Summary of Key Points
On August 19th, Yuzhu Technology, known as the "first stock in the humanoid robotics sector," went public on the STAR Market. Its opening price soared by 629% on the first day but then plummeted by 18% the next, resulting in a market value loss of over 160 billion yuan in just two days. On the same day, the 2026 World Robotics Conference opened, with 300 companies participating. Behind this industry event, the focus of competition has shifted from the "body" (the physical hardware of robots) to three main areas: the "brain" (embodied large models), "fuel" (data), and "hands" (dexterous manipulators). The competition for advanced AI technologies is particularly fierce, with various approaches to data collection. The industry also faces challenges such as valuation bubbles, difficulties in implementing technologies in real-world scenarios, and a lack of consensus on technical pathways.
I. Yuzhu's Public Offering: Calm After the Hype – The Bubble of the "First Humanoid Stock"
Yuzhu's public offering was like a rollercoaster ride: its opening price was 1,100 yuan (with an issue price of 150.8 yuan), and its market value briefly reached 444.9 billion yuan, only to fall to 687 yuan the next day, resulting in a loss of 160 billion yuan. There are two key issues behind this:
- Valuation Bubble: The price-earnings ratio based on the issue price is 219 times, more than five times the industry average of 38 times for general equipment companies. With only 7.44% of the shares being tradable, small-cap stocks are prone to extreme price fluctuations.
- Business Momentum: In the first three quarters of 2025, 73.6% of Yuzhu's revenue came from research and education clients, with only 9% from industrial applications (such as logistics and inspection). In other words, most of its robots are sold to laboratories as teaching tools and have not yet entered factories or households.
Yuzhu's pricing strategy is also typical: the price of its R1 model dropped by 95% from 650,000 yuan in 2023 to 29,900 yuan in 2026. However, this lower-priced model is designed for consumer or development purposes and is not comparable to industrial-grade, full-size robots, indicating that the competition in physical hardware has reached a point where profits are minimal, and the company may struggle to maintain its competitive edge.
II. Shift in Focus: From "Body" to "Brain" – Whose AI Can Be Installed in More Robots?
The focus of competition has shifted from the physical capabilities of robots to their artificial intelligence (AI). The core of this competition lies in "world models" – algorithms that allow robots to "pre-visualize" their environment and predict outcomes before taking action, unlike traditional reactive systems that act on immediate sensory input.
- Diverse Approaches:
- Jiexiaoshi: One of the earliest companies in China to develop world models, its products won first-place ratings globally and raised 3.5 billion yuan in funding within three months.
- Wujiedongli: A company that transformed from automotive autonomous driving technology, using "hidden space world models" to secure a 100 million-dollar order in one year.
- Zibianli: Funded by giants like ByteDance, Alibaba, Meituan, and Xiaomi, with a valuation of over 20 billion yuan, focusing on integrating both "small and large brains."
- Lingjingzhiyuan: Does not sell AI models but rather the computing infrastructure that processes sensory data, reducing latency to 5 milliseconds and collaborating with over 300 companies.
- The Competition for AI: Half of the 43.8 billion yuan in funding for embodied intelligence in the first half of the year went to companies developing AI models. Since robot bodies can come in various forms (wheeled, bipedal), only a few companies will emerge as winners, as they will be able to provide AI for all types of robots.
III. The Battle for Data: The "Fuel" of Robots
Robots need data to learn actions, just as humans need food. However, there are only 500,000 hours of high-quality interactive data globally, while training general-purpose AI models requires tens of millions of hours, creating a significant gap. Companies are exploring six approaches to generate data:
1. Remote Operation: Beijing Humanoid Technology built a 5,000-square-meter base with 120 robots for remote data collection; this method is accurate but slow, capable of collecting only 100 hours of data per day.
2. Crowdsourcing: JD.com mobilized 600,000 people (employees and external partners) to collect data in real-world scenarios (household services, logistics), aiming to gather 10 million hours of data within two years.
3. Motion Capture: Nuoyiteng uses motion capture technology from the film industry to collect data, releasing a 617-hour dataset for robot training.
4. Synthetic Data: Galaxy General has generated billions of hours of synthetic data, but the success rate of simulation in real-world applications is only 12%.
5. Internet Video Data: Mou Shen Intelligence uses 90% of publicly available video (with 3D processing) and 10% of remote operation data.
6. AI-Generated Data: AI is used to create videos from the robot's perspective, automatically tagged with action details, with a single GPU generating 40 frames per second.
Data has become a commodity, with the Jiangsu Data Exchange completing the first transaction of structured embodied data, selling 25,000 pieces for a specific price.
IV. A Critical Examination of Reality: Are Robots Really Ready?
At the conference, a embarrassing incident highlighted the industry's challenges: a robot lost control and fell, thrashing about uncontrollably. This highlights the difficulties in implementing robots in real-world scenarios:
- Difficulties in Implementation: Robots perform well in laboratories but fail in real-world conditions (for example, they may freeze in cold warehouses in the northeast).
- Excessive Hype: The pace of funding in the primary market far exceeds production volumes (funding in the first half of the year exceeded the entire previous year, yet only 40,000 robots were sold).
- Unstable Technology: Different companies use various approaches, with incompatible interfaces and lack of standardized testing standards.
- Ethical Issues: As robots become more human-like (e.g., with eye contact features), questions arise about their practical uses (e.g., replacing delivery workers).
Qian Xuesen pointed out 33 years ago that robots are not complex systems, and AI evolutionism has its limitations. Despite the current surge of investment, only those companies that can overcome practical challenges will survive.
Conclusion
At the exhibition, companies were competing not for physical robot bodies but for advanced AI and data. The "winter" of the robotics industry has not yet arrived, but the players have already begun to stock up on "weapons" (technologies) and "fuel" (data). However, most will likely be mere participants. In the end, only those who understand technology, can successfully implement it, and adhere to ethical guidelines will stand out and thrive in this industry.