--- title: "A tea robot picks a bud every 0.6 seconds. The daily arithmetic is less flattering" date: 2026-11-09 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-robot-tea-picking-robots language: en --- # A tea robot picks a bud every 0.6 seconds. The daily arithmetic is less flattering > Chinese labs have built tea-picking robots that pick a bud in about a second, but daily output still hovers near one skilled picker, for now. On a hillside in Yongchuan, the pickers start before the mist lifts. Each one works a row with both hands, thumb and forefinger closing on a single bud and pulling upward, the motion repeated thousands of times before lunch. It looks simple. It is one of the hardest things in agriculture to copy. In April, a small machine was driven between those same rows. It had two arms, a camera rig on its nose and a container behind, and it picked continuously without a tea break. The audience was impressed by the speed. The interesting question was what happened after eight hours. ## Key takeaways - A second-generation tea-picking robot developed by the Chongqing Academy of Agricultural Sciences and Chongqing University of Posts and Telecommunications picks one premium bud every 0.6 seconds, with 85.5% recognition accuracy and 94.8% of picked buds graded acceptable. - Its measured output is about 0.8 jin (400 grams) of premium fresh leaf per hour — across a full shift, roughly the daily haul of one skilled picker, not ten. - At least three other Chinese teams — Hunan Agricultural University, Sichuan Agricultural University and the Chinese Academy of Agricultural Sciences' tea institute — are testing machines aimed at the same one-bud-one-leaf standard. - The binding constraint is not speed but stability: sunlight, rain, slope and the economics of a 30-day harvest window. ## Why premium tea resisted machines China solved mechanised tea a decade ago, but only for the cheap end. The machines that sweep across a hedge cut everything in their path, which is fine for tea that will be ground into powder or blended into drinks. Premium tea (名优茶) is different: it is defined by which leaves are taken and how. A picker selecting one bud and one leaf is making a judgement about ripeness, damage and the next flush, several times a second, in glare, wind and rain. That is a perception problem wrapped in a manipulation problem — a soft target that bruises if you grip too hard, attached to a stem that resists if you pull too gently. The robots attack it the same way: a camera and a trained model to find the bud, then a gripper or a small shear to take it. The Chongqing team's second-generation machine moved from two rows to one, added a second arm so both sides work simultaneously, and replaced its camera with one that copes better with glare and low light, extending the working day at both ends. ## Three machines, three different numbers Chinese tea robotics is not one project but a cluster of them, each optimised for a different cost structure. - The Chongqing machine reports 0.6 seconds per bud, 85.5% recognition and 0.8 jin per hour of premium leaf, with 94.8% of buds meeting grade. - A student-led machine at Hunan Agricultural University, demonstrated in May, picks twice a second, climbs slopes up to 35 degrees, and adjusts its own height and width to different tea varieties. - The Chinese Academy of Agricultural Sciences' tea institute showed a near-adult-height robot in Hangzhou in May that takes one to three seconds per bud at about 90% accuracy, and is still in field trials. A fourth data point comes from Sichuan, where a team that has spent years collecting images across the province's tea districts reported more than 90% recognition and better than 80% picking success in a 2026 demonstration, with three to five machines roughly matching one worker. Notice that the last claim points the other way from the headline numbers. A machine that picks a bud in 0.6 seconds sounds overwhelming. A machine that needs three to five units to replace one person sounds like a prototype. ## The arithmetic of a 30-day season Both things can be true, because bud speed and daily output are different quantities. The Chongqing robot's 0.8 jin an hour, run for a full shift, is around 6 jin a day. In Anhui's Shexian county, a picker reported gathering about 5 jin of fresh leaf on a good day, and is paid ¥150–180 (≈ US$21–25 / HK$165–198) a day. In Guizhou's Duyun, the spring rate was ¥120–150 (≈ US$17–21 / HK$132–165) a day, up about 5% year on year. So the machine is roughly at parity with a good pair of hands, not ten times better. That matters because the labour it replaces is seasonal, cash-paid and increasingly hard to recruit: a township in Chongqing's Xiushan county organised 800 pickers a day this spring and still fell short of the 1,000 it needed. The tea case exposes a general rule in agricultural automation: the robot is judged on its worst hour, not its best second. A picker who is 30% slower but never jams is worth more in a 30-day window than a machine that is twice as fast and stops in the rain — and that rule travels well beyond tea. ## What actually breaks in the field The published failure modes are consistent across teams. Direct sun washes out the image; overcast light flattens the contrast between bud and leaf; rain puts droplets on the lens and makes the slope slippery. Terraced hillsides are narrow and irregular, so the machine's footprint and turning radius decide whether it can enter at all. Then there is damage. A bud that is crushed instead of cut oxidises early and shows up as a brown edge in the finished tea — a defect that costs more than the labour saved. The Chongqing team's 94.8% pass rate means roughly one bud in twenty still fails, which is acceptable in a trial and expensive at scale. ## Honest limitations The performance figures come from the developing institutions and from local government and state media coverage of their demonstrations; none has been independently benchmarked, and trial plots are not commercial fields. The 0.8 jin-per-hour output and the 5 jin-per-day human figure are not measured on the same site or the same cultivar, so the parity calculation is indicative, not precise. Machine price, service life and maintenance cost are undisclosed, which makes payback impossible to verify — and the national tea research system itself says stability and economics still need work. The machines described here are institutional prototypes, not products you can order. ## What readers can do now - If you buy premium Chinese tea, expect the labour story to change slowly: hand-picking will remain the quality signal for years, and "machine-picked" will keep appearing on cheaper grades. - If you farm, ask vendors for performance in rain and low light, not in a demo tent — those hours decide whether the machine earns its keep. - If you follow agricultural robotics, track the pass rate, not the cycle time. A bud every 0.6 seconds is easy to film; 95% acceptable output all day is the actual product. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-tea-picking-robots Free to quote with attribution and a link to the original.