Industries / Food, farming and forestry / Agriculture and farming / AgTech
Farm robots
Farm robots are machines sold to farms that drive, weed or tend crops with little or no human control.
6
Milestones achieved, from 1997 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1997
A windrower cuts hay alone
Demeter, a New Holland 2550 self-propelled windrower (a machine that cuts hay and lays it in rows) fitted with GPS, cameras and computer control, cut 100 acres (about 40 ha) of alfalfa in one continuous run in August 1997 without a driver steering it, stopping only to refuel. It planned the cutting for a whole field, followed the line of uncut crop with cameras and stopped if it saw an obstacle; Farmers Weekly reported in 1998 that sensor reliability was the main problem left to solve.
Credit
Carnegie Mellon University's National Robotics Engineering Consortium (Tony Stentz, Thomas Pilarski and colleagues) with New Holland
2017
Deere buys a thinning robot
Blue River Technology's LettuceBot photographed each lettuce seedling and sprayed only weeds and plants crowded too close together, a job called thinning that crews had done with hoes. It was sold as a pay-per-acre service and was thinning and weeding more than 10% of US lettuce production when John Deere agreed in September 2017 to buy the company for $305 million, by The Robot Report's account. Deere said the computer vision would help its machines perceive the crops they work on.
LessonBlue River chose to build the plant-recognition software and leave the heavy machines to an established maker.
Credit
Blue River Technology (Jorge Heraud and Lee Redden); John Deere
2017
Japan sells supervised robot tractors
In 2017 Japan's Ministry of Agriculture, Forestry and Fisheries (MAFF) issued safety guidelines for self-driving farm machines, and Kubota began selling tractors on the idea of "one driver, two tractors": a person drives one tractor and watches a second that works the same field without a driver. Researchers at NARO, Japan's national agricultural research body, write that full-scale commercial use of such supervised robot tractors began in 2018, that NARO began safety inspections of them that year, and that machines able to work without anyone in sight of the field were still in development.
Credit
Kubota; MAFF; NARO
2023
Weeding robots in one field
Südzucker, a German sugar maker, tested weeding robots against a conventional three-spray herbicide program in sugar beet on its research farm at Kirschgartshausen in April and May 2023. The spray program controlled 88% of weeds; the FarmDroid FD20, a solar-powered robot that remembers where it sowed each seed and hoes around it, controlled 90% alone and 96% with spot spraying, while the Farming GT robot reached 97%. Hand hoeing fell from 67 hours per hectare on untreated plots to 10–13 hours with the robots, but the FarmDroid moved at about 0.7 km/h, and Südzucker's head of research, Peter Risser, named cost and speed as the challenge.
LessonThe robots matched or beat herbicide on weeds; their speed and price, not their weed control, held them back.
Credit
Südzucker (Peter Risser); robots from FarmDroid, Farming Revolution and Strube; reported by Farmers Weekly
2024
A robot weeding service folds
Small Robot Company of the UK launched Tom, a robot that drove arable fields photographing every plant, on 29 April 2021; it covered 20 ha a day, and in one 6 ha field it recorded 12.7 million plants, 250,000 of them weeds, so that a second robot could kill each weed with an electric charge. The company planned to serve more than 100 farms in 2023. On 1 February 2024 its members resolved to wind it up in a creditors' voluntary liquidation, and the company was dissolved on 30 April 2026.
LessonA robot that mapped weeds plant by plant in the field still had to pay its way as a service before the money ran out.
agriland.iefind-and-update.company-information.service.gov.uk
Credit
Small Robot Company (a reversal)
2025
Lasers weed like herbicide
Carbon Robotics' LaserWeeder is a tractor-pulled machine that finds weeds with cameras and deep learning and burns them with lasers. In 2024 field trials at Rutgers University's research farm in Upper Deerfield, New Jersey, on spinach, peas and beets, it controlled weeds about as well as conventional herbicides, killing weeds about half a centimeter from seedlings without harming the crop; the study appeared in Pest Management Science in 2025, which Rutgers calls the first peer-reviewed study of AI-guided laser weeding on the US East Coast. The researchers note that it costs from about $500,000, works best on small weeds, needs several passes a season and runs at 1–2 mph (about 2–3 km/h).
Credit
Carbon Robotics (machine); Thierry Besançon (Rutgers) and Lynn Sosnoskie (Cornell)
Watch for
Tractors watched from afar
Kubota says its Agri Robo Tractor MR Series, to go on sale in Japan in April 2027, will be the first Japanese-made tractors able to work without a driver under remote monitoring: the operator watches camera views on a tablet over a mobile network, can start or stop work, and gets email alerts, without having to return to the field. MAFF extended its safety guideline to cover remote-monitored unmanned work in March 2024.
Achieved whenthe tractors are on sale, pass NARO's safety inspection for remote-monitored work, and a farm or researcher outside Kubota reports hours worked without anyone at the field and how often a person had to intervene.
Claimed by
Kubota, August 2026
Open
01
Pull weeds by demonstration
A low-cost robot arm that learns, from a few dozen remote-controlled demonstrations, to pull weeds by the root from beds or trays of vegetable seedlings without touching the crop, with every attempt filmed and counted.
The briefCredit · unclaimed02
Robots one person can watch
A simulator study that measures how many driverless tractors one person can safely supervise from a tablet, by planting obstacles, faults and stops in simulated fields and timing the person's responses as the fleet grows from one tractor to eight.
The briefCredit · unclaimed
Open 01 · the brief
Pull weeds by demonstration
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A low-cost robot arm that learns, from a few dozen remote-controlled demonstrations, to pull weeds by the root from beds or trays of vegetable seedlings without touching the crop, with every attempt filmed and counted.
Why it is still open
Even after robots passed, Südzucker's 2023 sugar beet trial still needed 10–13 hours of hand hoeing per hectare to clean the rows. Commercial weeding robots such as Aigen's Element remove weeds at the root with robot arms, but are trained mainly on simulated images, not by demonstration. A 2024 survey of imitation learning (robots learning from human demonstrations) in farm robotics reviews harvesting and food-handling studies but no weeding study, naming weeding only as a possible use, and a 2024 conference paper on a robot that pinches weeds with tweezers reported a low success rate, with weeds often torn. A November 2026 search of public LeRobot-format robot datasets on Hugging Face found none on weeding.
fwi.co.ukroboticsandautomationnews.compmc.ncbi.nlm.nih.govapi.openalex.org
Built with what exists
The SO-101 open-design robot arm; LeRobot to record demonstrations and train (see the Note "No download"); an open robot model such as NVIDIA GR00T 2 (see the Note "Five demonstrations per task") or Runway Praxis-1 (see the Note "Open robot model weights"); trays of lettuce or onion seedlings with weeds sown among them.
First step
Sow lettuce and weeds in ten trays and record 50 remote-controlled pulls with LeRobot.
Done looks like
100 filmed attempts on weeds growing within 2 cm of crop seedlings, with a published tally of weeds removed with the root, weeds snapped off, crop seedlings damaged and seconds per weed, and the data and trained model posted openly.
Open 02 · the brief
Robots one person can watch
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A simulator study that measures how many driverless tractors one person can safely supervise from a tablet, by planting obstacles, faults and stops in simulated fields and timing the person's responses as the fleet grows from one tractor to eight.
Why it is still open
Kubota's August 2026 announcement of remote-monitored tractors for 2027 does not say how many machines one person can watch, and NARO's 2024 review still placed remotely supervised machines in research and development. The closest study, by Adam Bogg and Stewart Birrell of Coventry University (2025), had 24 people supervise 3 to 9 simulated self-driving road vehicles and found monitoring best at 5 to 7. A 2022 University of Illinois simulation of field robots that call a remote operator for help compared ways of alerting the operator but did not vary the number of robots.
Built with what exists
The Gazebo robot simulator or NVIDIA Isaac Sim, driven through ROS 2; a tablet screen copying the camera view, stop and restart controls and alerts that Kubota describes; MAFF's levels of automation; the Coventry study's design as a template; the NASA-TLX workload questionnaire.
First step
Build one simulated field with one tractor and three scripted events, and time five volunteers.
Done looks like
A published study of at least 24 people, each supervising 1, 2, 4 and 8 simulated tractors, reporting missed events, response times and workload at each fleet size, with the simulator scenarios and anonymized data released.