Humanoid Robots Are Failing Basic Store Trials in China
NextFin News — Humanoid robots were supposed to look like the future of retail labor. In a recent 24-hour livestream, six machines from AgiBot (also known as Zhiyuan Robotics) took trial shifts in a convenience store. Two did not make it through.
That is an unusual ending for a product demo. Vendors usually script a clean success. Here the store’s standards were plain enough that hesitation and misjudgment counted as failure.
One unit assigned to patrol and restock waited for instructions instead of filling empty shelves on its own. Another, given cleaning duties, scrubbed wood grain as if it were a stain, left the store to hunt for a trash bin, and nearly treated a customer’s food as garbage.
Other machines performed better at greeting and packing but still dropped food or lost track of order status at peak times. The gap was not random quality control. Different jobs exposed different limits: initiative, visual judgment, and speed under pressure.
Convenience stores have become a favored proving ground. The work sits between the chaos of a home and the narrow repetition of a pure factory cell. Shifts mix routine tasks with surprise interruptions, which yields useful training data. City locations also put the machines in front of ordinary customers—part test, part advertisement.
Galaxy General has placed robots in FamilyMart stores and built unmanned retail cabins across dozens of Chinese cities. Unitree has sent units overseas, including night restocking trials with a Danish retail group.
None of this yet looks like a threat to human staff. Speed is still low; one reported store visit had a robot take about a minute to fetch a simple item.
Cost is worse. A base model priced near 700,000 yuan can equal well over a hundred months of a junior store clerk’s pay in Beijing. Until reliability and throughput improve, the economics favor people.
Vendors know the real prize is sustained work, not a viral shift. AgiBot’s leadership has framed the goal as measuring how much robots actually do—active units, hours of autonomous work, and useful output—rather than only how much they “think.”
The company has pushed machines into an electronics plant in Jiangxi, where a small team of units completed tens of thousands of operations over several days at very high success rates, and into a large theme-park deployment in Zhuhai for guidance and guest interaction. Management has tied that “put them to work” approach to strong first-half shipments and revenue claims against domestic peers.

The wider market is growing quickly from a small base. Industry trackers put global humanoid shipments in the first half of 2026 in the mid–twenty thousands, with China accounting for the large majority. What those machines do after delivery is less impressive.
Research breakdowns show entertainment, education, and data collection still dominating use, with guidance, manufacturing, and logistics as smaller slices. Secondary sellers describe many units cycling through one-off shows and exhibitions rather than staying on a payroll-like schedule.
That pattern is the quiet risk for the sector. Buyers who purchase curiosity eventually ask for results. Robots that cannot restock without prompting, clean without false alarms, or serve food without damage get rotated out—or resold.
Manufacturers that only know how to ship units into demos will feel the same pressure as the machines that failed their trial shifts. The job, for both, is to remain useful after the cameras leave.
更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App
内容来自 钛媒体:引领未来商业与生活新知,版权归原作者所有。
阅读原文 ↗