20 hectares per robot per season: what changes when the machine hoes instead of sprays

Both sources had the same robot in the field, the solar-powered drilling and hoeing robot FarmDroid FD20, and they measured different things.

The LfL, in its project “Testing, evaluation and optimisation of automated methods for mechanical weed control”, trialled four machines from 1 January 2020 to 31 December 2022: the FD20 in sugar beet, the Naïo Oz 440 in maize, the camera-guided Robovator hoe and the Agrointelli Robotti 150D carrier platform. Funding came from the Bavarian State Ministry for Food, Agriculture, Forestry and Tourism. Measured were yield, residual weeds and labour time.

Südzucker has for several years been running alternative weed-control methods in sugar beet on its Kirschgartshausen experimental farm near Mannheim under Dr. Peter Risser; the 2022 results were published on 31 May 2023. Five variants were compared side by side: broadcast spraying, hoeing with band spraying, hoeing with harrow, the FD20 hoeing alone, and the FD20 with a spot-spray unit. Measured above all was one thing: how many weeds are left at the end.

One source answers whether the robot gets the field clean. The other, what it costs and saves the farm. You need both answers, and neither is enough on its own.

Weed pressure at Kirschgartshausen in 2022 was very high, at 608,000 and 811,000 weeds per hectare at the final count. The results have to be read against that background.

• Broadcast spraying, four post-emergence applications: 97 and 99 percent control in the two trials. The best variant.
• FD20 mechanical only: 26,000 weeds per hectare left standing. Against the initial pressure, arithmetically around 96 to 97 percent.
• FD20 with spot spray: 19,000 plants per hectare, which corresponds to 98 percent control.
• Hoeing with band spraying, hoeing with harrow: according to the report between 68,000 and 111,000 weeds per hectare — the tractor-based hoeing variants were thus well behind the robot.

The robot with spot spray thus came close to the level of broadcast spraying and, according to Südzucker, saved “over 90 percent” of the plant protection product used compared with it.

That is the sentence that reframes the question. The robot did not get the field clean without a sprayer; the mechanical-only variant sat behind both chemical ones. What it changed is the area that is sprayed: no longer the whole field, but the strip immediately beside the beet that the hoe cannot reach. The product stays, the quantity drops by more than nine tenths.

Südzucker names the drawback in the same report without ceremony: the slow working speed of about 1 km/h at a maximum working width of 3 metres limits the output of one robot to around 20 hectares per crop and season. And: the robot needs intensive supervision despite its autonomy.

For a farm that is the more important of the two statements, because it fixes the unit of decision. A sprayer is a farm decision; it goes over every field. A robot with 20 hectares is a field decision. Anyone with 60 hectares of beet does not decide “robot or sprayer” but which 20 hectares the robot gets and what happens to the other 40 — or whether it has to be three robots, with three times the supervision.

And the supervision is no afterthought. A machine that stands on a field for weeks at 1 km/h is not driven, but it is visited: after the rain, after the deer, after the stone. Anyone who does not carry those visits in the labour calculation prices the robot too cheaply. What a pilot with a machine of this kind can prove and what it cannot is set out in the article on pilot projects for service robots in buildings; the logic in the field is the same.

LfL measured the farm side, and its three findings on the FD20 in sugar beet pick up exactly where the Südzucker report stops.

Yield. The yield was on the level of the organic reference variant. Against the method it replaces in organic farming, then, the robot cost nothing.

Labour. The saving across the trial years was 12 to 200 hours per hectare. The range is not a measurement error; it is the result. It depends on how much hand labour the reference method needed in the year in question. In a year of high weed pressure the robot replaces many hours of hand hoeing, in a quiet year few. Anyone pricing the robot therefore prices it over several years, not over the season in which the quotation arrives.

Break-even. Around 8 hectares, with subsidy. That is low, and it fits the 20-hectare capacity from Kirschgartshausen: between 8 and 20 hectares lies the range in which a single robot works economically on one farm.

One LfL result falls outside the pattern and is for that very reason the most useful for planning. In the first trial year the FD20 left “markedly higher residual weeds” than the tractor-based reference. In the following years the result was comparable.

The machine was the same in all three years. What changed was how it was handled: the drilling date, the setting of the hoeing unit, the distance to the beet, the response to the first rain after emergence. Those are not robot properties. They are farm routines, which a farm has to develop first for a machine it has never had.

For procurement that yields a simple rule: put the first year on a field whose result the farm can absorb, not on its best one. And read the first year's residual weeds as what they are — the price of learning, not the performance of the machine. That the learning stays with the operator rather than happening in the manufacturer's training centre is described for building robots; in this respect a field is a building without walls.

One detail of the method changes work organisation more than any performance figure. The FD20 drills and hoes from the same reference: at drilling it records the position of every seed by RTK and later hoes against those positions — in the row and between the rows. The hoe does not need to see the plant; it knows where it is.

That is the technical basis for the hoe getting closer to the beet than a camera-guided implement, and the reason the spot spray at Kirschgartshausen could be placed so precisely. But it has a consequence for the workflow: the robot can only hoe what it has drilled. Drilling with the farm's existing precision drill, which is four times faster, falls away for the robot's fields. The robot therefore blocks not only the hoeing weeks in May but already the drilling weeks in March — at 1 km/h there too.

Conversely, the 20 hectares are not 20 hectares of hoeing but 20 hectares of drilling and hoeing on the same machine. Anyone planning the capacity plans both windows, and both fall in the very weeks in which the farm has the least time anyway.

Three caveats, because the numbers carry exactly as far as the trials reach.

These are sugar-beet trials. The robot has been trialled in maize (LfL: Naïo Oz 440, 75-centimetre row spacing, RTK navigation) and in vegetables as well, but the control and cost figures above apply to beet on two sites. A different crop with a different row spacing and weed spectrum is a different trial.

It is one machine. All the figures come from the FD20. A camera-guided hoe on a tractor, or a robot that does not drill itself, has a different capacity, a different accuracy at the plant and a different dependence on drilling. The statements about supervision and the first year are probably transferable; the hectare and percentage figures are not.

The break-even includes a subsidy. LfL gives around 8 hectares “with subsidy”. Without it the threshold is higher, and how much higher depends on a subsidy rate that changes. The value describes an order of magnitude, not a decision. Under which regime a field robot is placed on the market at all is a separate question and has an article of its own; here the subject was only what happens in the field.