The greenhouse air pushes past 40°C by mid-morning, and the ripe fruit will not wait. In Dali County, Shaanxi, an aging workforce still climbs into plastic tunnels to hand-pick winter jujube (冬枣) one berry at a time. A grey robot on wheels has started doing the same job — slowly, and not always perfectly.
The machine is the "Lanxia" (蓝侠) winter-jujube harvesting robot, built by Shenzhen Lanxia Robotics together with Harbin Institute of Technology, Shenzhen (哈尔滨工业大学深圳). It is one of China's clearest real-world tests of fruit-picking robotics leaving the lab for the field, and a useful window into why farm robots are harder than factory robots.
Why a jujube, and why now
Dali County bills itself as China's top winter-jujube county. The scale behind that claim is large: roughly 420,000 mu of orchards — about 28,000 hectares — and annual output above 700,000 tons. The entire industry chain is worth more than 10 billion yuan (≈ US$1.4 billion / HK$11 billion).
That is exactly why automation is tempting. Winter jujube (冬枣) ripens in a tight window from roughly June to September. Picking it is hot, repetitive, and increasingly hard to staff. Ma Hongli, a National People's Congress deputy and village Party branch secretary in Dali, pushed for the robot after hearing farmers describe the physical toll of decades bent inside the tunnels — sore shoulders, bad knees, and a workforce that is literally graying.
What the robot actually does
The first field unit reached Dali in mid-2025. According to reporting from Shaanxi provincial broadcaster and China Agricultural Network, the generation shown in tests works like this:
- Vision first. A camera system judges ripeness, then a flexible fingertip reaches in and grasps only mature fruit, leaving unripe berries on the branch.
- Gentle handling. Early units picked a fruit about every 8 seconds, with an ~85% successful-capture rate and fruit damage kept under 8%.
- Fast iteration. By 2026 the third-generation prototype shown in local media used twin arms and an all-direction chassis that can turn on the spot and slip between narrow rows; in the lab it trims picking time toward roughly 3 seconds per fruit.
- Field senses. The latest units carry environmental sensors that log temperature, humidity, and light inside the tunnel, quietly building data that could later help with yield estimates and pest warning.
The team did not try to build one "universal" fruit robot. Instead they tuned the gripper, sensors, and software to winter jujube's specific shape, hang, and fragility — a reminder that agricultural robotics is usually crop-by-crop engineering, not a single magic machine.
The bigger picture: embodied AI comes to the farm
China's 2026 central "No. 1 document" for the first time named agricultural robots (农业机器人) as a standalone category, signaling policy interest beyond the lab. Researchers at institutions such as Nanjing Agricultural University and Northwest A&F University are pursuing similar picking robots for apples, kiwifruit, and tomatoes, with some multi-arm units reported at 1,700–3,130 fruit per hour in demonstration settings.
The Lanxia case fits a broader push toward embodied intelligence (具身智能) — robots that perceive and act in messy, unstructured real-world environments rather than controlled factories. Orchards are among the hardest such environments: variable light, occluding leaves, uneven ground, and fruit that bruises if squeezed wrong. A robot that can pick a jujube without crushing it is, in miniature, the same problem class as a home or hospital robot that must touch the world without breaking it.
Honest limitations
This is still a pilot, not a workforce. Key caveats:
- Speed and reliability. Even at ~3 seconds per fruit in the lab, a single machine is far slower than an experienced picker, and real greenhouses add leaf occlusion and uneven terrain that cut performance.
- Damage rate. An under-8% bruise rate may be acceptable for some grades but is not yet "packing-line clean" for premium fruit.
- Cost and business model. Multi-arm picking robots remain expensive; the realistic near-term path is a service-team model (like a harvester contractor) rather than every orchard buying one.
- Crop specificity. What works for winter jujube does not transfer directly to apples or grapes; each crop needs its own tuning and its own economics.
- Seasonality. A jujube picker sits idle most of the year unless it can be re-tasked or shared across crops and regions.
What readers can do now
- If you grow or buy specialty fruit: watch regional pilots in Shaanxi, Jiangsu, and Shandong — these are where picking-robot field data is being generated first, and where early commercial service models will appear.
- If you invest or build: track the "robot-as-a-service" leasing angle and the cost curve of vision modules and arms; the winner is unlikely to be the best gripper, but the cheapest reliable per-fruit cost.
- If you follow policy: the 2026 inclusion of agricultural robots in the No. 1 document is the signal to watch for subsidies and standards that decide how fast this scales from demo to deployment.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the real test for farm robots is not the demo video but the per-fruit cost versus a human on a hot afternoon.