The weed-spraying robot in a new field experiment did something that conventional broad spraying cannot easily do: it aimed at unwanted plants without treating the entire patch of farmland as a target. Working on terraced fields in the Indian state of Uttarakhand, the researchers reported using markedly less herbicide than a blanket-spraying comparison. That is the appealing prospect behind precision agriculture: protecting the crop while putting fewer chemicals where they are not needed.
In a peer-reviewed paper published in Scientific Reports on 15 September 2026, the team describes a compact robotic platform equipped with computer vision and a selective spraying system. Three paired field trials covering 0.7 hectares yielded an average herbicide-volume reduction of about 65 per cent compared with blanket treatment. The experiment also recorded misses and spray reaching places it was not intended to reach. Those qualifications matter because a working prototype is not yet a guarantee of lower chemical use on every type of farm.
Why a weed-spraying robot needs to see individual plants
Traditional spraying treats a field or crop row according to a planned coverage pattern. That can make sense when weeds are widespread, but the approach also means applying chemicals to ground where there may be nothing to kill. The idea behind site-specific spraying is to recognise the troublesome plants and trigger the nozzle only at those locations. The camera and software effectively decide whether the next patch of ground deserves treatment.
That is much harder than identifying a green object against bare soil. Crops and weeds are both plants. They overlap, cast shadows, change appearance through the growing season and can be partly hidden behind one another. Wind, mud and slopes make the problem more awkward. Researchers have to connect the act of recognising a weed to the physical timing of a moving sprayer. A correct label that arrives after the nozzle has passed is not especially useful.
Steep terrain turns a laboratory task into farm work
The Uttarakhand team designed its experiment around hilly agricultural land, where large machinery may be unsuitable and evenly treating irregular terraces can be difficult. The robotic base was intended to move across uneven ground while the camera identified crops, weeds and background vegetation. A controller then coordinated the spray mechanism with the detected position of a target. The trial therefore examined more than a classification model sitting on a computer.
The published work reported a 92.4 per cent spraying hit rate in the tested setup. It also measured off-target drift of 7.1 per cent, a reminder that sending liquid through a nozzle introduces practical uncertainty even when the visual decision is correct. Field measurements used water-sensitive paper to judge spraying outcomes. The experiment's herbicide saving was recorded against a matched blanket-spraying comparison rather than guessed from how many weeds the camera recognised.
A limitation is built into the sample size. Three paired field-level trials can demonstrate that a design functions under the tested conditions, but they do not reveal every complication that might emerge across seasons, soils, crops and farms. A prototype operating successfully on a small area may still need changes to its wheels, power supply, cleaning routine or detection model before it becomes a reliable agricultural product.
What happens when the camera gets a plant wrong
Every automatic detection system faces a trade-off. If it is too eager, it may mistake a valuable crop for a weed. If it is too cautious, it may leave unwanted plants untreated. The study compares different computer-vision approaches using measures of how often detections were correct and how often actual weeds were found. These are useful research indicators, but no single percentage captures the cost of making the wrong spray decision on a living crop.
The problem varies with the plants involved. A young crop can resemble a weed much more closely than a mature plant does. Leaves may overlap, and bright sunlight can create deep shadows between terraces. The team noted difficult cases where vegetation looked similar or where a target appeared briefly near the edge of the spray mechanism's reach. Those are precisely the details that determine whether an agricultural machine saves labour or creates extra work correcting its mistakes.
Could precision spraying genuinely reduce chemical use?
The trial provides a measured reason for cautious optimism: under its reported conditions, the robot used less herbicide while still targeting detected weeds. A separate 2025 study of weed detection in teff crops also explored how machine learning and aerial images might identify where herbicide is required. These are different technical routes to the same underlying question. Must a farmer treat a whole area when only particular plants need attention?
It would be premature to apply the percentage from the Uttarakhand experiment to all farms. Chemical savings depend on weed density, product choice, the accuracy of the model, weather and how the comparison treatment is delivered. The equipment itself has costs, and a farmer also needs to know whether reduced spraying affects subsequent weed growth or crop yield. A serious commercial assessment would consider all of those factors, not merely the volume of liquid released during a trial.
The environmental argument is still understandable. Better targeting may reduce unnecessary use, but that is not equivalent to proving a particular farm has no ecological impact. Herbicide chemistry, soil conditions and management practices remain important. The technology offers a tool for making a more selective decision, not an exemption from ordinary agricultural responsibilities.
Where this fits in the changing machinery on farms
Farm automation is not a single invention. Larger autonomous tractors attempt to remove some of the labour involved in driving and field operations. Other robots concentrate on much smaller tasks, such as finding organisms around crop roots. The selective sprayer belongs to that second tradition: rather than replacing the whole farmer, it tries to make one costly and repetitive job more precise.
There is a reason these machines are being designed for real terrain and awkward conditions. A field is not a tidy factory floor. Equipment must tolerate dust, uneven surfaces, wet leaves, plant growth and the consequences of incorrect decisions. The ability to explain what a machine saw and why it sprayed would also help an operator trust and maintain it. Farmers will want a reliable record of treated ground, not simply a striking demonstration video.
LiveAIWire has previously explored how digital tools are changing food production. This experiment adds a physical test of that idea. The camera identifies what matters, the robot acts only where needed, and the result can be measured against ordinary practice. If the approach becomes robust across crops and seasons, it could help make careful chemical use easier to achieve.
Farmers would also need a reliable way to intervene when the system is uncertain. An operator might reasonably want the machine to flag a difficult patch for review rather than applying a chemical dose based on a poor image. That becomes more important near crops, field boundaries and areas where unintended spraying could damage plants or harm nearby habitats. A machine that occasionally declines to act may be safer than one that insists on making a confident classification every time.
The economics will vary from farm to farm. Time spent setting up cameras, cleaning nozzles, maintaining a platform and checking disputed detections can reduce the gains from using less spray. On a large uniform field, one method may be practical; on a steep or irregular site, another may prove more useful. Those are operational questions to be answered by repeated field tests and farm accounts, not by the accuracy of a single image-recognition model.
The real test comes after the research demonstration
A useful next stage would involve larger, repeated comparisons across different farms, weather conditions and weed populations. Researchers would need to examine repairs, energy use, field speed, crop damage and long-term weed control. They would also need to discover whether a farmer can operate the system without specialist help whenever the camera fails to recognise a plant. These are not glamorous questions, but they determine whether a laboratory success becomes a trusted machine.
The new work has not made herbicides unnecessary, nor has it shown that a robot can care for an entire farm on its own. What it has shown is more concrete. A machine can distinguish many targets, move on difficult ground and reduce spraying in a small measured field experiment. That is a credible step towards agricultural technology that treats the weeds as the problem rather than treating every crop and every patch of soil alike.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.