The harvester comes down the paddy with nobody in the cab. On the ridge, a technician holds a tablet and lets it pass. A grain truck follows twenty metres behind, takes the load at the turn, and drives itself back towards the dryer.
This is Wanlü Smart Farm (万绿智慧农场) in Liucheng township, Dongyuan county, Heyuan, in Guangdong's hilly north-east. The late rice came off on 30 October, before visiting researchers from seven countries.
The machines are the visible part. The part that took five years is the ground they drive on.
Key takeaways
- Wanlü runs 5,300 mu (≈870 acres) of rice and seed production with a core crew of 10 to 20 people.
- Its team reports a 50-fold efficiency gain, measured in field labour, not in cost per tonne.
- The precondition was land: more than 2,000 mu transferred in under a month in 2020 and rebuilt into level, contiguous fields.
- A space-bred variety and variable-rate fertiliser cut nitrogen use by 28% in early rice and 22.5% in late rice, per the operator.
- None of the headline figures has been independently audited, and there is no side-by-side control field.
What is actually running
The farm splits into roughly 2,000 mu of seed production and 3,300 mu of rice. In the core block the whole cycle — ploughing, sowing, spraying, harvesting, grain transport — is handled by driverless tillers, seeders, harvesters and trucks, with drones doing plant protection and scouting. A cloud platform dispatches the fleet.
The number the operators cite is not speed. Mechanical harvest loss is held under 1.8%. One point of grain left on the ground is money, so that is what a manager watches.
What "50 times" is measuring
The efficiency claim comes from Luo Xiwen (罗锡文), a professor at South China Agricultural University and an academician of the Chinese Academy of Engineering, whose team built the technical base.
Field work at this scale used to need more than 2,000 labourers; the current chain needs 10 to 20 people in the back office and on maintenance. Efficiency, he says, is more than 50 times the traditional manual model.
Read the denominator carefully. It is field labour hours, replaced by machine hours and a small skilled crew. It is not cost per tonne, revenue per mu, or profit. A farm can post a 50-fold labour gain and still lose money, because the capital in the yard is expensive and the season is short.
It also assumes a field the machines can use. That was the expensive part.
The bulldozers came first
In 2020 the land was fragmented: plots at different heights, boundaries running through them. Luo recalls that the project took less than a month to transfer more than 2,000 mu. High-standard farmland construction (高标准农田) then merged small plots into large ones and levelled them.
A driverless harvester only made sense afterwards. A machine following a pre-set route needs a field with a route in it. A drone spraying a uniform block needs a block, not fourteen strips separated by bunds.
The farm's own account of the sequence: abandoned land, a single-point unmanned demo, then a full-chain smart farm, over five years. Automation is the third act.
Agronomy did as much as autonomy
The rice is Huahang Xiangyinzhen (华航香银针), a simiao rice (丝苗米) variety developed through space breeding (航天育种) at South China Agricultural University — seed flown to orbit, returned, and selected over generations on the ground. It is the only conventional simiao variety in Guangdong on the national super-rice list.
Average yield is reported above 550 kg per mu, with a best measured block of 723.5 kg. The national average is around 476 kg per mu. Some of that gap is variety, some water and soil, some management; the farm does not separate them.
The clearest automation win is not the harvester. It is fertiliser. Multispectral drone imagery produces a vigour map, and the spreader varies the dose across it — more where the crop is thin, less where it is heavy. The operator reports nitrogen down 28% in early rice and 22.5% in late rice, fertiliser down 25% to 30%, and savings above RMB 300 per mu (≈US$42).
It is a small, checkable kind of AI: a model turns an image into a rate, and a machine obeys it.
Who pays, and who collects
The farm reports annual output above 4,500 tonnes of paddy and more than 500 tonnes of quality seed, with output value above RMB 20 million (≈US$2.8M). Around 2,700 households receive roughly RMB 3.9 million (≈US$550,000) a year in land rent, and about 300 villagers are employed directly. Since 2024 the farm has run 16 training sessions for close to 2,000 people.
The model has spread to more than 20 townships across Heyuan. That matters more than the demo plot: one farm proves a machine works; twenty prove the model survives ordinary local budgets.
Honest limitations
- The numbers are the operator's. Efficiency, yield, fertiliser savings and loss rates all come from the farm and the university team behind it. No third party has audited them, and no control field of conventional rice on comparable soil is published.
- A labour gain is not a business case. A 50-fold cut in field hands says nothing about capital cost, depreciation, or the price the grain fetches.
- The labour that left is not the labour that stayed. The 10 to 20 people are technicians running software and maintaining drones. The 2,000 did work that no longer exists in that form, and this article cannot tell you where they went.
- Land consolidation is the binding constraint, and it is administrative, not technical. Transferring 2,000 mu in a month happened because a county government made it happen. Most places cannot.
- "Unmanned" is a claim about the field, not the farm. Scheduling, repairs, fuel and compliance still need people, and the published detail concentrates on a 1,500-mu core rather than the full 5,300.
What readers can do now
- When a vendor quotes an efficiency multiple, ask what the denominator is: labour hours, cost per tonne, or yield per hectare. The first is easy to win and rarely decides whether a farm stays solvent.
- If you design an agricultural automation programme, fund land consolidation and irrigation before equipment. Machines bought for fields that are not machine-shaped become expensive sheds.
- If you farm, start with multispectral scouting plus variable-rate application. It pays back in fertiliser before it pays back in labour.