Agricultural robotics

Agriculture was one of the first industries to adopt GNSS guidance at scale, and it is now one of the most active markets for field robots. Farms face shortages of skilled seasonal labour, pressure to cut herbicide and fertilizer use, concern about soil compaction from ever heavier machines, and narrow weather windows in which work must be done. This lesson surveys autonomous tractors and platforms, robots for weeding, spraying, harvesting and livestock, the coverage planning that underlies field operations, and the economics that decide whether robots replace, assist or simply change the shape of farm machinery. Product details are given as reported at the time of writing (October 2026) and change quickly.

From guidance to autonomy

Since the early 2000s, RTK auto-steer has let tractors follow straight or curved lines with a few centimetres of error, with the operator handling turns, implement control and safety. The next steps were automated headland turns, implement control linked to maps and machine-to-machine coordination, such as a combine harvester controlling the grain cart tractor that drives alongside it. Full autonomy removes the operator from the cab, which requires perception to detect people, animals and obstacles, a safety concept for the field, remote supervision and procedures for exceptions.

Several approaches coexist. Retrofit kits add cameras, computing and drive-by-wire interfaces to existing large tractors. Deere demonstrated a driverless 8R tractor for tillage in 2022 and at CES 2025 announced a second-generation perception kit with 16 stereo cameras for 8R and 9R tractors and tillage, plus kits for its 5ML orchard tractors and orchard sprayers using cameras and lidar, with wider availability planned for 2026. Purpose-built platforms take a different path. DOT, developed by DOT Technology in Regina, Saskatchewan, was a driverless U-shaped power unit that implements such as seeders and sprayers slide into; Raven Industries took control of the company in 2019 and rebranded it OMNiPower in 2021, and Raven was itself acquired by CNH. Raven's OMNiDrive kit, originally AutoCart, lets a combine operator command a driverless grain cart tractor. Other companies build smaller electric or diesel robots designed to work in fleets.

Safety of highly automated agricultural machines is addressed by ISO 18497, which sets requirements for the design of such machines and their obstacle protection systems; designers should use the current edition. In practice, autonomous farm machines operate in fields with controlled access, stop for detected obstacles, and are monitored remotely by an operator who can see camera feeds and intervene. Regulation of public-road travel, drones and mobile robots was covered in ROB-450; on the farm, occupational health and safety law and the manufacturer's safety case are the main controls.

Coverage planning for field operations

Most field operations, from tillage to spraying, require covering an area with an implement of fixed working width. The standard pattern is boustrophedon coverage, named after the back-and-forth way an ox ploughs: parallel passes across the field, connected by turns in headlands at each end. ROB-350 treated motion planning in general; here the questions are practical. How long will the job take? How much area is overlapped or missed? How should the passes be oriented to reduce turns? The answers depend on implement width \( w \), overlap fraction, speed, turning time and field shape.

For a rectangular field of length \( L \) along the passes and width \( W \) across them, the number of passes is about \( N = \lceil W / (w(1 - o)) \rceil \) for overlap fraction \( o \). The working time is the pass length divided by speed times the number of passes, and the non-productive time is the number of turns times the time per turn, plus time for refilling, unloading or recharging. Field efficiency is the ratio of productive time to total time. Orienting passes along the longest side of the field minimizes the number of turns; irregular fields are split into cells that are each covered in their best orientation.

# Coverage time and field efficiency for a boustrophedon pattern
import math
def coverage(L, W, width, overlap, speed_kmh, turn_s, refill_every_ha=None, refill_min=0):
    passes = math.ceil(W / (width * (1 - overlap)))
    work_h = passes * L / (speed_kmh * 1000)
    turn_h = (passes - 1) * turn_s / 3600
    area_ha = L * W / 10000
    refill_h = 0 if refill_every_ha is None else math.floor(area_ha / refill_every_ha) * refill_min / 60
    total = work_h + turn_h + refill_h
    return passes, total, work_h / total, area_ha / total
field = dict(L=800, W=400)   # m
cases = {"large sprayer 36 m":   dict(width=36, overlap=0.02, speed_kmh=16, turn_s=40, refill_every_ha=40, refill_min=20),
         "small robot 3 m":      dict(width=3, overlap=0.05, speed_kmh=6, turn_s=25, refill_every_ha=4, refill_min=10),
         "passes along short side": dict(width=3, overlap=0.05, speed_kmh=6, turn_s=25, refill_every_ha=4, refill_min=10)}
for name, c in cases.items():
    L, W = (field["L"], field["W"]) if "short" not in name else (field["W"], field["L"])
    n, t, eff, rate = coverage(L, W, **c)
    print(f"{name:24s} passes {n:4d}  total {t:5.1f} h  field efficiency {eff:4.0%}  {rate:5.2f} ha/h")

large sprayer 36 m       passes   12  total   0.7 h  field efficiency  83%  44.31 ha/h
small robot 3 m          passes  141  total  21.1 h  field efficiency  89%   1.52 ha/h
passes along short side  passes  281  total  22.0 h  field efficiency  85%   1.45 ha/h

The large sprayer covers the 32 ha field in under an hour of field time, while a single small robot needs about 21 hours. Running passes along the short side doubles the number of turns and lowers efficiency further. This is why small-robot concepts depend on fleets working in parallel and on operating many more hours per day, including at night, and why they suit tasks such as weeding, where precision and frequent passes matter more than raw work rate.

Weeding and precision spraying

Weed control is the task where robots have made the clearest commercial progress. Herbicide resistance, consumer demand for reduced chemical use, the cost of hand weeding in organic and vegetable crops, and the arrival of reliable deep-learning plant detection have all pushed in the same direction. Three main approaches exist. Precision spraying uses cameras on a boom to detect weeds and fire individual nozzles only where needed; Deere's See and Spray, which grew out of its 2017 acquisition of Blue River Technology, is the best-known example, and the company reports substantial average herbicide savings compared with broadcast spraying. Mechanical weeding robots use hoes, tines or rotating tools steered by vision to cut weeds between and within crop rows. Thermal and laser weeding robots kill weeds with heat or targeted laser energy; laser weeders have been sold for vegetable crops in North America since the early 2020s.

The core perception problem is detecting and classifying small plants, often at the two- to four-leaf stage, under changing light, with soil, residue and overlapping leaves, at the speed of a moving machine. The robot must also act within milliseconds: at 10 km/h a machine moves 2.8 metres per second, so a nozzle 0.5 m behind the camera has under 0.2 s between detection and spraying. Errors have asymmetric costs: missing a weed leaves competition and seed for future years; killing a crop plant loses yield directly.

# Herbicide savings and the cost of detection errors (illustrative values)
import numpy as np
rng = np.random.default_rng(4)
area_ha, weed_cover = 100, 0.12           # weeds occupy 12 % of the area on average
broadcast_l_ha, cost_l = 1.5, 12.0         # L/ha and dollars per L
recall, false_pos = 0.95, 0.05             # weeds sprayed, and clean area sprayed by mistake
fields = rng.normal(weed_cover, 0.04, 10).clip(0.02, 0.3)
for wc in fields[:5]:
    sprayed_fraction = recall * wc + false_pos * (1 - wc)
    saving = 1 - sprayed_fraction
    cost_saved = area_ha * broadcast_l_ha * cost_l * saving
    print(f"weed cover {wc:5.1%}  sprayed {sprayed_fraction:5.1%}  herbicide saving {saving:5.1%}  ${cost_saved:,.0f} per 100 ha pass")

weed cover  9.4%  sprayed 13.5%  herbicide saving 86.5%  $1,558 per 100 ha pass
weed cover 11.3%  sprayed 15.2%  herbicide saving 84.8%  $1,527 per 100 ha pass
weed cover 18.7%  sprayed 21.8%  herbicide saving 78.2%  $1,408 per 100 ha pass
weed cover 14.6%  sprayed 18.2%  herbicide saving 81.8%  $1,473 per 100 ha pass
weed cover  5.4%  sprayed  9.9%  herbicide saving 90.1%  $1,622 per 100 ha pass

Savings are largest when weeds are sparse, which is common in fields that are already well managed, and shrink as weed cover rises. Missed weeds (recall below 100 per cent) are the agronomic risk; false positives are mostly a cost. Growers often combine a pre-emergence broadcast application with targeted post-emergence spraying, so real savings depend on the whole programme.

Harvesting, orchards and greenhouses

Robotic harvesting of fresh fruit and vegetables is much harder than weeding. The robot must find fruit hidden among leaves, judge ripeness, plan a reach through branches, detach the fruit without bruising it or damaging the plant, and do so fast enough to compete with skilled pickers. Research and commercial systems for strawberries, apples, tomatoes, sweet peppers and mushrooms have used vacuum tubes, soft grippers and cutting end effectors, and some apple systems use flying or arm-mounted vacuum pickers. Many have struggled with throughput and cost. Orchard and greenhouse growers are adapting crops to robots: trellised two-dimensional fruiting walls, planar canopies and table-top strawberries make fruit far easier to see and reach. Greenhouses also offer controlled lighting, flat floors and rails, so mobile platforms that monitor plants, deleaf tomatoes or move benches are more mature there than in open fields.

Livestock and dairy robots

The most widespread farm robots are not in fields but in barns. Voluntary or automatic milking systems let cows choose when to be milked; the robot identifies the cow by RFID, cleans the teats, attaches the cups using cameras or laser sensors, monitors milk quality and records data for each animal. Robotic milking spread from the Netherlands in the 1990s and is now common on dairy farms in Canada and the United States, particularly family farms where labour is scarce. Barn robots also push feed back to the feed bunk, clean manure alleys and feed calves. These systems show a pattern seen across field robotics: success came from redesigning the whole system, including barn layout and cow traffic, rather than replacing a human motion by a robot motion.

Economics, data and policy

Farm robots must pay for themselves within narrow margins and short seasons. The value comes from labour savings, timeliness (getting the work done in the right window), input savings such as herbicide, yield gains from less compaction and more precise treatment, and data. Costs include purchase or subscription, connectivity and correction services, supervision, maintenance and the risk of downtime at critical times. Rural connectivity is a real constraint: RTK corrections, remote supervision and data uploads need cellular or satellite links, and low-earth-orbit satellite internet has improved coverage on many farms since 2021.

In Canada, Agriculture and Agri-Food Canada funds research and adoption programmes for agricultural technology, and provincial agencies and research centres run trials of autonomous equipment; check current programme names and terms, which change between federal funding frameworks. Labour shortages, particularly of seasonal workers who often come through temporary foreign worker streams, are a major driver of interest in horticultural robots.

Data ownership and the right to repair are live policy issues. Autonomous machines collect detailed field, yield and operational data, and farmers want control over who uses it. Repair of modern equipment often requires manufacturer software, and farm groups in both countries have pressed for repair rights; several US states have passed agricultural right-to-repair laws, and Canada amended its Copyright Act in 2024 to allow circumvention of technological protection measures for diagnosis, maintenance and repair. Students designing farm robots should expect customers to ask about these issues.