The next agricultural robot may spend more time watching what happens under a plant than driving across a field. At the University of Cambridge, researchers are using 3D-printed machinery and computer vision to study crop root worms, tiny parasites that invade plants beneath the soil. Their aim is to identify which plants naturally resist attack so future crops can be better protected.
These organisms are called plant-parasitic nematodes. Farmers may notice weaker growth or disappointing yields without immediately seeing the cause. A nematode can be difficult to spot without laboratory equipment, and studying its progress through a living root requires patience. That is what makes the Cambridge project’s combination of automation and detailed imaging significant.
The enemy that attacks where farmers cannot see
Many familiar agricultural threats announce themselves on a leaf or a fruit. Root parasites are less obliging. They operate underground, where the tissue being damaged is hidden from routine inspection. A farmer can respond to visible symptoms, but symptoms do not necessarily reveal which pest is responsible, when infection started or how a particular plant’s biology contributed.
For researchers trying to breed resistant plants, those unknowns create a slow, labour-intensive problem. Many plants have to be observed and compared, and the organisms affecting their roots may change over time. If only a small number of samples can be examined each day, finding reliable patterns in plant resistance is difficult.
The Cambridge team says its earlier screening capacity was about 90 plants per day. With the new automated system, it reports examining around 1,000 plants per hour. Those are the research group’s figures from a particular laboratory workflow, not a promise that commercial farm operations will suddenly work at the same rate.
What matters is the change in the kind of experiment scientists can attempt. Faster observation allows them to compare more plant varieties and watch how infection develops over time. It may reveal a pattern that would have been missed or taken an impractically long time to discover using manual inspection alone.
How machines study crop root worms over time
The system combines custom 3D-printed hardware with AI-powered image analysis. The machinery helps researchers collect images in a repeatable way; computer vision assists in interpreting what those images show. The combination is valuable because collecting a large number of pictures does not automatically make the biology understandable.
Cambridge researchers also describe using four-dimensional imaging, meaning observations that track the parasite over multiple days rather than treating each image as an isolated snapshot. Time matters here. A worm entering a root, growing and affecting the plant’s tissue is a process, not a single event. Watching that progression can help scientists connect a plant’s characteristics to its response.
The project’s plant-parasite research group studies the interaction between parasites and their hosts. Its objective goes beyond counting worms. Researchers want to understand which genetic characteristics make a plant susceptible and which may offer useful resistance. The new imaging method gives them another way of asking that biological question.
A key limitation is that the research is currently focused on model plants. Cambridge identifies food crops such as potatoes as a possible next area for application. It would be misleading to say that a new potato variety has already emerged from this system or that farmers can buy a commercial pest-finding robot based on the demonstration.
Why the goal is stronger plants, not simply more spraying
Farming has always involved managing pests, but a treatment that reduces a pest population is not the same as a crop that resists being harmed. Resistance is attractive because it can form part of protection before a visible outbreak forces a response. Researchers may investigate how different genetic traits alter the likelihood or severity of infection.
However, resistance is not necessarily permanent or universal. A plant variety may perform well against one parasite population and less well under another set of conditions. Breeding decisions also have to consider yield, growing conditions, taste, storage and other traits farmers and buyers need. Discovering a promising resistance marker is a starting point, not the end of crop development.
This is why laboratory screening complements rather than replaces field trials. A plant grown in controlled conditions is not exposed to the same mixture of soil conditions, weather and interacting organisms it faces outdoors. Researchers must eventually see whether the biological advantage remains useful under realistic growing conditions and across different environments.
The work has a different role from autonomous tractors and other field robots. Those machines aim to alter operations such as driving, sensing or tending crops. The Cambridge system concentrates on the earlier scientific task of discovering useful plant characteristics. Both can contribute to agriculture, but at very different points in the process.
Such a system is likely to matter most when biological comparisons must be repeated many times. A promising experimental plant is not convincing because it happened to look healthy in one picture. Researchers need to understand variation between samples, whether parasite counts change over time and how much of a difference is explained by plant characteristics. Speed matters only when it is paired with measurement quality and sound experimental design.
Why AI is useful when the images multiply
Moving from a modest set of observations to continuous high-volume imaging creates a data problem. A human expert could carefully inspect individual samples, but a large research programme needs consistent methods for interpreting many images. Computer vision can help identify features and measurements while leaving biological conclusions to researchers who understand the experiment.
Accuracy remains a critical question. If software systematically mistakes a root structure for a parasite, a larger dataset may simply produce more errors. Researchers need well-labelled examples, checks on difficult samples and repeatable results. They must be able to explain how the imaging tool performed before relying on differences between plant varieties.
Similar challenges appear in automated systems used around orchards, where machines have to distinguish fruit, flowers and other features of a complicated natural scene. Nature does not arrange itself like objects on a factory conveyor belt. The setting may be different, but dependable visual recognition remains central.
Potatoes provide an accessible example of why this work could be useful. A grower cannot choose a variety solely because it tolerates one pest; the crop still needs to grow well, survive storage and meet the market’s expectations. If screening helps breeders identify resistance sooner, they can incorporate that information into wider breeding programmes. The subsequent testing and commercial decisions remain substantial, but better early evidence can help narrow which possibilities deserve the most work.
The system could also change how experiments are repeated. When many samples are processed through the same imaging workflow, researchers may be better able to identify differences caused by the plant rather than differences in how an individual technician collected a picture. Standardisation does not make results infallible, and image-analysis models still need validation. It can nevertheless improve the consistency of a demanding kind of observation.
There is a reason the team is investigating the molecular interaction between plant and parasite instead of merely hunting for visible damage. Understanding what makes infection successful can point towards biological mechanisms worth studying in many varieties. A rapid screening platform may therefore be useful even when a particular plant fails the first tests: negative and unexpected results help researchers refine their explanation of resistance.
From a laboratory breakthrough to the dinner table
The potential benefit to consumers is indirect but important. Better crop resilience could, over time, help farmers reduce some losses and manage uncertainty. It does not establish that food will become cheaper, that pesticides can be abandoned or that any particular harvest is now protected. These outcomes depend on what scientists discover and what farmers can use.
There is a wider pattern in AI applications throughout the food system: sometimes the headline involves a machine picking or sorting a product, while the most useful work occurs in research and analysis long before a product reaches a shelf. A faster method for finding resistant plants could be valuable even if ordinary shoppers never see the machinery.
The Cambridge results are striking because they change the speed and scale of observation. They do not yet tell us which resistant varieties will emerge, how well those varieties will perform or when growers could benefit. For a problem hidden in plant roots, being able to watch many more infections closely is a practical scientific advance in its own right.
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.