A four-legged machine walks slowly between the rows of a vegetable greenhouse in Shouguang, its camera scanning every leaf.
The farmer who once spent mornings crouching to check for the first signs of disease now watches the findings arrive on a phone.
What changed is not the crop but who — or what — does the looking.
A robot dog named Tianquan
Inside the Shi'antian Agricultural Science and Technology Park, a four-legged inspection robot called Tianquan (天权) patrols the vegetable beds. It was built by Weifang Yuntu Technology and runs an agricultural large model (大模型) known as AIMP. What makes it unusual is the knowledge baked in: the model fuses roughly 40 years of Shouguang's accumulated planting experience with the know-how of more than 260 local veteran growers — the "earth experts" (土专家, tu zhuanjia) whose intuition normally takes a lifetime to build.
The dog carries a multispectral camera and a set of sensors. It reads leaf color and shape, spots the early symptoms of pests and disease, and tracks temperature, humidity, light and carbon-dioxide levels as it walks. Wang Yihong, who runs operations at Yuntu, put the shift bluntly: "We no longer need to train farmers. Now we train the robot dog."
What it does on a normal day
- Patrols the rows. The robot handles more than 60% of routine inspection work that a person would otherwise walk.
- Spots trouble early. By catching disease at the earliest visible stage, it gives the grower time to act before a problem spreads.
- Closes the loop. It can trigger linked spray carts, fertigation machines, ventilation and shade equipment, moving from "sense" to "judge" to "act" without a human in the middle.
- Cuts inputs. Early, targeted action means less blanket spraying; the park reports fertilizer and pesticide use down by more than a third.
Why Shouguang is the right place to try it
Shouguang, in Shandong province, calls itself China's vegetable capital. Its fields sow vegetables across roughly 650,000 mu (about 43,000 hectares) and ship around 5 million tons of produce a year. The greenhouses here have gone through seven generations of design, and the local IoT adoption rate in vegetable sheds now exceeds 85%.
The pressure behind the technology is demographic. Younger people left farming for cities, and those who remain are aging. A machine that walks the rows at any hour answers a real labor gap, not a science fair. Returns are concrete enough to interest growers: net income per mu is about RMB 20,000–30,000 (≈ US$2,800–4,200 / HK$22,000–33,000), and a new generation of "new farmers" is being drawn back by the higher-tech work.
None of this means farmers disappear. It means the scarce skill of spotting trouble early is copied into software, so a single trained operator can oversee far more ground than before. The robot does not replace the person who understands the crop; it extends how many plants that person can realistically watch in a day.
More than a single dog
Tianquan is one exhibit in a broader Shouguang push. Nearby parks run smart irrigation that cuts water use versus flood methods, and IoT platforms that let a phone control spraying, feeding and ventilation. The region is also exporting the model: Shouguang has set up more than 300 modern vegetable parks across over 40 countries, including Russia, the United Arab Emirates and Saudi Arabia, training technical staff and shipping its standards abroad.
For a global reader, that export story is the interesting part. A farming method refined in one Chinese county is now a product being planted in deserts and cold climates overseas.
What a model like this can and cannot do
- Can: monitor tirelessly, compare leaves against a deep memory of local disease patterns, and act fast.
- Cannot: harvest, prune or make the final call on a strange symptom — humans still decide.
- Needs: good baseline data from the specific crop and greenhouse; a model tuned to Shouguang tomatoes will not transfer perfectly to a Kenyan greenhouse without recalibration.
- Costs: the smart-greenhouse retrofit is still out of reach for many smallholders, which raises a fairness question the industry has not solved.
Honest limitations
The input-saving and labor-replacement figures come from the park operator and local media, not from independent field trials, so treat them as indicative. The robot handles monitoring and triggering, not the full cycle of growing and harvesting, so human oversight remains essential. Models tuned to Shouguang's crops and climate may need substantial recalibration elsewhere, and the upfront cost of a smart greenhouse excludes many small farms. Whether the published savings hold across a full season and across crop types is still an open question.
What readers can do now
- Greenhouse operators should pilot one inspection robot on a high-value crop before committing to a full smart-greenhouse retrofit.
- Agtech investors and buyers should ask for independent yield-and-input data rather than vendor-reported percentages.
- Policymakers in importing countries can study Shouguang's exported park model as a ready template for local vegetable production.