During the National Day holiday, a batch of workers that do not take breaks showed up on the job, some as mascots and some as real labour. Humanoid robots now appear at DQ ice-cream stores, FamilyMart and Chagee outlets, and at the showrooms of BYD, XPeng and Li Auto.
A humanoid from Sharpa runs a DQ store as the only employee, taking an order through the app and executing 55 steps to build an Oreo Blizzard, then overturning the cup to show its arm strength. At a Beijing FamilyMart, a Galbot G1 from Galbot greets customers, answers where items are, and makes coffee by voice, while admitting it has only two fingers and must calculate the best grip angle each time.
The machines rely on three abilities: perception through cameras and lidar, processing in the brain, and execution through legs and arms. They still lag humans on speed. One viewer noted a robot spilled one cup and knocked over another while a person would have made roughly ten in the same time.
Companies are collecting human motion to teach robots. JD.com announced a Suzhou data centre targeting 10 million hours of high-quality data with 100,000 staff and 500,000 public participants. Figure’s Index programme passed 1 million users, 16 million uploaded videos and US$15 million in rewards, aiming to grow collection roughly 100-fold in a year. Mifeng Tech launched a China version, Mifeng Pai, with 20,000 users and 13,000 valid submissions in a month.
Pushback is real. Some Tesla factory workers refused motion-capture suits, and Hyundai Motor’s union called robot deployment job theft and barred automation without a labour agreement. Most human worry about replacement is premature, given the maths.
A base Galbot G1 costs about RMB 699,000, equal to 155 months of a clerk earning RMB 4,500 a month, before electricity and wear. On a three-shift schedule that is the wage of three clerks for 51 months. Yet IDC’s global humanoid tracker shows first-half 2026 shipments near 25,000 units, up 432 per cent year on year, with a market above US$740 million, up 323 per cent.
The brain race is heating up too. After OpenAI released GPT-6 Astra on 3 September, Robocurve wired it to a robot and it topped the RoboDojo benchmark at 22.48 per cent average success across 42 tasks and 2,100 trials. Alibaba’s Qwen-Robot, ByteDance, Zhipu, DeepSeek and MiniMax are all building vision-language-action models, while embodied firms such as AgiBot and Zhizuan answer with their own cortex-cerebellum systems.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://it.sohu.com/a/1084261852_116132.