Plant Phenotyping Robot: From Mendel's Peas to ABB Arms

A plant phenotyping robot measures what a plant looks like and how it grows, from height and stem thickness to leaf angle, colour and temperature, so that breeders can link those traits to genes. For most of the history of plant breeding that measuring was done by hand, which is slow, labour intensive and often destroys the plant. This post follows the job from Gregor Mendel's pea garden to conveyor belt greenhouses and a 30 ton field gantry in Arizona. Then it looks at this autumn's research robots and at where ABB arms and mobile robots could fit into plant breeding.

Young poplar trees in pots on the conveyor system of ORNL's Advanced Plant Phenotyping Laboratory, 2020. Photo: Genevieve Martin, Oak Ridge National Laboratory / Wikimedia Commons (CC BY 2.0)

From Mendel's pea garden to conveyor belt greenhouses

Mendel did his measuring with his own eyes. Between 1856 and 1863 he grew and tested some 28,000 plants, most of them peas, in the 2 hectare experimental garden of St Thomas's Abbey in Brno. He followed seven traits, among them seed shape, flower colour, pod form and the height of the plant, and presented the results at two meetings of the Natural History Society of Brno in February and March 1865. The paper was published in 1866. The word for what he was recording came later: Wilhelm Johannsen proposed the distinction between genotype and phenotype in 1911.

Gregor Mendel, photographed around 1862, when he was growing and scoring pea plants in Brno. Photo: unknown photographer via Wikimedia Commons (public domain)

Measuring stayed the slow part of breeding. The TERRA-REF project at the University of Arizona says breeding is limited by the speed at which phenotypes can be measured, and that current instruments do not scale to the thousands or tens of thousands of plants a breeding programme has to evaluate. Companies such as LemnaTec, based in Aachen in Germany, set out to do the job with cameras and image analysis. LemnaTec has worked on plant phenotyping since 1998. It started with imaging tests on duckweed, moved on to seedlings and plant parts, and says its systems have gone to customers in more than 30 countries.

A scientist checks rice plants in a greenhouse at the IAEA plant breeding unit in Seibersdorf, Austria, 2000. Photo: IAEA Imagebank / Wikimedia Commons (CC BY-SA 2.0)

Greenhouse platforms bring the plants to the cameras. At the Bavarian State Research Center for Agriculture in Freising, a LemnaTec conveyor system that the staff call the moving fields facility has run almost continuously since April 2013, measuring above and below ground growth of barley, maize and potatoes. The Bellwether Foundation Phenotyping Facility uses a conveyor in a 70 square metre growth house to carry up to 1,140 plants per experiment to colour, fluorescence and near infrared imaging cabinets. Out in the field the sensors move instead. At Maricopa, Arizona, LemnaTec's Field Scanalyzer drives a 30 ton steel gantry along two 200 metre rails, imaging the crop below.

Conveyor lines carry potted plants through Purdue University's Controlled Environment Phenotyping Facility during a tour by USDA Secretary Sonny Perdue, April 2019. Photo: USDA via Wikimedia Commons (public domain)

Field robots, cheap gantries and ABB arms

Rail gantries give repeatable data but need costly permanent infrastructure, and drones see only the top of the canopy, so researchers keep building ground robots. On 2 October, Bioengineer reported on a self-driving vehicle from a team including Si Yang and Xinyu Guo that straddles two rows of maize on a 145 centimetre wheelbase and carries three synchronized depth cameras. It maps plants in real time on an ordinary CPU, while a structure from motion pipeline needed about 36 hours for 1,300 frames. Checked against hand measurements of 259 field plants, it estimated height with an error of about 41 millimetres, and its plant counts over nearly 3,900 plants had a mean accuracy of 94.55 percent.

A LIFT fluorescence sensor on a semi-automated phenotyping platform measuring winter wheat grown under elevated CO2. Photo: Knoli / Wikimedia Commons (CC BY-SA 4.0)

Price is the other pressure. On 6 October the same site described an open source phenotyping robot from Universidad de la República in Uruguay that costs 1,110 US dollars. It is a two axis gantry that moves a colour camera and a thermal camera over up to 24 potted plants in a growth chamber, so the plants stay where they are. Commercial systems such as PSI's PlantScreen or Phenospex's PlantEye F500 can exceed 100,000 dollars and approach a million in some setups, the article says, with conveyors, robot arms and pot handling adding much of the cost. Labs are automating the bench as well: Cornell's HyperBird, reported in September, checks hundreds of grape leaf samples in a few hours.

Poplar trees on the phenotyping system at ORNL, each pot sitting on its own carrier. Photo: Genevieve Martin, Oak Ridge National Laboratory / Wikimedia Commons (CC BY 2.0)

Robot arms already appear in some platforms. Iowa State University's Enviratron, described in 2019, sends one rover through eight growth chambers. The rover carries a Universal Robots UR10 arm with a 1.3 metre reach that takes overhead and side views and pushes a probe toward the leaf surface. We found no ABB robot in a phenotyping platform, but the same layout could be built from ABB parts. ABB says its GoFa CRB 15000 cobot can work directly alongside people without fencing, with protection up to IP67 and path accuracy down to 0.03 mm with the Ultra Accuracy option. Its Flexley Mover AMRs navigate with Visual SLAM and take different top modules, so a GoFa riding on one is a possibility.

An ABB GoFa cobot with its FlexPendant at a work table, shown at the launch of the Ultra Accuracy option in 2024. Photo: ABB

What plant phenotyping robots still have to solve

In breeding companies and research greenhouses, the next step is trusting the machine's scores. A 2025 study in Plant Methods, summarised by LemnaTec, compared automatic phenotyping with human raters in a screen of radiation induced Arabidopsis mutants. The camera system was more accurate overall but flagged more false positives, so the researchers suggested using automation as a first pass and sending a smaller set of plants to human experts. That division of work suits robot cells, which can image plants 24 hours a day, as the Enviratron does, while breeders check the shortlist. Seed testing went through a similar change, covered in our post on seed lab automation.

A member of Purdue's phenotyping team talks to visitors beside plants on the conveyor, April 2019. Photo: USDA via Wikimedia Commons (public domain)

The maize robot's authors list the problems still open in the field. Drift builds up over long runs and will need loop closure or RTK satellite positioning, the camera calibration can loosen with vibration and temperature swings, and the simple colour mask cannot tell maize from large green weeds. They plan deep learning segmentation and visual inertial navigation for the next version. Greenhouse lines face a mechanical choice: plants can keep riding conveyors to fixed cameras, gantries can carry the sensors to the plants as in Uruguay, or mobile arms can drive from bench to bench as in the Enviratron. ABB pitches GoFa as an alternative to the 2D gantries used in electronics manufacturing, saying the arm offers full 6D motion across its workspace without extra calibration work.

A GoFa cobot with ABB's Ultra Accuracy option, here fitted with a measuring probe. Photo: ABB

A commercial line with conveyors and arms costs far more than the Uruguayan gantry, and smaller labs will weigh that against how many plants they need to measure. Skills are the next limit. Every system needs people who can calibrate cameras, label training images and keep data pipelines running; the Uruguayan team trained its plant segmentation network on 100 hand annotated images. Mobile arms that work among greenhouse staff also need collaborative safety, such as the torque sensors ABB builds into GoFa to limit power and force. Cells like these are usually laid out and tested in simulation first, for example in RobotStudio.

An ABB Flexley Mover P603 carrying crates. ABB says the Flexley Mover range works without physical barriers around it. Photo: ABB

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com