Industries / Food, farming and forestry / Agriculture and farming
Harvest
Gathering the ripe crop from the field, orchard or greenhouse, by machine or by hand.
6
Milestones achieved, from 1831 to 2021.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1831
McCormick shows a reaper
Cyrus McCormick publicly showed a horse-drawn reaper, a machine that cuts standing grain, near Steele's Tavern, Virginia, in July 1831, and patented it on 21 June 1834. Sales were slow (29 in 1843 and 50 in 1844) until he moved to Chicago, where 700 were built and sold in a single year.
LessonA working demonstration took more than ten years to become a business.
Credit
Cyrus McCormick; Obed Hussey of Maryland patented a rival reaper
1836
A combine is patented
Near Climax, Michigan, Hiram Moore and John Hascall built a machine that both cut and threshed grain (knocked the grain loose from the stalk), the job a modern combine does, and patented it on 28 June 1836. The American Society of Agricultural Engineers calls it the first successful combined harvester-thresher.
Credit
Hiram Moore and John Hascall
1938
The combine drives itself
The Massey-Harris No. 20 had its own engine and drove itself through the crop instead of being pulled by horses or a tractor. The American Society of Agricultural and Biological Engineers calls it the first commercially successful self-propelled combine, first sold in 1938.
Credit
Thomas Carroll (chief engineer), Robert Ashton and Albert Luke, Massey-Harris Company
1944
Cotton picked by machine
Near Clarksdale, Mississippi, the Hopson Planting Company grew a 1944 cotton crop that machines planted, thinned and picked with no hand labour, after International Harvester engineers tested pickers there. Harvester began making commercial pickers in 1949, and machines picked virtually the whole US crop by the late 1960s.
LessonHistorian Donald Holley found that higher factory wages pulled workers out of the Cotton South more than machines pushed them out.
Credit
Hopson Planting Company; International Harvester; earlier work by John and Mack Rust
1959
A tomato bred for machines
At UC Davis, plant breeder Jack Hanna and engineer Coby Lorenzen paired a firm, blocky tomato with a machine that harvested and sorted it. After the US–Mexico Bracero guest-worker programme ended in 1964, the machine-picked share of California's processing tomatoes went from about 1% in 1963 to 95% in 1968.
LessonThe crop was redesigned to suit the machine.
Credit
Jack Hanna and Coby Lorenzen, UC Davis
2021
An apple robot shuts down
Abundant Robotics, a spin-out of SRI International, built a robot that found apples with cameras and pulled them off with a vacuum tube. It closed its business in June 2021 after it could not raise more money; its prototype filled a bin in 66 minutes against a 50-minute goal, co-founder Curt Salisbury said later.
LessonTree shape sets robot cost, and Salisbury said picking a 20-inch canopy costs twice as much as a 10-inch one.
Credit
Abundant Robotics (Dan Steere, Curt Salisbury, Michael Eriksen)
Watch for
Strawberry robot learns from demonstrations
A team adapted open robot models, which read camera images and output arm motions, to pick strawberries from tabletop plants in a real greenhouse, training on 227 remote-controlled demonstrations (3.71 hours). The best model reports 74.0% success, 32.6 seconds per pick and 4.1% damage over 50 trials, and no code or data release is listed.
Achieved whenanother group gets a similar success rate on its own greenhouse plants, or the data and code are released and re-run independently.
Claimed by
Ziyang Zhao, Shuheng Wang, Zhonghua Miao and Ya Xiong (HarvestFlex preprint), March 2026
USDA apple-picking robot
A USDA Agricultural Research Service team reports picking rates of 60–80% in Michigan orchards and field trials in commercial Washington orchards in the 2025 harvest, averaging 4–5 seconds per apple. It aims for 80% and 2–3 seconds, and has licensed the technology to Michigan State University, which formed a startup.
Achieved whena grower or independent evaluator publishes season-long results from commercial use near the 80% goal.
Claimed by
Renfu Lu, USDA ARS, January 2026
Raspberry robot fleets from 2027
Fieldwork Robotics raised £3 million for two years of trials on farms in Norfolk and Staffordshire, with Portugal and Australia to follow, and expects fleets of its robots on farms from 2027 if the trials succeed. It has published no pick rates.
Achieved whenrobots work commercially on farms and a grower or independent party reports pick rate and fruit damage.
Claimed by
Fieldwork Robotics, May 2026
Open
01
Count lost grain by photo
A phone tool that counts kernels on the ground inside a frame of known size behind the combine and converts the count to bushels per acre, plus a public test that compares it, sample by sample, with the hand method of collecting and weighing every kernel inside the frame.
The briefCredit · unclaimed02
Open strawberry-picking robot
A low-cost robot arm that picks ripe strawberries still attached to tabletop-grown plants, snaps or cuts the stem and places each berry in a punnet, with every demonstration and a per-berry results table published.
The briefCredit · unclaimed
Open 01 · the brief
Count lost grain by photo
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A phone tool that counts kernels on the ground inside a frame of known size behind the combine and converts the count to bushels per acre, plus a public test that compares it, sample by sample, with the hand method of collecting and weighing every kernel inside the frame.
Why it is still open
Mississippi State University Extension still teaches frame and drop-pan methods in October 2026 and says sifting chaff for grain "is time-consuming". CNH's Grain Cam Mobile app counts kernels from a photo, but its page gives no accuracy figures, and no independent check of it or open image set of grain on stubble was found.
Built with what exists
A phone camera; a PVC frame (Mississippi State uses 76 by 38 inches); a kitchen scale; Label Studio or CVAT for labelling; an open object-detection model such as Ultralytics YOLO; Grain Cam Mobile for comparison.
First step
Northern harvests are over, so scatter counted kernels (50, 100 and 200) on chaff inside a one-square-foot frame and photograph them in sun and shade to make a first labelled set.
Done looks like
200 frames across at least two crops, each photographed and then hand-collected and weighed, with the tool's estimate within 0.5 bushels per acre of the weighed value on at least 80% of frames, and all photos, labels and weights published alongside Grain Cam Mobile's scores on the same frames.
Open 02 · the brief
Open strawberry-picking robot
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 picks ripe strawberries still attached to tabletop-grown plants, snaps or cuts the stem and places each berry in a punnet, with every demonstration and a per-berry results table published.
Why it is still open
HarvestFlex calls itself the first study of these robot models on real greenhouse tabletop strawberries and lists no code or data. In a July 2026 Hugging Face forum thread, a researcher reviewing public robot datasets found "Almost everything is indoor tabletop", with almost no farm data.
Built with what exists
The SO-101 open-design arm; LeRobot to record and train (Notes: "No download"); an open robot model such as NVIDIA GR00T 2 (Notes: "Five demonstrations per task") or Runway Praxis-1 (Notes: "Open robot model weights"); everbearing strawberry plants under LED grow lights.
First step
Assemble an SO-101, pot two everbearing plants, and record and post 50 remote-controlled picks with LeRobot.
Done looks like
50 filmed attempts on ripe berries still on the plant, with a published tally of successful picks, bruised or dropped berries, unripe berries picked by mistake and seconds per pick, set beside HarvestFlex's 74%, 32.6 seconds and 4.1% damage.