Unmanned farms as a state delivery system — China builds the ground-robotics layer through promoted-technology catalogues, not commercial products

East-Asia (China); national frame with Heilongjiang, Jiangsu, Xinjiang, Shandong, Ningxia and Beidahuang (XPCC) as the documented provincial and enterprise carriers

Content

China’s ground-robotics layer is not a vendor market that the state later regulated. It is a state delivery system: MARA identifies a technology gap, publishes a promoted-technology catalogue with published performance figures and explicit application preconditions, provincial governments write matching subsidy rules, and state enterprises (Beidahuang, the XPCC) provide the land that makes the preconditions satisfiable. The distinctive artefact is not a machine. It is the precondition clause — the catalogue tells you the plot must be contiguous and at least 5 mu, that machine roads must be at least 4 m wide for large equipment, that ramps must be under 15°, and that satellite and network coverage must be good. Autonomous farming in China is gated on land consolidation and infrastructure, not on software.

The headline number needs its source discipline. MARA reported in February 2025 that more than 1,000 “unmanned farms” (无人农场) were in operation across 23 provinces and municipalities, naming Beijing, Shanghai, Guangdong and Jiangsu. That figure circulates widely through secondary Chinese outlets; the corpus could not locate the MARA primary document and treats it as V1 ministry-reported. The Jiangsu case is the one with a documented, itemised breakdown and is the better anchor: as of January 2025 Jiangsu reported 283 unmanned farms, ¥872M total project investment, 3,643 intelligent machines, and 330,000+ mu-times of unmanned operation in 2024 (MARA 农业农村部 site, 16 January 2025, reporting the provincial agriculture department). The same report gives the economic claim directly: unmanned machinery across the main tillage–planting–management–harvest stages saves about ¥300 per mu in cost, adds about 25 kg per mu in yield, and reduces pesticide use by about 5%. These are provincial self-reported figures with no independent audit surfaced.

The catalogue as the mechanism

MARA’s 2025 十大智慧农业主推技术 (ten promoted smart-agriculture technologies) is the most useful single Chinese artefact on this topic, because each entry ships with technical points, claimed cost-and-benefit effects, and application conditions. The ten span soil testing, machinery, pigs, aquaculture, layer breeding, field crop management, winter wheat, soybean breeding, fruit-and-vegetable sorting, and lychee processing — which is itself the finding: the catalogue is a whole-value-chain instrument, not a machinery instrument.

The ground-robotics entry, 大田农机装备自主作业技术 (autonomous field-machinery operation technology), states five core problems — efficient autonomous driving, precise parameter sensing, intelligent operation regulation, cooperative operation, and swarm control — and claims:

The rice-transplanting exception is the substantively important line in that whole list. China has automated every other stage of rice and wheat production and left transplanting out. That is not a vendor gap; it is the hardest stage — paddy water, mud, and a delicate seedling — and it is the same stage the Heilongjiang 15th Five-Year Plan names as its flagship (“laser weeding robots, driverless rice transplanters”). The corpus should read the transplanting exception as the honest boundary of the current deployment envelope.

The same catalogue makes the precondition logic explicit for the soybean and field-crop-management entries: field crop management requires contiguous cultivated land of at least 5,000 mu and specified machinery ratios (one 375-hp combined tillage unit per 30,000 mu; one self-propelled sprayer per 40,000 mu; one 300-hp+ harvester per 10,000 mu). Large-scale mechanisation is a precondition for the digital layer, not a consequence of it.

The four documented deployment carriers

Heilongjiang — the research-institute-to-field path. This is the corpus’s clearest case of a university laboratory product entering production. The HIT State Key Laboratory of Robot Technology and Systems (付宜利 / Fu Yili team) has a self-propelled intelligent laser weeding robot in working use, identifying and burning weeds in 0.3 seconds with weed-recognition accuracy over 95%; the robot collects crop–weed data autonomously to iterate its model (“the more it is used, the smarter it gets”). A parallel HIT group (姚蔚然, 航天学院自主智能无人系统实验室) has a “plough–plant–manage–harvest” integrated scheme with a BeiDou-navigation system, using deep learning to recognise seedlings and plan routes; a 180 kg-payload agricultural UAV for spraying and lifting claims cumulative service of 900 million mu-times and 70% labour saving. Separately, HGTECH’s Harbin subsidiary 华工智耘 (HGTECH is Wuhan-based, self-described as China’s first laser stock) reports its all-weather intelligent laser weeding robot has completed field validation and entered engineering mass production, with weed-recognition over 99% and removal over 95% — materially better figures than the HIT unit’s, from a company rather than a laboratory, and with no published evaluation basis for either. Heilongjiang’s 15th Five-Year Plan press conference commits the province to build “future farms” supported by laser weeding robots and driverless transplanters, and to expand what it calls China’s first agricultural full-industry-chain large-model application scenario.

Xinjiang — the vendor’s proof-of-scale. XAG’s 超级棉田 (Super Cotton Field) at Yuli county, Bayingolin, has run since 2021 on 3,000 mu of government-consolidated, irrigated, electrified and networked high-standard farmland, managed end-to-end by two people (艾海鹏, 凌磊) with no other manual labour. The season record is documented year by year and includes failures: 2021 wind damage left only ~50% harvestable area at 254 kg/mu; 2022 reached 100% harvestable at 403.6 kg/mu with Micronaire grade A and ~30 mm staple length; 2023 a rare May snow killed more than half the seedlings, 1,500 mu were replanted, and the season still finished at 1,191 t / 420.9 kg/mu; 2024 reached 1,233.95 t at 529 kg/mu, with fine-cotton on 2,332 mu and 668 mu of trial plots, and superior-grade cotton above 96%. By 2026 the management model was reported at 75% unmanned operation rate and promoted to more than 2 million mu across Xinjiang. The corpus’s drone unit corrects XAG’s marketing position; this is the more careful evidence — a vendor-managed, government-consolidated, four-season record with documented weather failures, which is what a deployment claim should look like.

Beidahuang / XPCC — the state-farm carrier. Beidahuang has run 30 smart-agriculture demonstration projects across 21 farms since 2025 and is exploring the “future farm” model at Qixing and Youyi farms, with satellite remote sensing, UAVs and sensors for sensing, a digital platform for aggregation, AI-assisted decisions, and intelligent machinery and UAVs for execution — a closed sense-decide-execute loop in which execution results feed the next round (per 北大荒’s own description, via a company representative). MARA’s action plan names Beidahuang in its addressee list, so the state-farm system is a designated delivery channel rather than an independent actor.

Shandong and Jiangsu — the provincial infrastructure bet. Jiangsu is the most instrumented case: ¥100M per year since 2021 for full-season mechanisation projects, ¥27M since 2020 supporting 18 unmanned-farm integrated demonstrations, a provincial intelligent-machinery industry alliance, first-in-China construction standards for unmanned grain farms, and a subsidy for connecting unmanned farms to the 苏农云 provincial data platform — the data relationship is itself a policy instrument. Jiangsu’s 14th Five-Year digital agriculture plan targets 100 digital innovation parks, 200 digitalised farms, and a 70% provincial digital-agriculture development level by 2025. Shandong reports over 1,000 smart-agriculture scenarios built, more than 20,000 mu of smart greenhouses, 168 smart pastures, 71 national marine pastures, and over 18,000 plant-protection drones. The Wujiang district demonstration in Suzhou specifies the operational decomposition the catalogue abstracts: unmanned tillage (ploughing, deep loosening, harrowing, ridging, levelling), unmanned transplanting with terrain and paddy-boundary recognition, unmanned top-dressing and plant protection combining drones with self-propelled sprayers, and unmanned harvest–unload–transport requiring harvester and grain-cart coordination.

The verification posture, stated plainly

Every efficiency figure in this unit is government- or vendor-reported. None has an independent peer-reviewed evaluation, and the corpus did not surface one. Two things make that more rather than less serious. First, the catalogue’s own precondition clauses mean the deployments are selected for suitability, so the ¥300/mu saving is a saving on land that was already consolidated, contiguously laid out, roaded and signal-covered — it does not transfer to fragmented or unroaded farms, and the catalogue says so. Second, the two laser-weeding figures in circulation (HIT’s >95% recognition, HGTECH’s >99% recognition and >95% removal) are not comparable and neither is audited; both are in production claims, from a laboratory and a commercial vendor respectively.

Against that, the verification infrastructure is genuinely improving and worth naming: MARA’s December 2025 reporting records the first-ever ministry-level quality inspection and testing centre for smart-agriculture equipment and technologies (农业农村部智慧农业技术装备质量检验测试中心), 104 key technologies and 62 whole-machine intelligent equipment items through breakthrough, 14 published standards including 农机田间作业感知数据采集规范, 6 new standards in development including BeiDou positioning/navigation performance test methods, and 16 typical-scenario case templates including a rice smart-field-farm planting model. A ministry that has just built its first inspection centre has not yet used it to test the catalogue’s own numbers.

Two counterweights the corpus already holds. The drone-drift unit documents MARA’s own institute finding ~25% fine-droplet drift and multi-kilometre herbicide movement — the state’s own institution contradicting the deployment narrative is the best available precedent for how this evidence will surface. And the subsidy-integrity cases (Wuhan Audit Bureau using BeiDou terminal data in an investigation; Jilin machinery-subsidy fraud cases) show the terminal data exists in a form that is auditable by the state — the same infrastructure that makes subsidy fraud detectable makes claimed performance checkable, and that check has not yet been run on the performance claims.

Why this is structurally different from the corpus’s other robotics

The corpus already has nine robotics units. None of them is a state delivery system. john-deere-see-and-spray.md and cnh-raven-autonomous-portfolio.md are manufacturer product lines sold into farm markets. advanced-farm-tech-farmwise-carbon-robotics-specialty-crop-weeders.md is a venture-backed weeding cluster. niqo-robotics-india.md is Indian precision hardware. marfrig-agrorobotica-brazil-cattle-carbon.md is corporate carbon accounting.

China’s case differs on three axes, and the differences are what make it a cluster pattern rather than another vendor unit:

  1. The unit of deployment is the province, not the vendor. 23 provinces, a national catalogue, provincial subsidy rules, and provincial data platforms. The vendor supplies a machine; the province supplies the land, the roads, the signal and the plot consolidation that make autonomy feasible.
  2. The binding constraint is documented and physical. Every promoted technology ships with its application conditions. This is a deployment programme that knows what it cannot yet do — the rice-transplanting exception, the 5-mu minimum plot, the 4-m road, the 15° ramp. The corpus’s North American robotics units do not carry this class of self-limitation.
  3. The data relationship is a subsidy, not a contract. Jiangsu pays farms to connect to 苏农云. Farm data in this layer moves because the state attaches payment to the connection, which is the opposite of the vendor-terms problem the corpus documents elsewhere.

The corollary matters for anyone reading the corpus as a diffusion story: China’s autonomous-farming capability is concentrated in the state-farm and consolidated-land tier, and the ability to reproduce it elsewhere is a question of land structure, not of exporting machines. XAG is the one actor in this unit that exports a model rather than a machine, and its 2-million-mu promotion happened inside Xinjiang’s own land-consolidation programme, not outside it.

What the field guide should take from this

New gaps surfaced by this unit

New contested claims surfaced