Agricultural Engineering · Farm Machinery & Power · also Sensors, AI and Robotics in Agriculture, Mechatronics, Advances in Automation and Robotics
Spray drones, machine vision, robots and autonomous tractors — the newest part of the curriculum.
In the exams: Core in the ICAR 6th Deans' Committee B.Tech, with a skill module on drone operation and maintenance; not yet a large exam topic
Agricultural drones, robotics and automation is the part of agricultural engineering where machines begin to sense, decide and act for themselves. It covers unmanned aerial vehicles (UAVs) that map fields and spray crops, the cameras and machine-learning models that let machines recognise plants, weeds and fruit, robot arms and grippers that pick, field robots that weed and scout, the motors and valves that turn computer decisions into movement, and tractors that drive and work without an operator in the seat.
The engineering in this subject is concrete and calculable. How long can a 30 kg spray drone hover on a 30 Ah battery? What flow rate gives 20 L/ha at a 6 m swath and 5 m/s? Why do 50 µm droplets drift a hundred metres? What flight height gives a 2 cm ground sample distance? How far does a weeding robot move while its neural network is thinking? Each chapter gives the relation, works an example, and says where the simple model stops being right. Positioning by GNSS, variable-rate technology and ISOBUS belong to Precision Agriculture and are only referred to here; control theory belongs to Instrumentation, Sensors and Control.
The subject is new in most curricula. The ICAR Sixth Deans' Committee B.Tech curriculum in India makes 'Sensors, AI and Robotics in Agriculture' a core course and adds a skill module on drone operation and maintenance; universities in Europe, North America, Australia, China and Japan teach it as mechatronics, automation or robotics electives. It is not yet a large topic in national entrance or licensing papers, so the practice questions here follow course examinations and the machinery parts of PG, PhD and specialist examinations. Law matters as much as engineering: drone and pesticide rules are national, differ widely and change often, and this manual describes their structure rather than any one country's current text.
An unmanned aerial vehicle (UAV) is an aircraft flown without a pilot on board. Regulators often call the whole set-up, aircraft plus ground station plus data link, an unmanned aircraft system (UAS) or remotely piloted aircraft system (RPAS); 'drone' is the everyday word. In agriculture UAVs do two quite different jobs. Sensing drones carry cameras and other sensors and produce maps: crop vigour, plant counts, weed patches, drainage problems, stand gaps, damage after a storm. Application drones carry a liquid tank or a granule hopper and spray pesticides, foliar nutrients or spread seed and fertiliser. The two jobs pull the design in opposite directions. A sensing drone wants long endurance and a stable, vibration-free camera; its payload is a few hundred grams to a few kilograms. An application drone wants to lift as much liquid as possible for a short flight close to the crop; its payload can be several tens of kilograms, and its whole design is dominated by the power needed to hover with that load.
Multirotors lift and steer with several fixed-pitch propellers: four (quadcopter), six (hexacopter) or eight (octocopter), sometimes in coaxial pairs. They take off vertically, hover and fly slowly at constant height, which is exactly what spraying and close inspection need, but they are inefficient: all the lift comes from rotor thrust, so endurance is short. A multirotor is controlled entirely by changing motor speeds. Speeding up all motors climbs; speeding up the motors on one side tilts the aircraft and makes it translate; and because half the propellers spin clockwise and half anticlockwise, yaw is produced by speeding up one set and slowing the other so that the reaction torques no longer cancel. Hexacopters and octocopters have more redundancy; with the right control software some can land safely after one motor fails, a quadcopter cannot. Fixed-wing UAVs fly on a wing like an aeroplane. Lift costs far less power, so they cover hundreds of hectares per flight and suit mapping large farms, but they cannot hover, need space or a launcher to take off and land, and must keep a minimum airspeed. Single-rotor unmanned helicopters use a main rotor with cyclic and collective pitch control and a tail rotor; they lift heavy payloads efficiently and have been used for rice spraying in Japan for decades, but they are mechanically complex and need skilled maintenance. Hybrid VTOL (vertical take-off and landing) aircraft add lift rotors to a fixed wing so they take off like a multirotor and cruise like an aeroplane.
The propulsion set is a brushless DC (BLDC) motor, an electronic speed controller (ESC) that switches the motor phases, and a propeller. A motor's KV rating is its no-load speed per volt, so a 100 KV motor on a 50 V battery spins about 5,000 rpm unloaded; lower-KV motors with larger propellers are more efficient at producing static thrust. The energy store is usually a lithium-polymer (LiPo) or lithium-ion pack made of cells in series; a lithium-ion polymer cell is 3.7 V nominal and about 4.2 V full, so a '12S' pack is 44.4 V nominal. Capacity is in ampere-hours, energy in watt-hours (volts times ampere-hours), and the C-rating gives the safe continuous discharge current as a multiple of capacity. The flight controller is a small computer running the stabilising control loops. It reads an inertial measurement unit (IMU: three-axis gyroscopes and accelerometers), a magnetometer for heading, a barometer for height, and a GNSS receiver for position. Spray drones add a downward radar or laser rangefinder for terrain following, forward and rear obstacle-avoidance radar, a pump and flow meter, a tank level sensor and the spraying system. A command-and-control radio link connects the aircraft to the ground station or handheld controller, and failsafe logic decides what happens if that link or the battery is lost, typically hover, return-to-home or land.
Payload is what the aircraft carries beyond itself and its energy store: a camera, a spray tank with its liquid, a granule hopper. The maximum take-off mass (MTOM) is fixed by thrust, structure and regulation, so every kilogram of battery is a kilogram not available for liquid, and the trade-off between battery and payload is the central design decision. Designers keep a thrust margin above the hover requirement so the aircraft can climb, brake, resist gusts and still control attitude; a thrust-to-weight ratio of about 2 at full load is a widely used rule for multirotors. Sensing payloads include RGB (visible colour) cameras, multispectral cameras with a few narrow bands including red-edge and near-infrared, thermal infrared cameras, hyperspectral imagers and LiDAR (laser scanners). Gimbals stabilise a camera against the aircraft's motion. The choice of payload follows the question being asked of the field, not the other way round.
A rotor makes thrust by pushing air downward. Treating the rotor as an ideal actuator disc that accelerates a column of air (momentum theory), the induced power needed to hover is P = T^1.5 / √(2ρA), where T is thrust in newtons (equal to the weight mg in hover), ρ is air density (1.225 kg/m³ at sea level in the standard atmosphere) and A is the total disc area swept by all rotors. Real rotors lose more power to blade drag, tip losses and non-uniform inflow; the figure of merit (FM), the ratio of ideal to actual rotor power, is commonly 0.6 to 0.75 for small propellers. Motors and speed controllers lose more, so electrical power drawn from the battery is P_ideal / (FM × η_drive). Three consequences follow directly from the formula. First, hover power rises with the 1.5 power of mass, so a 50% heavier aircraft needs about 84% more power. Second, larger total disc area lowers power, which is why heavy-lift spray drones use large propellers. Third, thinner air at altitude or in heat raises power, so a drone that lifts its full tank at a cool sea-level site may not at a hot upland site.
Take a hexacopter with six 0.76 m propellers (total disc area 2.72 m²). At 20 kg take-off mass the thrust is 196.2 N and the ideal hover power is 1,064 W. With FM 0.7 and a drive efficiency of 0.8 the battery must supply about 1,900 W. A 14-cell, 30 Ah pack stores 14 × 3.7 × 30 = 1,554 Wh; if only 80% is used to protect the battery and keep a landing reserve, 1,243 Wh is available and hover endurance is about 39 minutes. Load 10 kg more liquid (30 kg take-off) and the electrical power rises to about 3,491 W, cutting endurance to about 21 minutes. The extra 50% mass nearly halves endurance. These are idealised hover figures; forward flight, manoeuvring, wind, battery ageing and cold weather reduce them, and real spray drones at full tank are often planned around flights of ten minutes or so. The lesson for an operator is that payload and endurance are traded against each other in a strongly non-linear way.
A fixed-wing aircraft in level flight has lift equal to weight and thrust equal to drag. Drag is weight divided by the lift-to-drag ratio (L/D), so the power needed is P = D × V = (W / (L/D)) × V. A small mapping aircraft with an L/D of 10 flying at 15 m/s with a weight of 30 N needs only 45 W of aerodynamic power, far less than a multirotor of the same weight, which is why fixed-wing machines map larger areas per battery. Range is endurance multiplied by ground speed, and wind changes ground speed: flying into a headwind shortens the distance covered per battery.
Battery current is power divided by voltage, which is why heavy drones use high-voltage packs: doubling voltage halves current and cuts resistive loss in wires and connectors. The C-rating limits continuous current (a 30 Ah pack rated at 10C may supply 300 A). Lithium batteries lose usable capacity in cold weather and with age; deep discharge, over-charging, physical damage and high temperature can cause swelling and fire. Field operations need a charging plan: with charge times often longer than flight times, a spray team typically rotates several battery packs per aircraft and runs a generator or a vehicle-based charging station. Batteries should be stored and transported at partial charge in fire-resistant containers, and damaged packs should be discharged and disposed of under local hazardous-waste rules.
A spray drone carries a tank (often 10 to 50 L on current machines), a pump (diaphragm, peristaltic or impeller), a flow meter, and either hydraulic nozzles on a short boom or rotary (centrifugal) atomisers that throw liquid off a spinning disc or cage. Hydraulic nozzles set droplet size through nozzle type, orifice and pressure; rotary atomisers set it through rotation speed, faster rotation giving finer droplets, which lets the operator change droplet size without changing nozzles. The rotor downwash, the strong downward air stream beneath the rotors, carries droplets into the canopy and is the drone's equivalent of an air-assisted sprayer; its strength depends on aircraft mass, height and speed. The flight controller links the pump to ground speed so that the output per hectare stays constant as the aircraft speeds up and slows down at the ends of passes, and it shuts off spray when the aircraft stops or turns. Granular spreaders replace the tank with a hopper and a metering device feeding a spinning disc.
The effective swath width is the lateral distance between adjacent flight passes that gives an acceptably uniform deposit across the field. It is not the width of the visible spray cloud and it is not the rotor span. Swath is measured, not assumed: collector cards or strings are laid across the flight line, the aircraft flies over them, and the deposit pattern is analysed. Overlapping the measured patterns at different lane spacings gives a coefficient of variation (CV) of deposit for each spacing; the effective swath is the widest spacing that keeps CV within the accepted limit (limits of 20 to 25% are common in aerial-application practice, but the target is set by the standard or authority being followed). Swath depends on flight height, speed, wind, nozzle or atomiser setting and the aircraft's downwash, so a manufacturer's figure is only a starting point. Flying too wide leaves stripes of under-dose between passes; flying too narrow wastes product and double-doses the overlap.
The basic relation is the same as for any boom sprayer: volume per hectare = flow rate ÷ area covered per unit time. With flow Q in L/min, swath W in m and speed v in m/s, area per minute is W × v × 60 m², so application rate AR (L/ha) = Q × 10,000 / (60 × W × v). With speed in km/h the familiar form is AR = 600 Q / (W v). Example: Q = 2.5 L/min, W = 5 m, v = 5 m/s gives AR = 2.5 × 10,000 / (60 × 5 × 5) = 16.7 L/ha. Rearranged, the flow needed for a target rate is Q = AR × W × v × 60 / 10,000. The theoretical field capacity is W × v; at 5 m and 5 m/s it is 25 m²/s, or 9.0 ha/h, before turns, refills and battery swaps reduce it. A 20 L tank at 16.7 L/ha covers 20 / 16.7 = 1.2 ha per load. Drone application volumes are low compared with ground sprayers, often in the range of roughly 10 to 30 L/ha for many uses, so the pesticide is more concentrated in the tank. That is why the label matters: in many countries a product may be applied by drone only if its label or a regulator's approval allows that use and volume.
Calibration checks that what the controller reports is what the aircraft actually does. Run the system on the ground with water, collect the output of each nozzle or atomiser for a timed period, and compare with the displayed flow; replace nozzles whose output differs markedly from the mean. Confirm speed and swath from the flight log and a pattern test. Then plan the job: field area, target rate, tank volume and area per load give the number of loads; area per load and effective field capacity give flight time per load; battery endurance tells whether one battery covers one tank or more batteries are needed. Mix only what the job needs, record batch, product, rate, weather and treated area, and triple-rinse and wash the system after use.
A spray is a spectrum of droplet sizes, not one size. It is described by the volume median diameter (VMD or Dv0.5), the diameter for which half the spray volume is in smaller droplets and half in larger, along with Dv0.1 and Dv0.9. The proportion of spray volume in droplets smaller than 100 µm is a common index of drift potential. Standards such as ASABE S572 classify nozzles by comparison with reference nozzles into categories from extremely fine and very fine through fine, medium, coarse, very coarse and extremely coarse to ultra coarse, and labels in many countries specify a minimum category. Smaller droplets give more droplets per unit volume and better coverage: halving droplet diameter gives eight times as many droplets from the same volume, because volume scales with diameter cubed. Smaller droplets also fall more slowly, evaporate faster and drift further. Every spray choice balances coverage against drift.
A small droplet falling in still air quickly reaches its terminal (settling) velocity. For very small droplets Stokes' law applies: v = ρ_d g d² / (18 μ), where ρ_d is liquid density, d the diameter and μ the air viscosity (about 1.81 × 10⁻⁵ Pa·s at 20 °C). A 50 µm water droplet settles at about 0.075 m/s. Released 3 m up in a 3 m/s crosswind and ignoring downwash and evaporation, it would travel roughly 3 / 0.075 × 3 ≈ 120 m sideways before landing. A 100 µm droplet, which settles four times faster, would travel about a quarter of that. Stokes' law is accurate only at low Reynolds number, so for larger droplets it overestimates settling speed, but the squared dependence on diameter is the key idea: fine droplets stay airborne long enough for wind to carry them off the target. Evaporation makes it worse in hot dry air, as droplets shrink while they fall.
Drone drift depends on the same factors as ground and aircraft spraying: droplet size, wind speed and direction, release height, air temperature and humidity, atmospheric stability and the formulation. Drones add their own. The rotor downwash drives droplets downward and can reduce drift compared with a free-falling spray, but the downwash weakens with height and speed, and the rotor wake and wing-tip-like vortices at the edge of the swath can lift fine droplets and throw them sideways. Flying higher than needed to be safe over the crop, flying fast, and flying in a crosswind all widen the drift plume. Temperature inversions, common at dawn, dusk and on calm clear nights, trap fine droplets in a stable layer near the ground and let them move a long way, so spraying in calm, inverted air is a drift hazard even though it looks ideal. Practical controls are the coarsest droplet category that still gives adequate coverage, the lowest safe height, moderate speed, spraying only within the wind range the label or national rules allow, no spraying during inversions, downwind buffer zones, and drift-reducing adjuvants where approved.
Deposit and drift are measured with collectors: water-sensitive paper cards that stain where droplets land, filter papers or strings that are washed and analysed for a tracer dye, and passive samplers downwind at set distances. Card images give droplet density (droplets per cm²) and percent cover; tracer recovery gives the mass deposited per unit area. Wind-tunnel and laser-diffraction measurements characterise the droplet spectrum itself. Drift studies report deposit as a percentage of the applied rate at distances downwind; these curves underpin the buffer-zone distances regulators set, and because drone studies are newer and fewer than those for ground and manned-aircraft spraying, many regulators still treat drone spraying case by case.
Ground sample distance (GSD) is the ground length represented by one image pixel. For a nadir (straight-down) camera, GSD = (sensor width × flight height) / (focal length × image width in pixels). Example: a sensor 12.8 mm wide with 4,000 pixels across, an 8 mm lens and a flight height of 80 m give GSD = 12.8 × 80 / (8 × 4,000) = 0.032 m, or 3.2 cm per pixel. GSD is directly proportional to height: halve the height, halve the GSD. The image footprint on the ground is GSD multiplied by the pixel count, here 128 m across the 4,000-pixel side and 96 m along a 3,000-pixel side. Choose GSD from the object to be detected: a rule of thumb in image analysis is that an object should span several pixels to be recognised reliably, so finding small seedlings needs a GSD of a centimetre or less, while mapping field-scale vigour patterns can use tens of centimetres. Lower height means finer GSD but smaller footprints, more images, more flight time and longer processing.
Photogrammetry builds a 3D model and an orthomosaic (a geometrically corrected, map-accurate image) by matching the same ground features in many overlapping images. Forward overlap (along the flight line) and side overlap (between lines) are set in the flight plan; recommendations of about 70 to 80% forward and 60 to 70% side are common for crop mapping, with more over uniform canopies where matching is harder. Line spacing = footprint width × (1 − side overlap); distance between exposures = footprint length × (1 − forward overlap). With the footprint above, 70% side overlap gives lines 38.4 m apart, and 80% forward overlap with the 3,000-pixel side along track gives an image every 19.2 m; at 8 m/s that is an image every 2.4 s, which the camera must be able to sustain. Motion blur in pixels equals ground speed × exposure time ÷ GSD; keep it below about one pixel by shortening exposure or slowing down.
Without help, image positions come from the drone's standard GNSS and the resulting map may be shifted or scaled by metres. Accuracy is improved by ground control points (GCPs), surveyed targets visible in the images, or by RTK or PPK positioning of the camera itself (the GNSS principles are covered in Precision Agriculture). Independent check points, not used in the processing, measure the real accuracy. Radiometric quality matters for crop indices: multispectral cameras need a calibration panel image and an irradiance sensor so that reflectance, not just brightness, is mapped, and flights should be at steady light, ideally near solar noon, avoiding broken cloud. Vegetation indices such as NDVI are computed from the calibrated reflectance maps; their agronomic interpretation is treated in Precision Agriculture and in Remote Sensing and GIS.
Drone law is national, written by each country's civil aviation authority, with pesticide law written separately by agriculture or environment regulators. Rules differ by country and change often, so an operator must read the current text for the country, and sometimes the state or province, where the job is done. Most frameworks share the same building blocks: registration of the aircraft or operator; a remote-pilot qualification by test or training; categories of operation set by aircraft mass and risk; a height limit for routine flights, commonly around 120 m (400 ft) above ground; a requirement to keep the aircraft within visual line of sight (VLOS) unless specially approved for beyond visual line of sight (BVLOS); no-fly or restricted zones around airports, military sites and other sensitive places; rules about flying over people; insurance; and logbooks. Spraying adds a second layer because the drone is dispensing a pesticide: the product must be approved for that use, the applicator may need a pesticide licence, and buffer zones, notification of neighbours and record-keeping may apply.
In the United States, small unmanned aircraft under 55 lb (about 25 kg) are flown under 14 CFR Part 107, and an operation that dispenses agricultural products also needs an agricultural aircraft operator certificate under Part 137; heavier aircraft need exemptions. The European Union applies Implementing Regulation (EU) 2019/947, which sorts operations into 'open', 'specific' and 'certified' categories by risk; spraying does not fit the open category, and the EU Sustainable Use Directive (2009/128/EC) prohibits aerial spraying of pesticides except under derogations granted by member states, so drone spraying there depends on national decisions. India's Drone Rules, 2021 classify aircraft by mass, require type certification and a remote pilot certificate, and use an airspace map of green, yellow and red zones; drone spraying of pesticides also follows the agriculture ministry's procedures and needs products approved for that use. Japan, China, Australia, Brazil and other countries have their own regimes, several of them well developed for spray drones. These examples are a guide to the structure of the rules, not a substitute for the current legal text.
Most incidents come from a small set of causes: loss of the control link, battery failure, compass or GNSS errors, collision with wires, trees and poles, and pilot error in a hurry. A safe operation starts with a site survey for power lines, masts, buildings, people, livestock, water bodies and sensitive crops; a pre-flight checklist (propellers, arms and locks, battery voltage and health, sensors calibrated, failsafe and return-to-home height set above obstacles, geofence set, spray system leak-free); a briefing for any visual observer; and a defined take-off and landing area kept clear of bystanders. During flight the pilot watches wind, battery state and the aircraft, not just the screen. Pesticide handling follows the label: chemical-resistant gloves and the stated protective equipment for mixing, closed transfer where available, and clean water for washing. After the job: decontaminate the aircraft, inspect propellers and motors, log the flight, the product and the area treated, and report any accident as national rules require.
Maintenance follows the manufacturer's schedule and the flight log: propellers replaced when nicked or cracked, motors checked for bearing play and heat, frame arms and folding joints inspected for cracks, connectors cleaned of chemical residue, pumps and flow meters flushed, filters cleaned, nozzles or atomisers checked for wear, and firmware kept current. Spray chemicals are corrosive to electronics, so washing and drying after every day's work matters. Batteries are logged by cycle count and internal resistance and retired when they swell, lose capacity or show high resistance. A maintenance log tied to the airframe and each battery by serial number is what lets an operator show an inspector, an insurer or a client that the aircraft is airworthy; many national rules require such records for commercial work. Spare propellers, arms, a spare pump and nozzles kept in the field kit turn a lost day into a lost hour. Calibration of the compass, IMU and flow meter should be repeated after any crash, repair or firmware change, and whenever the aircraft is moved to a distant site.
Machine vision is the use of cameras and image processing to measure or decide something automatically. A digital camera focuses light through a lens onto a sensor of photosites; each produces a pixel value. Colour cameras commonly use a Bayer filter, a mosaic of red, green and blue filters, and interpolate the full colour image. Field of view depends on focal length and sensor size; shorter focal length gives a wider view. Global-shutter sensors expose every pixel at once and avoid the skew a rolling-shutter sensor produces on moving machines. Exposure time sets motion blur, aperture sets depth of field, and gain trades brightness for noise. Outdoors, light changes by orders of magnitude from cloud to sun, and shadows and specular reflections from wet leaves confuse algorithms; many field systems therefore use shrouds and their own lights, or strobes bright enough to overwhelm sunlight. Beyond visible colour, near-infrared, multispectral, hyperspectral and thermal cameras measure what the eye cannot, and depth sensors (stereo cameras, time-of-flight cameras and LiDAR) measure distance.
Segmentation separates the pixels of interest, for example plant from soil, from the rest. In RGB images green vegetation is separated well by colour indices computed per pixel. The excess green index ExG = 2g − r − b, using chromatic coordinates (each channel divided by the sum R + G + B), was introduced for weed and plant detection and is still widely used. Transforming RGB to HSV (hue, saturation, value) separates colour from brightness and makes thresholds less sensitive to light level. A threshold turns the index image into a binary mask; Otsu's method chooses the threshold automatically by maximising the variance between the two classes of the grey-level histogram. Morphological operations then clean the mask: erosion removes small specks, dilation fills small holes, and opening and closing combine the two. Connected-component labelling groups the remaining pixels into objects whose area, shape and position can be measured. Classical pipelines like this are fast and explainable and still run in many in-row weeders and sorting machines, but they struggle when crop and weed look alike or leaves overlap.
A stereo camera has two cameras a known baseline B apart. A point appears at slightly different horizontal positions in the two images; the difference is the disparity d. Depth Z = f × B / d, with focal length f and disparity in the same units (pixels). With f = 1,400 pixels, B = 0.12 m and d = 56 pixels, Z = 3.0 m. Because Z is inversely proportional to d, depth resolution worsens with the square of distance: ΔZ ≈ Z² Δd / (f B). A wider baseline improves far-range accuracy but makes close objects harder to match. Time-of-flight cameras measure the return time of modulated light, and LiDAR scans laser pulses to build point clouds; both struggle in strong sunlight or with dust and spray in the air to different degrees. Depth is what a harvesting robot needs to reach a fruit and what a tractor needs to detect an obstacle.
Examples across the subject: camera-guided inter-row hoes that shift the implement sideways to follow crop rows; in-row weeders that open and close blades around crop plants; spot sprayers that open a nozzle only over detected weeds; fruit-grading lines that sort by colour, size and blemish; yield estimation by counting fruit on trees; plant phenotyping; and the perception systems of harvesting robots and autonomous tractors. The engineering questions are always the same: how many pixels on the target, how much time between image capture and actuation, how stable the lighting, how the system copes with dust, vibration and dirty lenses, and how its errors were measured in the field rather than on clean test images.
Machine learning (ML) builds a model from labelled examples rather than from hand-written rules. In agricultural vision the tasks are image classification (what is in this image: diseased or healthy leaf), object detection (where each object is, as a bounding box with a class: each weed, each fruit), semantic segmentation (a class for every pixel: crop, weed, soil), and instance segmentation (a mask for each separate object). Convolutional neural networks (CNNs) learn layered filters directly from pixels and are the usual starting point; transformer-based vision models are also used. Deep learning needs many labelled images, so transfer learning, starting from a network pre-trained on a large general image set and fine-tuning it on crop images, is standard practice. Classical machine learning (support vector machines, random forests) on hand-crafted features remains useful for small data sets, spectral data and tabular sensor data.
A model is only as good as its data. Images must cover the conditions the machine will meet: growth stages, varieties, soils, light, weather, camera heights and seasons. Labels must be consistent, and labelling protocols should be written down. Data are split into training, validation (for tuning) and test sets; the test set must be kept out of all training and tuning decisions. Data leakage, for example near-identical frames from the same video or the same field appearing in both training and test sets, gives test scores that collapse in a new field. Splitting by field, farm or season is the honest test of generalisation. Augmentation (flips, rotations, brightness and colour changes, crops) widens the data cheaply. Class imbalance is common: a field may have one diseased leaf in a thousand, so a model that always says 'healthy' scores 99.9% accuracy and is useless. That is why accuracy alone is a poor metric.
From a confusion matrix: true positives (TP), false positives (FP) and false negatives (FN). Precision = TP / (TP + FP) is the share of detections that are real; recall = TP / (TP + FN) is the share of real objects found. F1 is their harmonic mean, 2PR / (P + R). For detection, a predicted box counts as a true positive only if its intersection over union (IoU), the overlap area divided by the union area, with a ground-truth box exceeds a threshold, often 0.5. Average precision summarises the precision–recall curve for one class, and mAP averages it over classes. Which error matters more depends on the job: a spot sprayer that misses weeds (low recall) leaves weeds to seed, while one that sprays crop (low precision) wastes herbicide; a harvesting robot that reaches for leaves wastes cycle time. The operating threshold is chosen from the cost of each error.
On a moving machine the model must run on embedded hardware within a latency budget. The machine moves a distance equal to speed × total latency (capture, inference, communication, actuator response) between seeing a weed and acting on it, so the camera must look far enough ahead. At 2 m/s with 0.15 s total latency the machine moves 0.3 m. Models are made smaller and faster by choosing compact architectures, reduced input resolution and lower-precision arithmetic, at some cost in accuracy. Models also drift: a new variety, a new weed or a new camera can reduce performance, so field systems need monitoring, fresh labelled data and retraining. Being able to explain why a model failed matters more in agriculture than a benchmark score.
A robot manipulator is a chain of links joined by joints, revolute (rotating) or prismatic (sliding). Each independent joint adds a degree of freedom (DOF); positioning and orienting a tool freely in space needs six. Common arrangements are Cartesian or gantry robots (three prismatic axes, simple, rigid, used over beds and in greenhouses), SCARA arms (fast in a horizontal plane), articulated arms (human-arm-like, flexible reach), and parallel or delta robots (very fast, for sorting and packing). Forward kinematics computes the tool position from joint values. For a two-link planar arm with link lengths L1 and L2 and joint angles θ1 and θ2, x = L1 cos θ1 + L2 cos(θ1 + θ2) and y = L1 sin θ1 + L2 sin(θ1 + θ2). Inverse kinematics finds the joint values that place the tool at a target; for the two-link arm there are generally two solutions (elbow-up and elbow-down), and points beyond L1 + L2 or inside |L1 − L2| are unreachable. The reachable region is the workspace. Path planning must also keep the arm clear of branches, wires and supports in a cluttered canopy, which is the hardest part in orchards and greenhouses.
The end-effector is the tool at the end of the arm: the part that touches the crop. Grippers hold by friction (fingers), by enclosing the object (cages and soft fingers), or by suction. A friction grip with two opposed fingers must supply enough normal force that friction carries the weight and the acceleration: N ≥ m (g + a) / (μ n), where n is the number of friction surfaces, usually multiplied by a safety factor, but too much force bruises soft fruit. A vacuum cup holds with force F = ΔP × A; a 40 mm cup at 60 kPa below atmospheric pressure holds 75 N in theory, less in practice because of leakage over hairy or curved skin. Soft grippers made of flexible materials driven by air pressure conform to irregular shapes and limit contact pressure. Detachment methods copy hand picking: pulling, twisting, bending at the abscission layer, or cutting the peduncle with scissors, blades, heated wire or lasers. End-effectors often combine suction to draw the fruit in, fingers to hold it and a cutter to sever the stem, with sensors for contact, force and vacuum to confirm each step.
A harvesting robot repeats a cycle: detect and locate the fruit (vision, often with depth), decide whether it is ripe and reachable, plan a collision-free approach, grasp, detach, and place the fruit in a container without damage. Performance is measured by detection rate, harvest success rate (fruits picked out of fruits present), damage rate and cycle time per fruit. Throughput per hour is 3,600 ÷ cycle time × number of arms × success rate. Reviews of harvesting robots for high-value crops found that, despite decades of research, success rates and cycle times in unstructured canopies lagged human pickers, mainly because fruit are hidden by leaves, clustered, and variable. Progress has come as much from changing the crop as the robot: trellised, planar or high-wire training systems that present fruit in a narrow wall, and greenhouse crops grown in controlled conditions. Robots succeed most readily where the task is repetitive, the environment structured and labour scarce or expensive.
Small field robots for weeding, scouting and seeding are commonly skid-steer or differential-drive platforms. For differential drive with wheel speeds vR and vL and track width b, the forward speed is (vR + vL)/2 and the yaw rate is (vR − vL)/b; equal speeds give a straight line and opposite speeds a turn on the spot. Localisation combines wheel odometry (which drifts with slip), an IMU, GNSS with RTK where the sky is open, and relative sensing of the crop rows by camera or LiDAR, which works under canopies where GNSS fails. Simultaneous localisation and mapping (SLAM) builds a map and locates the robot in it at the same time. Swarms of small, light robots are studied as an alternative to large tractors because they compact the soil less and can work in wet conditions, at the cost of more machines to manage.
Mechatronics joins mechanics, electronics, sensing and software into one controlled machine. Actuators are the part that moves. A permanent-magnet brushed DC motor produces torque proportional to current (T = k_t I) and generates a back-EMF proportional to speed (E = k_e ω); speed is set by voltage and falls as load torque rises. Brushless DC motors replace the commutator with electronic switching by a controller that senses rotor position; they are efficient, durable and almost universal in drones and electric robots. Stepper motors move in fixed steps, commonly 1.8° (200 steps per revolution), and can be positioned open-loop as long as the load never makes them skip steps; microstepping divides each step further for smoother motion. Servo motors are motors with position feedback (an encoder or resolver) in a closed loop and are used where position must be held against varying loads. Mechanical power is torque times angular speed (P = Tω); a gearbox multiplies torque by the ratio, minus losses, and divides speed by the same ratio.
A solenoid converts current into linear pull over a short stroke and is the actuator inside on/off valves, including the nozzle-body valves of sprayers. Pulse-width modulation (PWM) switches a supply fully on and off at a fixed frequency; the fraction of time on is the duty cycle, and the average voltage or flow follows it. PWM drives motor speed controllers and also the pulsing nozzle valves that change each nozzle's flow by changing the duty cycle while the pressure, and so the droplet size, stays nearly constant. Proportional and servo valves meter hydraulic flow in proportion to an electrical signal, linking electronic control to the hydraulic power of a tractor (hydraulic circuits are treated in Hydraulic Drives and Controls).
Rotary motors drive linear motion through lead screws or ball screws: linear speed equals rotational speed times lead, and position resolution equals lead divided by steps per revolution. Belt drives move gantries fast over long spans. Pneumatic cylinders give force F = p × A from compressed air; they are light, clean, fast and naturally compliant, which suits grippers and soft actuators, but are hard to stop precisely mid-stroke. Hydraulic cylinders give much higher forces in a small package and dominate tractor and implement actuation; electro-hydraulic control lets software command them. Each actuator choice balances force, speed, precision, compliance, power source, cost and the dirty, wet, vibrating environment of a farm.
Closing the loop needs feedback. Incremental encoders give pulses as a shaft turns; with quadrature (two channels 90° apart) the controller counts four edges per line and detects direction. Potentiometers and Hall-effect sensors measure angle; load cells measure force; current sensing estimates torque. The controller compares the measurement with the set point and drives the actuator to reduce the error, commonly with a proportional-integral-derivative (PID) law, as covered in Instrumentation, Sensors and Control. Good mechatronic design matches actuator, sensor and controller to the job: no faster, finer or stronger than needed, but with margin for the worst field conditions. Sensors on farm machines live with mud, water jets, fertiliser dust and vibration, so sealing, connector choice and cable routing decide reliability as much as the sensor specification does. Diagnostics matter too: a controller that reports an open-circuit encoder or a stalled actuator lets a technician find a fault in minutes instead of hours.
Automation on farm machines grew in steps: hydraulic draft and position control on the three-point linkage, automatic header-height control on combines, electronic seed-metering drives, auto-steer guidance along GNSS lines, automatic section control, and now machines that can run a field operation without an operator in the cab. Guidance, GNSS and ISOBUS are treated in Precision Agriculture; this chapter takes them as given. Autonomy is described in levels, from an operator who steers with assistance, through an operator who supervises from the seat while the machine steers, turns and controls the implement, to machines supervised remotely from the field edge or a farm office, with one person supervising several machines. Leader–follower systems, in which an unmanned machine follows a manned one, are an intermediate step used for grain cart and multi-machine operations.
An autonomous tractor needs to know where it is (RTK GNSS, IMU, wheel and radar speed), what is around it (cameras, stereo cameras, LiDAR, radar and ultrasonic sensors fused together), and what to do (a mission plan). Coverage path planning divides the field into passes, usually back-and-forth (boustrophedon) parallel tracks, and plans headland turns that respect the minimum turning radius and the implement's width; the pass direction affects the number of turns and so the non-productive time. Obstacle detection must see far enough ahead to stop: stopping distance = speed × reaction latency + speed² / (2 × deceleration). At 10 km/h (2.78 m/s) with 0.5 s latency and 2 m/s² braking, the tractor needs 3.3 m, before the length of any implement or the effect of a slope. A machine that cannot classify an obstacle with confidence must stop and call the supervisor.
Functional safety standards define how control systems that protect people must be designed. ISO 25119 covers safety-related parts of control systems on tractors and agricultural machinery and grades required performance in agricultural performance levels (AgPL a to e); ISO 18497 deals with the safety of highly automated agricultural machines, including obstacle protection and the role of the supervisor. Design principles include a hardware emergency stop that removes power independently of software, fail-safe states (a lost link or failed sensor must lead to a controlled stop), geofences that keep the machine inside the field, redundant sensing for people detection, and clear warning lights and sounds. Legal frameworks for driverless machines on farms and roads differ by country and many are still being written, so an engineer must check national law on unattended machinery, road travel between fields and liability.
Automation is not only full autonomy. Electric drives on seed meters allow row-by-row control and turn compensation; load-sensing hydraulics and electronic draft control reduce operator workload; automatic steering cuts overlap and fatigue; on-board weighing and flow control make rates exact. For any automation the engineer asks: what task does it remove or improve, what does it cost per hectare, what does it save in labour, input, timeliness and compaction, and what new failure modes does it bring? Small autonomous machines can run long hours with low labour cost but low work rate each; large autonomous tractors keep high work rates but need high-value safety systems. The answer depends on farm size, labour cost, field shape, and the reliability of service support in the region.
A B.Tech student uses this subject to master a small set of relations and practise them until they are automatic: application rate from flow, swath and speed; area per tank; hover power and endurance; GSD and overlap; Stokes settling; stereo depth; precision and recall; two-link kinematics; stopping distance. In the drone skill module the student learns the pre-flight checklist, calibrates a sprayer with water, and plans a small mission. In projects, a student builds a simple vision pipeline (ExG, Otsu threshold, morphology) or a small differential-drive robot and measures how well it actually performs in a field, not just on a bench.
An engineer five years into practice may run a drone spraying service, work for a machinery dealer, or test equipment. The daily decisions are which jobs a drone should take and which it should refuse (wind, inversions, sensitive neighbours, products not approved for drone use), how to set swath and rate from pattern tests, how many batteries and chargers a team needs for a day, and how to keep aircraft and records in a state that satisfies the aviation and pesticide regulators. A dealer engineer commissions auto-steer and automated implements and diagnoses faults in sensors, actuators and wiring.
A senior engineer with 25 years' experience designs services and fleets: the economics of drone versus ground spraying across a district, the safety case for an autonomous tractor on a large farm, training programmes for drone pilots, and standards and procedures for testing spray drones and robots. Such engineers advise regulators on buffer zones and operating limits, write procurement specifications that ask for measured swath, drift and safety performance rather than brochure claims, and judge when a new automation is mature enough to recommend to farmers who cannot afford a failure in their one season.
A researcher with a PhD and 25 years of work sets the questions the field has not answered: how rotor downwash and canopy structure govern deposition and drift; how to make detection models that work in a new country without new labels; how to design grippers that pick soft fruit as fast as people without bruising; how to certify the safety of machines whose perception is learned rather than programmed; and how small robot swarms compare with large machines on soil, energy and cost. Such a researcher designs field trials with proper controls, publishes negative results, and trains engineers who can tell a demonstration from a dependable product.
This manual does not give any country's current drone or pesticide rules in operational detail: weight categories, pilot-licence syllabi, distances from airports and buffer-zone widths change often and must be read from the current national text. It gives no product-specific spray volumes, wind limits or droplet categories, which come from the pesticide label and national guidance. Endurance and hover figures are idealised momentum-theory estimates; it does not cover propeller blade-element design, battery chemistry in depth, granular spreading patterns from drones, LiDAR processing, reinforcement learning, livestock robots (milking, feeding), or detailed legal liability for autonomous machines. Published harvest-robot success rates and adoption statistics are left out because figures vary widely by crop, study and year.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions. Drone and pesticide rules differ by country and change often; regulation questions describe the structure of named frameworks and must be checked against the current legal text.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.