On June 30, the topic "embodied intelligence data collectors earn 200+ yuan per day, serving as robot teachers" went viral in Chinese AI circles. Behind it lies 2026's most visible talent shortage in China's embodied AI industry — a three-tier pyramid structure:
The Three-Tier Talent Pyramid
Top tier: Chief scientist level — UBTech reportedly offered a 124 million yuan annual salary for a chief scientist in 2025, setting a domestic AI executive compensation record.
Middle tier: Algorithm/engineering hybrid roles — JD.com opened positions such as "data annotation engineer" and "embodied data test development engineer" at 15K-50K yuan/month. A Shenzhen robotics company listed "robot data analysis engineer" at 12,000 yuan/month, "senior robot data analyst" at 300,000-350,000 yuan/year, and "robot after-sales product design expert" at 25,000-50,000 yuan/month with 16-month annual pay.
Base tier: "Robot teachers" — basic operator positions starting at 200 yuan/day (day shift, weekly settlement), using joystick-like controllers to guide robots through tasks such as grasping objects, opening/closing bottles, and folding clothes, plus data recording.
Meanwhile, the domestic true-robot data market price holds at 500-1,000 yuan/hour (per Mifeng Technology CEO Yao Maqing, revealed in April).
JD.com's 600,000-Collector Drive
- Two-year goal: collect 10 million hours of real-world human video data + 1 million hours of robot-body data
- Scene coverage: eight categories — home, office, factory, logistics, retail stores, restaurants, healthcare, sanitation
- Scale: 100,000 internal employees + 500,000 external industry participants = up to 600,000 people
- Positioning: "the largest data collection campaign in human history"
- Job seekers: low entry barrier (200 yuan/day) plus technical roles at 15K-50K yuan/month make this a high-growth track
- Entrepreneurs: data service providers are the scarcest link in China's embodied AI supply chain
- Traditional annotation firms: pure text/image annotation is being automated away by LLMs, but embodied trajectory data requires humans — a structural safe haven
- VLA researchers: data scale and quality determine VLA model ceilings; expanding Chinese data pools will accelerate iteration on RT-2 / OpenVLA / Pi0 architectures
- Policymakers: a 600,000-person collection effort involves labor, data security, and privacy regulation that needs dedicated rules
- Data quality: collectors with under one week of training raise consistency and trajectory cleanliness issues — Yao Maqing's own "high-quality" figure of 500,000 hours suggests much collected data is unusable
- Coverage bias: JD's eight scenes concentrate on home/office/factory; long-tail scenarios like outdoors, disasters, and extreme weather are barely covered, capping VLA generalization
- Cost control: at 500-1,000 yuan/hour, JD's 10 million hours of video could cost 5-10 billion yuan — who pays?
- Privacy boundaries: real-scene video inevitably contains faces, documents, and home layouts; unclear how informed consent compares to GDPR
- Hardware bottleneck: 1 million hours of robot-body data requires roughly 340 robots running 24/7 for a year (at 8 hours/day each); Mifeng's 2026 target implies thousands of robots
- Open-source leakage: will AGIBOT WORLD 2026 be used directly by US companies to train their own VLA models — a "data export" for Chinese firms?
- Baidu Baijiahao "124 million yuan for a scientist, 200 yuan per day for robot teachers": https://baijiahao.baidu.com/s?id=1863614066088886618
- BOSS Zhipin robot data collector listing: https://www.zhipin.com/job_detail/9ed98e99870356c903N42t-7E1FZ.html
- Tencent News "Stay-at-home mom as data collector": https://browser.qq.com/mobile/news?doc_id=1286a09157b20452
- CSDN 2026 Q2 embodied AI startup roundup: https://blog.csdn.net/2611_95864581/article/details/160337993
Mifeng Technology (incubated by Agibot) plans for tens of millions of hours of data capacity by 2026 and tens of billions of hours by 2030.
Yao Maqing offered a striking comparison: GPT-5's training corpus equates to roughly 10 billion hours, while industry-wide high-quality embodied data totals only about 500,000 hours — a gap of tens of thousands of times.
Agibot simultaneously open-sourced AGIBOT WORLD 2026, positioned as "the world's first real-scenario dataset covering full-domain embodied intelligence research," 100% collected from real environments like commercial spaces, hotels, and supermarkets — unlike prior datasets such as DROID/Robomimic that come mainly from labs.
Companies actively hiring: OEMs — Agibot, Galbot, Unitree, Galaxea, Tastone Intelligence (raised over $455 million Pre-A in April), Spirit AI (2 billion yuan), Qianxun Intelligence (nearly 2 billion yuan); data service providers — Mifeng Technology, Suwei Technology; internet giant — JD.com.
Deep Analysis
The essence of this "robot teacher" hiring wave is that embodied AI competition has shifted from "algorithm + hardware" to "data scale". Companies used to compete on robotic arm degrees of freedom, VLA model architecture, and Sim2Real transfer; now they compete on who can collect enough data in real scenes.
Reason 1: VLA models' data hunger is exposed. VLA (Vision-Language-Action) is the core paradigm in embodied AI, fusing visual perception, language instructions, and action execution. Lab data cannot cover real-world complexity like lighting changes, object occlusion, and pedestrian movement.
Reason 2: Imitation learning is now the dominant training paradigm. Unlike LLMs that passively learn from internet text, embodied AI must actively learn from human demonstrations. Teleoperated trajectory data is imitation learning's fuel — every data point must come from a real human operator and cannot be scraped like text.
Reason 3: China's embodied AI industry has entered a "battle of a hundred robots" phase. From April-May 2026, funding density was extremely high. The first thing funded companies do is buy data — hardware can wait on supply chains, algorithms can be hired, but data either gets collected or it doesn't — a hard constraint.
JD.com's drive marks the first time an internet giant formally treats embodied intelligence as an "infrastructure-level" business. If completed on schedule, China's embodied data pool could jump from 500,000 hours to the 11-million-hour level by 2027 — narrowing the gap from tens of thousands of times to roughly hundreds.
For the job market, "robot teacher" could be one of China's largest new job categories in 2026-2027, potentially creating 1-2 million positions industry-wide — a fourth million-scale flexible employment scenario after food delivery, ride-hailing, and express delivery.