# Agricultural robotics and autonomous tractors Why perception and liability, not propulsion, are the binding constraints on field robotics; what ISO 18497 actually asks a manufacturer to prove; and why the case for swarms of small machines is an argument about soil compaction. Driving a machine across a field was solved long ago. The hard part is knowing what is in front of it, and who answers for the mistake. Source: https://en.bioecon.ru/docs/agri-food/agri-tech-precision/agricultural-robotics-autonomous-tractors/ Updated: 2026-09-06 An autonomous field machine is not solving the problem it appears to be solving. Moving across a field is done: the RTK auto-steer described on the [precision agriculture page](../precision-agriculture-agritech/) has guided implements to within a couple of centimetres since the 1990s, and it needs no machine learning whatsoever. What remains unsolved is everything else — recognising what lies ahead, deciding whether it is dangerous, and carrying responsibility for the error with nobody in the cab. ## Perception in an unstructured scene A factory robot works in an environment designed around it. A field is designed by weather. Illumination swings across four orders of magnitude between noon and dusk, dust and canopy scatter a laser return, vegetation moves in the wind, and the object that must not be driven over could be a child in tall grass, a fallen post or a resting animal — while 99.9% of similar sensor returns are stems. Hence sensor redundancy: lidar gives geometry but distinguishes grass from obstacle poorly; cameras give semantics but depend on light; radar sees through dust at coarse resolution. No single channel satisfies the safety requirement alone, so they are fused, and the cost of that fusion, not engine power, sets the price of the machine. The false-alarm rate is a separate difficulty. A machine that halts for every suspicious shadow needs an operator standing by, which erases the point of autonomy. The threshold has to be placed between an unacceptable miss and an economically unacceptable stoppage frequency, and that trade is the engineering substance of the sector. ## The standard, and liability ISO 18497 is the safety standard for highly automated agricultural machinery; the 2018 edition was a single document, and since 2024 it is a multi-part series separating design principles, obstacle-detection performance and the requirements on a supervised working zone. Its central move is that a manufacturer does not demonstrate "the machine sees a person" but rather that, within a stated and bounded operating condition, the system delivers a specified level of risk reduction. Hence the practice of restricted operational domains: a field with a verified perimeter, daylight, a known crop. Outside the declared domain liability is undefined, and it is this legal boundary, more often than any technical one, that stops deployment. ## The case for swarms The second motive for autonomy is mass. A heavy tractor compacts soil, and subsoil compaction below the tilled horizon is effectively irreversible on a decadal timescale: it reduces infiltration, restricts rooting depth and drives denitrification losses from waterlogged layers. The classical answer is controlled-traffic farming, which confines all wheelings to fixed lanes. The robotic answer is different: a machine that carries no human needs no cab, no chassis to support one and no power reserve for road speed, so several light units can do the work of one heavy one. Weeding and seeding platforms weighing hundreds of kilograms rather than ten tonnes are in serial production. The economics do not follow automatically. Ten machines mean ten sensor suites and ten failure points; the gain appears where the operation is valuable per plant — weeding, spot spraying — not where it is valuable per hectare.