Jiyuan and the inspection robots — China puts AI inside the state grain reserve

East-Asia (China); national

Content

Grain storage is where China’s post-harvest AI is most developed, and it is an unusual case: the deployer, the researcher, the regulator and the beneficiary are the same state apparatus.

The institution. The intelligent grain depot technology R&D platform is led by Sinograin’s Chengdu Grain Storage Research Institute, a central-level research institution under China’s central grain reserve group, sitting inside the National Engineering Research Centre for Grain Storage and Transport. Reported scale: 42 research and management staff; 30 active projects in the evaluation period including 2 national science projects; R&D expenditure of CNY 67.6 m; technical income of CNY 373 m.

What it built, in four parts.

1. Intelligent sampling and inspection for grain purchase (粮食收购智能扦检系统) — a “robot technology + modular” system performing multi-indicator, whole-process unmanned inspection of grain at intake, with reported detection efficiency more than doubled. It was selected into the 2025 list of ten major science and technology innovation achievements in the grain circulation sector, is reported deployed at 80 grain enterprises nationally with over 4 million tonnes of procurement executed through it, and was inspected by Vice-Premier Ding Xuexiang in July 2025.
2. “Jiyuan” (稷元) grain-storage large model — built on more than 200,000 high-quality grain-storage knowledge units, described by the institute as the foundation for intelligent management across storage enterprises, and admitted to SASAC’s AI “Renewal Community” (焕新社区) platform — that is, into the central state-asset regulator’s own AI showcase.
3. Grain-condition monitoring and warning — a multi-parameter online system covering temperature, humidity, moisture, insects, mould and gas, automatically identifying seven grain-condition modes (including loading, unloading, heating and moulding) and providing 21-day dynamic forecasting, with an embedded AI pest-monitoring system that recognises more than 20 stored-grain pest species and integrates remote collection, identification and warning for pests on the grain surface.
4. A new silo type — air-film reinforced-concrete low-carbon silos: airtightness more than six times the national silo standard, thermal insulation three times a traditional silo, 33% shorter build cycle, over 20% lower operating and maintenance cost, at 9,000-tonne scale.

The work has produced 26 national invention patents and 6 national standards, with three first prizes from the Chinese Cereals and Oils Association.

The commercial parallel from Sinograin’s peers. At the 2026 World AI Conference, three Sinograin AI achievements were selected for SASAC’s showcase of excellent results, including inspection equipment that compresses single-sample testing to under three minutes with imperfect-kernel recognition accuracy up to 97.6%. (Figure from the Sinograin announcement’s own summary line; the page returned an error on fetch, so this is recorded at snippet level pending re-verification — see G-462.) COFCO, the commercial state group, separately introduced AI into large-scale grain procurement and processing, including a digital market-intelligence system. And Beijing’s municipal programme for 2026 continues iterating its “smart grain depots” with the stated aim of raising the share of off-site supervision and mining grain-temperature and grain-condition data for predictive risk warning — a shift, in its own words, from “watching the site” to data-based supervision.

What this unit is doing in the taxonomy

The corpus’s post-harvest/storage unit for China — the first anywhere in the corpus in that sector for this country, and one of its few post-harvest units at all. It also fills the corpus’s G-032 gap (Chinese AI at the processing level) on the storage side.

Distinct from:

• china-food-processing-smart-factories.md — the processing and manufacturing layer; this is the reserve and storage layer.
• china-mara-agricultural-data-resources-2026.md — the ministry’s data estate; this is the central reserve group’s operational system.
• grain-storage-type material elsewhere — the corpus’s other post-harvest units are mostly commercial cold-chain and quality systems; China’s is a state reserve integrity system, with supervision as a first-class purpose.

Why it matters for talks

• Supervision is the point, not productivity. The Beijing programme says plainly that it is shifting from on-site inspection to off-site data supervision. AI in China’s grain reserve is primarily a principal-agent control technology for the state over its own reserve — a purpose that barely appears elsewhere in the corpus.
• Four million tonnes of grain through an automated inspection system at 80 enterprises is one of the largest verified operational reaches of any Chinese agricultural AI system in the corpus, and unlike most of it, it is at intake rather than in-field.
• Jiyuan’s 200,000 knowledge units and 21-day forecasting show what a domain model looks like when built on a narrow, well-documented process instead of on general agricultural text.
• Six national standards came out of this work, which is how the corpus’s Chinese standards thread keeps surfacing: the same institutes that deploy also write the rules.
• The 97.6% imperfect-kernel recognition figure — pending verification — is the kind of number that makes a vision system’s value legible; the corpus should hold it until re-checked.

Critical context

• Every figure is from the deploying institution. The 80 enterprises, 4 million tonnes, 2× efficiency and model specifications are Sinograin’s own or its research institute’s. maturity-verification: V1 reflects a state research institute’s publication, not an audit.
• One number in this unit is snippet-level only. The WAIC imperfect-kernel accuracy (97.6%) and the three-minute detection claim come from a Sinograin page that failed to load; they are flagged in G-462 rather than presented as verified.
• “Seven grain-condition modes” and “20+ pest species” describe a rule-and-classifier system as much as a model. The AI label sits on top of a long-standing sensor and inspection discipline; the corpus should not read modern machine learning into what may be improved threshold logic.
• The knowledge-unit count (200,000) is not a benchmark. It measures corpus construction, not accuracy.
• Reserve integrity is not farm-level food loss. This unit is about the state’s ability to know what is in its silos; it says nothing about losses on farms, in transport, or in commercial storage.
• The sandbox does not extend to the commercial layer. COFCO and private mills buy their own systems; the Corpus should not generalise Sinograin’s deployment reach to Chinese grain storage as a whole.