Agricultural Robotics — Weeding, Harvesting, Spraying
Scope. Ground and aerial robots that operate in fields, orchards, vineyards, greenhouses, and livestock barns. The four anchor tasks are weeding, harvesting, spraying, and planting, with secondary tasks of milking, livestock handling, and field scouting. Mechanical mobility is borrowed from
[[Robotics/mobile-base-wheeled]](Ackermann tractors, skid-steer field robots) and[[Robotics/multirotor-design]](sprayer drones); per-plant identification borrows from[[Robotics/computer-vision-robotics]]. This note focuses on what is agriculture-specific: terramechanics, row-following, per-plant intervention economics, ISOBUS, NDVI, the regulatory + environmental envelope, and the named machines a 2026 farm actually buys.
1. At a glance
Agriculture is the second-largest robotics market by deployed unit count after warehouse logistics, and the fastest-growing as of 2026. The progression has been mechanisation (1900–1950) — replacing draft animals with tractors — then automation (1950–2010) — hydraulics, GPS-guided autosteer, ISOBUS implement control — and now autonomy (2018–present) — true unmanned operation, per-plant decisioning, and aerial spray fleets. Three forces drive the 2026 wave:
- Labour cost and supply. US H-2A guest-worker minimum wage was USD 17.20/hr in 2024 (USDA AMS), up 6%/yr since 2020. California specialty-crop growers report 50%+ of seasonal positions unfilled in peak years (CFB 2023 survey). EU situation is similar (FNSEA, France 2024).
- Sustainability and regulation. The EU Farm-to-Fork strategy targets 50% pesticide reduction by 2030; the EPA tightened glyphosate use 2024. Carbon Robotics LaserWeeder field trials report >95% herbicide elimination in row crops by replacing chemical with thermal/laser weeding.
- Precision economics. Per-plant intervention (spray only weeds, pick only ripe fruit, fertilise only N-deficient zones) reduces input cost 20–90% depending on crop. John Deere See & Spray claims 60–90% herbicide reduction at parity yield.
Real platforms in production or commercial pilot as of 2026: John Deere See & Spray Premium / Ultimate (Blue River acquisition 2017) and autonomous 8R tractor (2022 launch); CNH Industrial Raven autosteer + autonomous CNH Industrial New Holland T6; AGCO Symphony Robotics; Naïo Technologies Dino (field-scale veg), Ted (vineyard), Oz (small-farm); Carbon Robotics LaserWeeder (CO₂ laser, US row crops); FFRobotics and Octinion Rubion for strawberries; GUSS Automation autonomous orchard sprayer (CA); Saga Robotics Thorvald (UV-C strawberry powdery-mildew); Tortuga AgTech (table-top strawberry pick); Iron Ox greenhouse cells; DJI Agras T50 (50 L spray drone, replacing the T40); Lely Astronaut and DeLaval VMS milking; Solinftec Solix scouting; Ecorobotix ARA ultra-precision spot sprayer.
Where it sits in the design stack. A field robot is a wheeled mobile base ([[Robotics/mobile-base-wheeled]]) carrying a perception payload ([[Robotics/computer-vision-robotics]], [[Robotics/sensors-perception]]), a manipulation tool (boom, end-effector, laser turret), and a domain-specific decision layer (weed-vs-crop classifier, ripeness regression, NDVI-zoned spray map). The agricultural-specific layers above are: ISOBUS for tractor-implement comms, a farm-management system (FMS, e.g., John Deere Operations Center) for fleet + agronomic data, and a regulatory wrapper (EPA, FAA Part 137 for aerial spray, USDA, FDA for raw produce).
First ask before any ag-robot project:
- Crop and task. Row crop (corn, soy, cotton) vs. specialty (strawberry, apple, lettuce) vs. tree crop (orchard, vineyard) vs. greenhouse vs. livestock. Each cluster has a different robot.
- Indoor or outdoor? Greenhouse is structured (rails, climate-controlled); open field is unstructured (terrain, weather, GNSS-denied under canopy).
- Whose tractor base? Retrofit an existing Deere/CNH/AGCO platform via ISOBUS, or build a clean-sheet small robot (Naïo, Burro)?
- Per-plant or per-zone? Per-plant requires fast perception + actuation; per-zone (variable-rate prescription) is solved by autosteer + map.
- Connectivity? Rural 4G/5G coverage drives whether the FMS round-trip is online or store-and-forward.
- Regulatory class. US drone spraying is FAA Part 137; EU pesticide robots are under SUD revisions; autonomous tractors on public road require state-level approval.
- Throughput economics. Compute the seasonal hour-budget — most ag tasks have a 2–6 week window. Robot must finish the field within the window.
2. Why it matters
The robotics community often treats agriculture as a niche, but the addressable market is the second-largest after warehouse logistics. As of 2026:
- Global agricultural-robot market: USD 13–17 B (multiple firms, 2025 estimates), CAGR 22–25%
- US farm labour: ~2.4 M workers, of which ~25% are H-2A foreign seasonal; median wage USD 17.20/hr (2024)
- LaserWeeder eliminates ~95% of herbicide in trials (Carbon Robotics 2024 field reports)
- See & Spray Premium reduces herbicide use 60–90% in soy, corn, cotton (John Deere 2023)
- Lely + DeLaval combined milking-robot install base > 30,000 farms globally (Lely investor day 2024)
- DJI Agras T-series sales > 250,000 units cumulative (DJI 2024)
- Field-corn and cotton acreage with autosteer adoption > 70% in the US (USDA NASS 2022)
For an individual farm, the unit economics break down approximately as:
| Task | Manual cost (USD/ha) | Robotic cost (USD/ha) | Notes |
|---|---|---|---|
| Soy weeding (broadcast spray) | 60–90 | 25–40 | See & Spray spot-spray savings |
| Lettuce weeding (hand-hoe) | 300–500 | 80–150 | LaserWeeder, Naïo Dino |
| Strawberry pick (hand) | 0.30–0.60/lb | 0.20–0.40/lb | FFRobotics, Octinion, Tortuga |
| Orchard spray | 50–80 | 30–50 | GUSS autonomous tractor |
| Aerial spray (rotary-wing) | 30–50 | 10–25 | DJI Agras vs Ag-Cat |
| Dairy milking | 4–6 USD/cow/day labour | 2–3 USD/cow/day amortised | Lely / DeLaval |
The picker rate gap is dramatic in strawberries: a skilled human worker picks ~600 berries/hr; a 7-arm FFRobotics gantry sustains ~3,360 berries/hr (≈5×). The savings appear when the labour is unavailable, not (yet) when it is cheap.
3. First principles
Terramechanics — soil-tire interaction
Agricultural mobility happens on deformable soil. The standard model is Bekker (1956, extended by Wong 1989):
Pressure-sinkage:
p = (k_c / b + k_φ) · z
where p is contact pressure (kPa), z is sinkage (m), b is the smaller contact-patch dimension, and (k_c, k_φ, n) are soil parameters from a bevameter test. Tractive force is bounded by Mohr-Coulomb shear:
τ = c + σ · tan(φ)
with c soil cohesion (kPa), φ internal friction angle (°), σ normal stress. For autonomous tractors and field robots this matters because:
- A 6,000 kg autonomous 8R tractor with 4×0.5 m² contact patches generates ~30 kPa ground pressure — comparable to a manned tractor. Soil compaction is identical.
- A 400 kg Naïo Dino on 4× narrow tires generates ~80 kPa contact pressure but the total integrated compaction is far less because the machine is light.
- Slip ratio s = 1 − v_actual / v_wheel; field tractors run 10–15% slip in normal till, 25%+ when struggling. Higher slip = wasted fuel + soil disturbance.
Same model governs planetary rovers (see [[Robotics/legged-robotics]] if the library has it).
GNSS-RTK for row-following
The dominant positioning stack on a 2026 ag robot is RTK-GNSS:
- Base station + rover, L1+L2+L5 multi-frequency receiver (Trimble BX982, Septentrio Mosaic, u-blox ZED-F9P)
- RTK fix accuracy: 1–3 cm horizontal, 2–5 cm vertical (open sky, < 30 km baseline)
- Refresh: 10–20 Hz
- Failure modes: tree-canopy multipath (orchards), urban-corridor blockage (greenhouse perimeter), ionospheric scintillation (low latitude, solar max)
For row-following at < 1 km/h slow weeding work, ±5 cm is sufficient. For 10 km/h broad-acre work, ±2 cm is needed to stay in-row.
When GNSS is degraded (under orchard canopy or in low-light greenhouse), the robot falls back to visual or LiDAR-based row detection — Hough-transform row-line fit on near-IR or RGB, or learned crop-row segmentation (CNN). Bonirob, Naïo Dino, and Carbon Robotics all maintain a vision fallback.
Inter-row navigation
Three layered methods, used together:
- GNSS-RTK absolute — global path waypoints in field-frame
- Vision-based row centerline — RGB or NIR camera looks down at the rows ahead, finds two parallel green stripes via Hough transform or learned segmentation, generates a centerline cross-track error
- LiDAR-based crop-volume — 2D or 3D LiDAR finds the canopy as obstacles, fits the lane
A practical heuristic: GNSS-RTK if the satellites are healthy, vision row-detection as primary servo, LiDAR for emergency-stop on canopy gap / obstacle.
Per-plant identification
For weed-vs-crop or fruit-ripe vs unripe:
- Closed-set deep learning: YOLOv8/v10 detector trained on labelled in-field RGB at 1–3 ms inference per 720p image on Jetson Orin. Classes: crop sp., 6–12 weed species, soil. Mask R-CNN or SAM-2 for instance segmentation when overlap matters.
- Active illumination + multispectral: NDVI = (NIR − RED) / (NIR + RED), threshold > 0.4 = vegetation, < 0.4 = soil. Cheap, weather-robust, but cannot distinguish weed from crop.
- Hyperspectral: Cubert / Headwall imagers spread 200+ bands; sufficient signal to discriminate corn from pigweed via chlorophyll absorption — used in research, rarely in production due to cost.
- 3D pose + ripeness: RGB-D (Realsense D435i, ZED 2i) + colour-based ripeness regression (red-channel histogram). Octinion Rubion claims 95% ripeness classification accuracy.
A common pipeline: detector → tracker → per-instance ripeness/weed regression → spray valve actuation. End-to-end latency budget < 50 ms at 10 km/h means 14 cm of vehicle motion per cycle, so the actuator must be predictively triggered on the projected ground position, not the current one.
Mechanical vs laser vs chemical weeding
Three competing physical mechanisms, each with a champion:
- Mechanical (Naïo Oz / Dino, FarmDroid FD20) — rotating hoe or finger-weeder between rows; in-row weeds removed by camera-guided rotating disk. ~3 km/h.
- Laser (Carbon Robotics LaserWeeder) — 30× CO₂ lasers, 150 W each, kill weed apex meristem via thermal denaturation. ~3 km/h, 200,000 weeds/hr. No chemical, no soil disturbance.
- Chemical spot-spray (John Deere See & Spray, Ecorobotix ARA) — narrow nozzles fire 1–5 mL of herbicide at each detected weed. 10–25 km/h achievable. 60–90% herbicide reduction at field scale.
The mechanical and laser options are pesticide-free; the spot-spray option requires herbicide but enables broadcast-equivalent throughput.
Picking and grasping
See [[Robotics/end-effectors]] for general gripper kinematics. Ag-specific grasping constraints:
- Bruising threshold — strawberries bruise at ~0.5 N contact force; tomato 1–2 N; apple 5–10 N
- Stem detachment — twist-pull (apple), abscission (ripe tomato), cut (strawberry pedicel)
- Occlusion — leaves hide fruit; ~30% of ripe strawberries are not visible from the top
- Soft-grip strategies: silicone suction cup (Octinion), fingered soft pneumatic gripper (Soft Robotics Inc. mGrip), under-actuated rigid (FFRobotics)
The “vine touch” sensor (capacitive or force) tells the gripper it has contacted the fruit and triggers detachment.
Crop modelling and decision support
Closing the loop on a robot requires an agronomic model of the field:
- NDVI / NDRE zoning — field map of N status drives variable-rate fertiliser prescription
- APSIM / DSSAT — process-based crop growth models, predict yield + harvest timing from weather + soil + cultivar
- Yield prediction — pre-harvest scout-drone counts fruit, regressor predicts total tonnage ±10–15%
- Disease detection — Saga Thorvald uses UV-C lamps at night to kill strawberry powdery mildew (Podosphaera aphanis); 80%+ reduction in fungicide
4. Worked examples
A — Robot tractor field coverage
Scenario. 100 ha cornfield, autonomous Deere 8R with 6 m boom sprayer.
Implement effective swath: 6 m. Forward speed: 8 km/h = 2.22 m/s. Coverage rate:
A_dot = 6 m × 2.22 m/s = 13.3 m²/s = 4.8 ha/hr
Including 15% overlap and 10% turnaround at headlands:
A_effective = 4.8 × 0.85 × 0.90 = 3.67 ha/hr
Time per hectare: 1 / 3.67 ≈ 0.27 hr ≈ 16 min/ha → 100 ha × 0.27 hr = 27 hr.
If the autonomous tractor runs 18 hr/day (refuel + RTK base maintenance), the field is covered in 1.5 days. A manned tractor at 10 hr/day completes the same field in 2.7 days — autonomy gains ~1.8× by extending shift, not by going faster.
Diesel. Deere 8R 410 burns ~25 L/hr at 80% load → 27 hr × 25 L/hr = 675 L = 670 USD at USD 1.00/L. Per-hectare fuel: 6.75 L = USD 6.75/ha — small compared to herbicide (USD 40–100/ha).
B — Drone spray budget
Scenario. DJI Agras T50, 50 L tank, 12 m effective swath, 7 m/s ferry / 5 m/s spray speed; nominal spray rate 30 L/ha.
Single-tank coverage:
A_tank = 50 L / 30 L/ha = 1.67 ha per tank
Spray time per tank:
t_spray = 1.67 ha / (12 m × 5 m/s) = 1.67 × 10000 / 60 = 278 s ≈ 4.6 min spraying
Add 2 min ferry + 2 min swap-and-recharge battery + refill tank → ~9 min total cycle. So 1 ha covered in:
t_per_ha = 9 min / 1.67 ha ≈ 5.4 min/ha
In an 8-hr field day:
A_day = 480 min × (1 ha / 5.4 min) ≈ 89 ha/day
With a 50% duty cycle (battery swap delays, wind hold), realistic = 45 ha/day, or 75 ha/day with two parallel drones and an attentive ground crew. A traditional fixed-wing ag plane (Air Tractor 502, 1900 L hopper, 50 m swath) covers ~300 ha/hr but costs USD 1.5 M; the T50 costs USD 18,000 (2026 price).
C — Strawberry picker rate
Scenario. FFRobotics 7-arm gantry over open-field strawberry rows.
Each arm sustains ~8 berries/min average (including pose, vision recognition, pick, place, retract). 7 arms in parallel:
R_robot = 7 × 8 = 56 berries/min = 3,360 berries/hr
Human picker rate (skilled, peak season, optimal row): ~600 berries/hr
Ratio: 5.6× per machine vs per worker. Assuming 12 hr operation:
Daily robot throughput = 3,360 × 12 = 40,320 berries/day Daily human throughput = 600 × 8 = 4,800 berries/day per worker
Robot replaces ~8.4 workers per shift. With berry mass 25 g: 40,320 × 0.025 = 1,000 kg/day ≈ 1 t. At USD 5/kg field price = USD 5,000/day revenue per gantry. Gantry capital cost USD 250,000 → 50-day payback at full utilisation, ~120-day payback realistically.
5. Vehicle and system categories
Autonomous tractors (broad-acre)
| Platform | Vendor | HP class | Year | Comms | Notes |
|---|---|---|---|---|---|
| 8R 410 Autonomous | John Deere | 410 hp | 2022 | ISOBUS + JD Link | First series-production fully autonomous large tractor |
| New Holland T6 Methane Power Autonomous (concept→pilot) | CNH Industrial | 175 hp | 2023 pilot | ISOBUS + Raven | Methane-fuelled |
| Symphony Robotics OmniPower | AGCO / Fendt | 130–200 hp | 2024 | ISOBUS | Driverless transplant + cultivation |
| Solectrac eFarmer | Ideanomics | 70 hp electric | 2023 | CAN | Small electric, retrofit autonomy add-on |
| Monarch MK-V | Monarch Tractor | 70 hp electric | 2022 | proprietary | Driver-optional, electric, fleet learning |
Field robots (small-format, sub-1 t)
| Platform | Vendor | Country | Task | Drive |
|---|---|---|---|---|
| Dino | Naïo Technologies | FR | Vegetable weeding | 4× DC, RTK |
| Ted | Naïo Technologies | FR | Vineyard inter-row | 4× DC |
| Oz | Naïo Technologies | FR | Small-farm weeding | 4× DC |
| Robotti | Agrointelli | DK | Tool-carrier, 150 hp diesel | Ackermann |
| Thorvald | Saga Robotics | NO/UK | UV-C, scouting, spray | 4× swerve |
| Burro | AugLogica (Burro) | US | Carry-companion in orchards | diff-drive |
| FarmDroid FD20 | FarmDroid | DK | Solar-powered seed + weed | 4× DC, RTK |
| BoniRob (legacy) | Bosch Deepfield Robotics | DE | Research weeding platform | 4× steer |
Greenhouse and indoor
Iron Ox (US, vertical-farm cells with mobile manipulator + rail), Bowery (NY, indoor vertical), Plenty (CA, automated leafy greens), MetoMotion GRoW (IL, tomato picking on rail), Priva / Ridder (NL, climate + crop control software backbones).
Harvesters
| Crop | Vendor | Mechanism | Status 2026 |
|---|---|---|---|
| Strawberry | FFRobotics (IL) | 7-arm gantry, cut-and-place | Commercial pilots |
| Strawberry | Octinion / Rubion (BE) | Soft-suction, single arm | Commercial in EU |
| Strawberry | Agrobot E-Series (ES) | 24-arm row carrier | Commercial Spain |
| Strawberry (tabletop) | Tortuga AgTech (US) | Dual-arm cobot | Commercial CA, FL |
| Apple | Abundant Robotics (US) | Vacuum tube pick | Defunct 2021 |
| Apple | FFRobotics | Multi-arm gantry | Pilot |
| Lettuce | Earth Rover AVL | Selective laser thinning + harvest | Pilot UK |
| Mushroom | ABB + Aitia | Soft-pneumatic | Pilot HU |
| Asparagus | AGS by AvL / AgroBot | Vision + cut | Commercial EU |
| Sweet pepper | Sweeper EU FP7 | Single arm + Cartesian | Research |
Sprayers (ground + aerial)
| Platform | Vendor | Class | Coverage | Notes |
|---|---|---|---|---|
| See & Spray Premium / Ultimate | John Deere (Blue River) | Tractor-mounted boom | 36–48 m boom | Per-plant herbicide |
| ARA | Ecorobotix (CH) | Tool-carrier sprayer | 6 m boom | 95% chemical reduction claim |
| GUSS | GUSS Automation (US) | Autonomous orchard sprayer | 5 m boom | Almonds + citrus |
| Solix | Solinftec (BR/US) | Scouting + spot-spray | 3 m | Solar + electric |
| T50 / T40 | DJI Agras | Multirotor 50 L | 12 m swath | FAA Part 137 in US |
| V40 | DJI Agras | Coaxial 8-rotor 40 L | 11 m swath | High-density |
| MG-1P (legacy) | DJI Agras | 10 L hex | 6 m | Early-gen |
Weeders
Carbon Robotics LaserWeeder (US — CO₂ laser, no chemical), Naïo Dino (FR — mechanical + camera), Ecorobotix ARA (CH — spot-spray), FarmDroid FD20 (DK — mechanical + solar), FarmWise Titan (US — mechanical, 12-row).
Milking and dairy
| Platform | Vendor | Country | Approach |
|---|---|---|---|
| Astronaut A5 | Lely | NL | Free-traffic robotic stall |
| VMS V300 | DeLaval | SE | Free-traffic with body-condition scoring |
| MR-S1 | BouMatic Robotics | US/NL | Stall + arm |
| Merlin² | Fullwood | UK | Stall + arm |
Combined install base > 30,000 farms; ~5% of EU/NL dairy cows milked robotically.
Livestock and other
Halter Cow Collar (NZ — virtual fencing); SwagBot (AU — mustering); BoviSync (US — health monitoring); Cattle-Eye (US/UK — body-condition video).
6. Sensing and perception
| Sensor | Purpose | Typical model | Cost (USD) |
|---|---|---|---|
| GNSS-RTK | Cm positioning | u-blox ZED-F9P, Trimble BX982 | 500–4,000 |
| RGB-D | Plant 3D, ripeness | Realsense D435i, ZED 2i, Femto Bolt | 300–1,500 |
| Multispectral | NDVI, NDRE, crop status | MicaSense RedEdge-P, Parrot Sequoia+ | 4,500–8,000 |
| Hyperspectral | Disease, weed species | Cubert FireflEYE, Headwall Nano | 30,000–80,000 |
| Thermal | Plant stress, animal health | FLIR Vue Pro, Workswell WIRIS | 4,000–15,000 |
| LiDAR | 3D canopy, terrain | Velodyne Puck, Ouster OS1-32, Livox Mid-360 | 800–8,000 |
| IMU | Roll-pitch, deadreckoning | VectorNav VN-100, Xsens MTi-680G | 1,500–5,000 |
| Soil EC / pH / NPK | Variable-rate input | Veris MSP3, Stevens HydraProbe | 2,000–10,000 |
| Weather station | Microclimate | Davis Vantage Pro2, METER ATMOS-41 | 1,000–3,500 |
NDVI = (NIR − RED) / (NIR + RED). Values: 0.7–0.9 = healthy crop; 0.4–0.7 = moderate; 0.2–0.4 = sparse / stressed; < 0.2 = soil. NDRE substitutes red-edge band for higher sensitivity at high biomass.
Most ground robots fuse RTK-GNSS + IMU + wheel odom in a tightly-coupled EKF (e.g., robot_localization in ROS 2), with vision/LiDAR for in-row and obstacle detection. Drone sprayers add baro + airspeed + downward-looking ToF for terrain-following.
7. Software and standards
ISO 11783 (ISOBUS) is the canonical tractor-implement communication standard, built on SAE J1939 CAN. Parts 1–14 cover physical layer (250 kbps CAN at 9-pin), network management, virtual terminal (operator UI on implement, displayed via tractor screen), task controller (TC-BAS for variable-rate prescription, TC-GEO for geo-tagged, TC-SC for section-control). 2026: ISO 11783-10 (TC-GEO) is universally implemented; high-speed ISOBUS (ISO 11783-7 over 100BASE-T1 Automotive Ethernet) is rolling out for vision-heavy sprayers like See & Spray.
AEF (Agricultural Industry Electronics Foundation) governs compliance + certification (“AEF Functionality” labels: ISB, AUX-N, TC-BAS, TC-GEO, TC-SC, TIM). Multi-vendor interoperability depends on AEF database lookup.
Farm-management systems (FMS):
| FMS | Vendor | Strength |
|---|---|---|
| Operations Center | John Deere | Tightest tractor + implement integration |
| FieldOps | CNH Industrial | Multi-brand, Raven legacy |
| FieldView | Climate (Bayer) | Agronomic decision support |
| AgriWebb | AgriWebb (AU) | Livestock + paddock |
| Agrian | Agrian | Crop protection compliance |
| Granular | Corteva | Profit + recordkeeping |
APIs and integration: most FMS expose REST + occasional GraphQL. Common identifiers: field GUID, ADAPT framework (AgGateway) for the lossless tractor-implement task interchange. CropX, Trimble Ag, AgX provide middleware.
Drone flight planning: DJI Terra, DJI Pilot 2, Pix4Dfields, Sentera FieldAgent, DroneDeploy, UgCS for ag.
Open-source robotics:
- ROS 2 Jazzy / Lyrical (Lyrical Luth LTS, 2026) — most research platforms (Thorvald, FarmDroid, BoniRob legacy)
- Nav2 stack for autonomous in-field navigation
- PCL + Open3D for 3D point clouds (LiDAR canopy)
robot_localizationfor RTK-GNSS + IMU fusion- Gazebo / Ignition for sim
- Agri-specific ROS packages:
agrobot_ros,naio_ros(community), Bonn UoB perception datasets
8. Edge cases and failure modes
- Weather. Rain interrupts spray ops (drift) and vision (water on lens). Wind > 5 m/s aborts drone spraying. Heat (> 40 °C) derates compute and batteries. The robot must have a “park-and-wait” mode plus weather-tight covers (IP65+).
- Uneven terrain. Furrow crossings can lift one wheel; pitch > 15° risks rollover for narrow robots. Naïo Dino has roll-over detection that triggers e-stop and lowers stabilising outriggers.
- Crop occlusion. Leaves hide ~30% of ripe strawberries from top-down view. Solutions: under-canopy camera, dual-arm gentle leaf-displacement, multiple visits per row.
- GNSS denial. Tree canopy in orchards drops RTK fix; cliffs and silos cause multipath. Fallback: VIO + LiDAR-based SLAM (
[[Robotics/slam]]), or pre-mapped UWB beacons in greenhouses. - Battery and diesel range. Small electric robots (FarmDroid, Naïo Oz) limited to 4–10 hr per charge; field-day operation requires hot-swap or in-field DC fast-charge. Diesel platforms (Robotti, Deere 8R) limited by fuel tank but can refuel in 5 min.
- Regulatory. US drone agricultural spraying is regulated under 14 CFR Part 137 (“Operation and Certification of Agricultural Aircraft Operators”) — requires commercial pilot certificate, agricultural-aircraft operator certificate, and EPA-label compliance for each chemical. EASA equivalent is national-competent-authority approval under EU 2019/947 (Specific category, PDRA).
- Theft and vandalism. Remote fields, no operator; the robot is a USD 50,000–500,000 asset. GPS tracking, geofence alarms, cellular kill-switch are now standard.
- Connectivity. Rural 4G coverage gaps force store-and-forward; the robot must operate fully autonomously when offline and sync only when in range. CNH and Deere have private LoRaWAN / mesh for in-field telemetry.
- Vibration and dust. ISO 16750-3 vibration class V2 minimum; IP66+ for dust; PM2.5 from soil pulverises camera lenses → washer nozzles + hydrophobic coatings.
- Pesticide drift and cross-contamination. Spray drone wash > 8 km/h crosswind drifts off-target; spot-spray on weeds 5 cm from crop risks crop phytotoxicity. EPA-label buffer zones must be respected by the path planner.
- Animal interaction. Cattle approach moving robots; livestock-monitoring robots must detect proximate animals (PointPillars on LiDAR + thermal). Halter cow collars and SwagBot include explicit animal-aware controllers.
- Soil compaction. A 12 t fully-loaded autonomous combine causes lasting subsoil compaction; controlled-traffic farming (CTF) keeps heavy wheels on fixed lanes. ISOBUS-shared lane database lets all vendors’ equipment follow the same tramlines.
- Mixed-fleet planning. Multiple vendors in one field (autonomous Deere tractor + Carbon Robotics LaserWeeder + DJI drone) lack a common task-scheduling layer; AEF TIM (Tractor Implement Management) is the partial answer but cross-OEM still requires custom integration.
9. Tools and simulators
| Tool | Use |
|---|---|
| Webots | Open-source ag robot sim; Clearpath Husky model + crop rows |
| Gazebo / Ignition | ROS 2 standard; Bonn UoB sugar-beet field model is free |
| Isaac Sim | NVIDIA; Omniverse-based; Isaac for AMR + Isaac for Robotics extensions usable for ag |
| PolyVerif / Cognata | Originally automotive, used for ag with custom plug-ins |
| AGX Dynamics | Algoryx; precise soil-tire (Bekker) sim |
| FarmSim | Crop-canopy + tractor sim (Wageningen, Bonn) |
| Pix4D / Agisoft Metashape | Drone photogrammetry, orthomosaic + 3D model |
| DroneDeploy | Field mapping + NDVI |
| APSIM | Process-based crop growth model |
| DSSAT | USDA process-based crop model |
| OpenDroneMap | Free photogrammetry alternative to Pix4D |
| QGIS + GRASS | Free GIS for field-level analysis |
| ROSAg (community) | Lightweight ROS 2 ag-perception bundles |
A sim-to-real pipeline for in-row weeding (Bonn UoB, 2022): synthesise sugar-beet rows in Gazebo with domain randomisation (lighting, weed density, soil texture), train YOLOv5 weed detector, deploy on Jetson Xavier in BoniRob, achieve 75% precision / 82% recall on real Heeslingen plots after one round of real-data fine-tuning.
10. Case studies
John Deere See & Spray Premium (2022) and Ultimate (2023)
Stack: 36 cameras (10 fps each) on a 36-foot (~11 m) boom; per-camera Jetson AGX Xavier; ResNet-based per-plant weed-vs-crop classifier; individual nozzle solenoids at ~20 cm spacing. Spray decision-to-actuation latency: < 100 ms. Speed: up to 15 mph (~24 km/h). Reported chemical reduction: 60–90% in soy, corn, cotton at parity yield (John Deere 2022, 2023 case-study white papers). Acquisition: Blue River Technology, USD 305 M, 2017. The Premium kit is a retrofit on existing 400/600 Series Sprayers; Ultimate is a full carbon-fibre boom system. As of 2026, See & Spray Ultimate is offered as a service (acres-priced) in addition to capital sale, validating the per-plant economics.
Carbon Robotics LaserWeeder (2024)
Stack: 30× 150 W CO₂ lasers in a 6 m towed implement; 12 cameras with onboard GPU per camera; YOLOv8-based weed classifier at 30 Hz; galvanometer mirror per laser steers the beam to the weed apex. Field rate: ~3 km/h, ~200,000 weeds/h per unit; 5,000 weeds eliminated per row-foot. Power: 30 kW peak via tractor PTO + onboard genset. Chemical use: zero. Reported in carrot, onion, lettuce, broccoli trials 2023–2024; “G2” model launched 2024 reduces unit cost ~40%. Carbon Robotics has > USD 70 M in venture funding (2024). The LaserWeeder is the first commercially-successful fully chemical-free row-crop weeder.
Naïo Dino (open-field vegetable weeder)
Stack: 1,000 kg, 4WD with independent steer (4WIS), RTK-GNSS + 2× front cameras + 1× rear; battery-electric (3 hr per swap); modular tool-bar (Treffler harrow, finger weeder, mechanical hoe, spot-spray module). Reported sales 250+ units across France, Spain, Italy, US (2024 Naïo announcement). Open ROS-based stack with proprietary safety layer. Reference customer: French organic carrot growers report 60% reduction in hand-weeding labour. Naïo also produces Ted (vineyard) and Oz (small-farm) on the same software platform. Acquired strategic investment from BASF in 2022 in joint development of robot + biological-input integration.
11. Cross-references
[[Robotics/mobile-base-wheeled]]— Ackermann, skid-steer, 4WIS; the chassis under a field robot[[Robotics/multirotor-design]]— DJI Agras T-series and other spray drones[[Robotics/computer-vision-robotics]]— per-plant detection (YOLO, Mask R-CNN, SAM)[[Robotics/sensors-perception]]— RGB-D, multispectral, hyperspectral, LiDAR[[Robotics/end-effectors]]— picking grippers; soft suction; finger sensing[[Robotics/path-planning]]— coverage planning, boustrophedon, headland turns[[Robotics/slam]]— vision + LiDAR SLAM under canopy when GNSS fails[[Robotics/rl-for-control]]— sim-to-real for in-row driving and harvest grasping[[Robotics/safety-standards]]— ISO 18497 (agricultural automated machinery)[[Engineering/transportation-engineering]]— off-road vehicle dynamics[[Engineering/environmental-engineering]]— pesticide drift, soil compaction, water[[Engineering/classical-control]]— autosteer outer loops[[Engineering/supply-chain-management]]— farm-to-fork traceability stack
12. Citations
- Pedersen, S. M.; Fountas, S.; Have, H.; Blackmore, B. S. (2006). “Agricultural robots — system analysis and economic feasibility.” Precision Agriculture 7(4): 295–308.
- Pedersen, S. M.; Lind, K. M. (eds.) (2020). Precision Agriculture: Technology and Economic Perspectives. Springer.
- Bechar, A.; Vigneault, C. (2016). “Agricultural robots for field operations: Concepts and components.” Biosystems Engineering 149: 94–111.
- Bechar, A.; Vigneault, C. (2017). “Agricultural robots for field operations. Part 2: Operations and systems.” Biosystems Engineering 153: 110–128.
- Oliveira, L. F.; Moreira, A. P.; Silva, M. F. (2021). “Advances in agriculture robotics: A state-of-the-art review and challenges ahead.” Robotics 10(2): 52.
- Reid, J. F. (2011). “The impact of mechanization on agriculture.” The Bridge 41(3): 22–29. (National Academy of Engineering).
- Erdle, K.; Mistele, B.; Schmidhalter, U. (2011). “Comparison of active and passive spectral sensors in discriminating biomass parameters and nitrogen status in wheat cultivars.” Field Crops Research 124(1): 74–84.
- Slaughter, D. C.; Giles, D. K.; Downey, D. (2008). “Autonomous robotic weed control systems: A review.” Computers and Electronics in Agriculture 61(1): 63–78.
- Sørensen, C. G.; Bochtis, D. D. (eds.) (2014). Conceptual model of a future farm management information system. CIGR.
- Bonirob: Ruckelshausen, A. et al. (2009). “BoniRob – an autonomous field robot platform for individual plant phenotyping.” Precision Agriculture ‘09: 841–847.
- HortiBot: Jørgensen, R. N.; Sørensen, C. G.; Maagaard, J. et al. (2007). “HortiBot: A system design of a robotic tool carrier for high-tech plant nursing.” Agricultural Engineering International IX, manuscript ATOE 07 006.
- Bekker, M. G. (1956). Theory of Land Locomotion. University of Michigan Press.
- Wong, J. Y. (1989). Terramechanics and Off-Road Vehicles. Elsevier.
- ISO 11783-1 to -14 (2017–2023). “Tractors and machinery for agriculture and forestry — Serial control and communications data network (ISOBUS).” International Organization for Standardization.
- ISO 18497 (2018, revised 2024). “Agricultural machinery — Safety of highly automated agricultural machines.”
- AEF Functionality Specifications (2024). Agricultural Industry Electronics Foundation.
- 14 CFR Part 137. “Agricultural Aircraft Operations.” Federal Aviation Administration.
- John Deere (2022). See & Spray Premium and Ultimate Product Brochure.
- John Deere (2023). Autonomous 8R 410 Tractor — Technical Overview.
- Carbon Robotics (2024). LaserWeeder G2 — Field Trial Report 2023–2024.
- Naïo Technologies (2024). Dino and Oz Product Datasheets.
- DJI (2024). Agras T50 Product Specification, V1.4.
- Lely (2024). Astronaut A5 Investor Day Briefing.
- USDA NASS (2022). Farm Computer Usage and Ownership.
- USDA AMS (2024). Adverse Effect Wage Rate — H-2A Schedule.
- AGCO Corporation (2024). Annual Report 10-K.
- CNH Industrial (2024). Annual Report 20-F.
- Deere & Company (2024). Annual Report 10-K.
- Lowenberg-DeBoer, J.; Erickson, B. (2019). “Setting the record straight on precision agriculture adoption.” Agronomy Journal 111(4): 1552–1569.
- Liakos, K. G.; Busato, P.; Moshou, D.; Pearson, S.; Bochtis, D. (2018). “Machine learning in agriculture: A review.” Sensors 18(8): 2674.