a controlled trial, 2022–2024)¹². A nitrogen robot would face the same dial between saving input and protecting the crop, with the difference that under-feeding shows up only in yield.

Weather moves the result. In the 2025 reduced-nitrogen programme, two farms hit by drought lost about 5 bushels per acre and $10–12 per acre in profit on their low-nitrogen plots (self-reported, programme data)⁷.

Environmental gains are assumed, not measured. None of the economic studies in the 2019 review estimated environmental impacts, and its authors noted that robots may or may not be a cost-effective way to reach environmental goals compared with other methods⁹.

Reasoning: the challenge

The strongest objection is that much of the prize may be available without robots. Splitting applications, testing reduced rates in strips and insuring farmers against the downside are cheap, and the reduced-nitrogen programme above produced profit gains with none of the hardware⁷. If that holds more widely, per-plant robots would earn their place only where they do something those methods cannot, such as late-season doses in tall crops or fine-grained response on very variable fields.

The risk would also sit unevenly. The farmer bears the yield loss if a robot misreads a plant or breaks down at the wrong moment, while a per-acre fee is charged whatever the result; a contract that tied fees to measured savings would align incentives better. The plant-by-plant data the robots collect would be valuable, and who owns it would matter. Finally, the evidence is narrow: Ontario corn, one research platform in Madrid, and one seed company's trials including Illinois sites. Nothing comparable was found for Europe's large-scale arable systems.

Signs to watch

• Ontario 2026 harvest results. Upside's 31-farm season and the Ontario Soil Network's reduced-nitrogen plots report after this harvest; the test is whether downtime fell and whether robot fields beat reduced uniform rates⁷.
• First commercial sales. Upside aims to have products for sale by 2027 (operator-stated)⁶.
• Independent ExactShot data. The trigger is publication of a fourth year of Beck's Practical Farm Research results⁵.
• EU Machinery Regulation. It applies from 20 January 2027; watch for harmonised standards covering autonomous mobile machinery¹⁰.
• German single-plant research. The EXACT project at Martin Luther University Halle-Wittenberg is developing single-plant fertilisation for rapeseed and maize sown in equidistant patterns, running from January 2025 to January 2028 (search extract)¹³.
• A controlled, independent trial. The trigger is publication of a peer-reviewed field trial comparing a per-plant robot with a reduced uniform rate and measuring nitrogen losses as well as yield.