Robots weed ten million acres. They still cannot pick.

TL;DR. By May 2026 one Australian operator had more than 250 robots that had worked over ten million acres. A European seed-and-weed robot worked in 26 countries and weeded more than 26,000 hectares in the 2025 season. Laser weeding systems destroy 5,000 weeds per minute in commercial fields. This is not a pilot. It is deployed agriculture at scale. Meanwhile apple-picking robots manage one fruit every 5 to 10 seconds against a human's roughly one per second, and a greenhouse tomato system reports 86.7% success at 32.5 seconds per pick. The split is not between easy and hard tasks. It is between acting on the environment and acting on the crop, and where automation succeeded in agriculture, the plants were changed to suit the machines decades before the robots arrived.

---

Status: established, with market figures attributed. Deployment figures are from operator and industry reporting and are attributed where used. Harvest performance figures are from published system evaluations. Neither category has the regulatory reporting that autonomous vehicles do, so the numbers are less well-audited than that article's.

---

What is actually working

Worth being specific, because agricultural robotics gets discussed as a future and much of it is a present.

SwarmFarm, an Australian operator, had more than 250 robots that had collectively worked over ten million acres by May 2026, with the Australian Clean Energy Finance Corporation committing A$7 million in 2025 to expand production.

FarmDroid, a lighter solar-powered seed-and-weed platform, operated in 26 countries and weeded more than 26,000 hectares during the 2025 season.

Carbon Robotics laser weeders destroy 5,000 weeds per minute, raised $70 million in October 2024, and planned to scale manufacturing to 500 units.

Autonomous tractors run commercially for defined tillage, with major manufacturers extending to orchard spraying and retrofit kits for mixed fleets. Targeted spraying systems report chemical reductions of 60 to 80%.

These are working machines doing paid work at scale, and the scale is not modest: ten million acres is roughly the agricultural area of a small country.

What is not

Harvesting.

Apple-picking robots operate at one fruit every 5 to 10 seconds. A human picker manages roughly one per second. That is a factor of five to ten, and it is the gap that has kept commercial deployment marginal despite sustained investment and acute labour shortage.

A deep-learning greenhouse tomato system reported 86.7% success with an average pick time of 32.5 seconds. In a greenhouse, which is the most controlled agricultural environment that exists.

Strawberry harvesting remains substantially developmental after more than a decade of effort, in a crop with among the highest labour costs per acre in agriculture.

The economic incentive could not be stronger and the technology has not arrived.

The line is not difficulty. It is the target.

Look at what the working machines do and what the struggling ones do.

Weeding, spraying, tillage: the robot acts on the environment. Kill the thing that is not the crop. Apply chemical where the sensor says. Move soil. The target is a nuisance or a substrate, and the crop is defined negatively, as the thing to avoid.

Harvesting: the robot acts on the crop itself. Assess whether this specific fruit is ripe, which is a continuous judgement with no clean boundary. Grip it without bruising, which is force control on a deformable object whose properties vary between individuals. Detach it without damaging the plant, which will produce again.

Three properties make the second category harder in a way that speed alone does not capture.

Judgement is continuous, not binary. A weed either is or is not the crop. Ripeness is a spectrum, assessed differently by market, and getting it wrong is not symmetrical: picking unripe loses the fruit, leaving ripe loses it to spoilage.

Failure is irreversible. A missed weed is caught next pass. A bruised strawberry is unsellable, permanently, and the damage is often invisible at the moment of picking.

And every instance differs. Weeds vary, but the response does not: destroy it. Fruits vary in size, firmness, position, occlusion by leaves, and each variation changes the correct action.

Where the automation actually came from

This is the finding that ties agriculture to the rest of Territory 7, and it predates robotics entirely.

Row crops are heavily mechanised because the plants were changed to suit machines.

Uniform height so a cutting bar works. Simultaneous ripening so one pass harvests the field. Tough skins that survive mechanical handling. Determinate growth so the plant stops rather than producing continuously. The processing tomato was bred specifically for machine harvest in the mid-twentieth century, alongside the harvester, as a joint programme.

The machine did not learn to handle the plant. The plant was redesigned for the machine.

Specialty crops were not. Strawberries ripen unevenly and continuously, bruise under light pressure, and hide under foliage. Apples on a standard tree present at varying heights, angles and occlusions. These plants were selected over centuries for flavour, yield and appearance, by and for human hands.

Which makes agriculture the clearest case of the pattern this territory keeps finding. Industrial robots work because the workspace was engineered around them. Autonomous vehicles work because the operating domain is drawn and mapped in advance. Agriculture went further and engineered the organism.

Three things this establishes

Automation success tracks whether the target can be standardised, not whether the task is intellectually hard. Recognising a ripe strawberry is trivial for a person and hard for a machine; navigating a field is hard for a person and straightforward for a machine with GPS. The distribution of difficulty does not match intuition, and the deciding factor is variability in the thing being acted on.

Irreversible failure raises the required accuracy sharply. A weeding robot at 95% accuracy is excellent, because the remaining 5% is addressed next pass. A harvesting robot at 95% on a delicate crop damages one fruit in twenty, permanently. The same accuracy figure means different things depending on whether the error can be undone, which is the false positive cost argument in a physical setting.

And the environment can include the biology. Where the workspace cannot be rebuilt and the domain cannot be narrowed, agriculture changed the organism instead. That is a strategy unavailable to most fields and it explains a large share of what looks like robotic success.

What it does not establish

That harvesting will not be solved. Substantial capital has entered the sector, over a billion dollars into harvesting and weeding start-ups between 2022 and 2025 on industry estimates, and pick rates are improving.

That the deployment figures are audited. Unlike vehicle crash reporting, agricultural robotics has no regulatory disclosure requirement. Acreage and unit counts come from operators and industry analysts, and should be read as such.

That breeding for machines is costless. Crops bred for mechanical harvest have often been criticised for flavour and texture, and the processing tomato is the standard example on both sides of that argument.

And that labour displacement follows. The tasks being automated are ones with acute labour shortage, and the relationship between automation and agricultural employment is contested rather than settled.

What is unresolved

Whether pick rate closes or asymptotes. Five to ten times slower than a human is the current position. Whether that is a engineering trajectory or a limit imposed by force control on deformable objects is not known.

What the real field success rates are. Greenhouse figures like 86.7% come from the most controlled setting available. Open-field performance is less reported and is the number that matters.

Whether crops will be bred for robots again. The obvious response to hard harvesting is to change the plant, and some breeding programmes target machine compatibility. Whether consumers accept the result is a separate question with history behind it.

And what the durable economics are. Robotics-as-a-service models are spreading, which changes the capital question, and there is little public data on renewal rates.

The counter-argument

The comparison to human pick rate is the wrong measure. A robot works at night, does not tire, and does not require housing or seasonal visas. Five times slower over twenty hours beats one times faster over eight, and the relevant figure is throughput per day per dollar, not per second.

Greenhouse results are not a ceiling. Controlled-environment agriculture is expanding independently, and a system that works well in a greenhouse is commercially useful whether or not it transfers to open fields.

Breeding for machines is not a concession. It is how nearly all agricultural mechanisation happened, it is ongoing, and treating it as an admission that robots failed misreads the history. Co-design of crop and machine is the normal path, not a workaround.

And the environment-engineering framing may prove too general. If every success can be described as the environment being adapted, the claim risks becoming unfalsifiable. The test is whether it predicts which of the current harvesting efforts succeed, and that prediction has not yet been made or checked.

The short version

More than 250 robots working over ten million acres. 26,000 hectares weeded across 26 countries in one season. Laser systems destroying 5,000 weeds per minute in commercial fields. Agricultural robotics is deployed, not prospective.

And apple pickers manage one fruit every 5 to 10 seconds against a human's roughly one per second. A greenhouse tomato system reports 86.7% success at 32.5 seconds per pick, in the most controlled agricultural setting available. Strawberry harvesting remains largely developmental after a decade, in a crop with among the highest labour costs in farming.

The line is not difficulty. It is what the robot acts on. Weeding, spraying and tillage act on the environment, where the crop is defined negatively as the thing to avoid, judgement is binary, and a miss is caught next pass. Harvesting acts on the crop, where ripeness is a continuous judgement, the grip is force control on a deformable object that varies between individuals, and a bruise is permanent and often invisible at the moment it happens.

And the mechanisation that did succeed came from changing the plant. Row crops were bred for uniform height, simultaneous ripening and mechanical tolerance; the processing tomato was developed alongside its harvester as a joint programme. The machine did not learn to handle the plant. The plant was redesigned for the machine.

Which makes agriculture the clearest case of this territory's pattern. Industrial robots got an engineered workspace. Autonomous vehicles got a drawn and mapped domain. Agriculture engineered the organism.

Common questions

Is agricultural robotics actually deployed or still experimental? Deployed, at substantial scale, for a specific set of tasks. By May 2026 one Australian operator had more than 250 robots that had worked over ten million acres. A European seed-and-weed platform operated in 26 countries and weeded more than 26,000 hectares in the 2025 season. Laser weeding systems destroy 5,000 weeds per minute in commercial fields, and autonomous tractors run commercially for defined tillage. These are working machines doing paid work.

Why can robots weed but not harvest? Because weeding acts on the environment and harvesting acts on the crop. In weeding the crop is defined negatively as the thing to avoid, the judgement is binary, and a missed weed is caught on the next pass. In harvesting the machine must assess ripeness, which is a continuous judgement with asymmetric costs, grip a deformable object whose properties vary between individual fruits, and detach it without damaging a plant that will produce again. A bruise is permanent and often invisible at the moment of picking.

How much slower are harvesting robots than people? Apple-picking systems operate at roughly one fruit every 5 to 10 seconds against a human picker's approximately one per second, a factor of five to ten. A deep-learning greenhouse tomato system reported 86.7% success with an average pick time of 32.5 seconds, in the most controlled agricultural environment that exists.

Why are row crops so heavily mechanised? Because the plants were changed to suit machines, largely before robotics existed. Row crops were bred for uniform height so a cutting bar works, simultaneous ripening so one pass harvests the field, tough skins that survive mechanical handling, and determinate growth. The processing tomato was developed specifically for machine harvest in the mid-twentieth century alongside the harvester itself. The machine did not learn to handle the plant; the plant was redesigned for the machine.

Does the accuracy figure mean the same thing in both cases? No, and this is the important part. A weeding robot at 95% accuracy is excellent, because the missed 5% is addressed on the next pass. A harvesting robot at 95% on a delicate crop damages one fruit in twenty, permanently. The same number means different things depending on whether the error can be undone, which is why irreversible failure raises the required accuracy sharply.

Is comparing pick rate to a human fair? It is the standard comparison and it is not the whole picture. A robot works at night, does not tire, and does not require seasonal housing or visas, so throughput per day per dollar may favour a slower machine. That is the strongest counter-argument, and it depends on capital cost, utilisation across a season, and maintenance, none of which is well reported publicly.

How reliable are these deployment numbers? Less reliable than the vehicle figures in this series. Autonomous vehicles report crashes under regulatory obligation; agricultural robotics has no equivalent disclosure requirement. Acreage and unit counts come from operators and industry analysts, and should be read as company-reported rather than audited.

What would change the picture? Open-field harvesting success rates, which are far less reported than greenhouse figures and are the number that matters. Whether pick rate is on an engineering trajectory or approaching a limit set by force control on deformable objects. And whether breeding programmes targeting machine compatibility produce crops consumers accept, which is the historically proven path and also the one with the longest record of complaints about the result.