Summary of Key Points
Embodied intelligence (simply put, intelligent robots that can physically interact with the world) is currently experiencing a stark contrast in perception: on one hand, it is hailed as a trillion-dollar market and a key to humanity's future advancement towards interstellar civilization; on the other hand, even basic tasks like picking up a cup or moving a chair prove challenging for these robots. The issue isn't a lack of an "intelligence" (AI model) but rather a deficiency in real-world "physical experience data"—such as the appropriate amount of force needed to grasp an object or the ability to avoid obstacles. Wang Mingyue, founder of the Moon Exploration Embodied Intelligence community, drew on Xiaomi and Hammer's experience in integrating software and hardware to build a platform that connects professionals from various disciplines and resources, promoting industry development through events like hackathons. The biggest bottleneck in the industry is the shortage of talent. While China's supply chain and contextual advantages can accelerate innovation, challenges such as valuation bubbles and data accumulation still need to be addressed.
Detailed Explanation
1. Embodied Intelligence: Beyond the Trillion-Dollar Hype, Real-World Challenges
You might have heard about ChatGPT’s capabilities in chatting and drawing, but embodied intelligence focuses on making robots physically active—performing tasks like dealing cards, shooting basketballs, or moving objects. The potential is vast: if robots could replace some human labor, they could create value comparable to that of the global GDP and even undertake tasks beyond human capabilities (such as space exploration). However, in reality, robots are still quite clumsy. For example, they may slip while picking up a cup, struggle to estimate the weight of an object, or fail to avoid obstacles. The reason is that although vast amounts of textual data have been collected over the internet, data on how robots perceive their environment and apply force has not been systematically recorded. Just as children learn to walk by falling countless times, robots must also learn through real-world interactions.
2. The Moon Exploration Community: A Talent Connector for Embodied Intelligence
Wang Mingyue previously worked on projects like Xiaomi's Mi Home smart speakers and whole-house automation systems, understanding the complexity of integrating software and hardware. After the emergence of GPT, she realized that while robot “brains” had become smarter, the industry needed a platform to bring together experts from different fields (mechanics, automation, materials, law, etc.) to solve practical problems. She started with the Tsinghua University Embodied Intelligence Club and founded the Moon Exploration community, which now has 50,000 members and over 300 doctoral researchers in embodied intelligence. The community plays a crucial role by facilitating connections between engineers, factories, and potential investors; some members simply seek someone who understands their work. She didn’t start a company directly but focused on building a network because embodied intelligence requires collaborative innovation across multiple disciplines.
3. Embodied Models vs. Language Models
Language models like ChatGPT are trained on decades of internet data (text, images, videos), and their technology is well-established. However, there is no unified standard for embodied models, which require data collected through real-world interactions—e.g., the pressure exerted when picking up a cup or the direction adjustments needed to avoid obstacles. This data is difficult to obtain directly from online sources and often requires expensive and time-consuming methods like 3D scanning and simulation. Autonomous driving is a limited example of embodied intelligence, as it operates in a specific context (driving). Generalized embodied models need more diverse data, including actions and forces in various scenarios. Some experts predict that the “GPT moment” for embodied intelligence could arrive in 1-3 years, but this depends on sufficient data accumulation.
4. Hackathons as a Platform for Learning
The Moon Exploration hackathon featured projects like robot croupiers and ice hockey robots, as well as teams collecting shooting data to train robots. Critics question the feasibility of completing such tasks within 48 hours. Wang Mingyue views hackathons as opportunities for learning, not just for creating immediate startups. She believes that even participants without extensive experience can gain valuable insights and possibly change their career paths through these events. The focus is on the creative process, not on winning prizes; the goal is to temporarily forget about academic or job pressures and work together on solving problems. Some projects may not make it to the top ten, but they can lay the foundation for future success.
5. China’s Advantages in Embodied Intelligence
China has a significant advantage in its supply chain, which allows for rapid development of robot components and testing in real-world scenarios. However, challenges such as a shortage of talent (especially among top universities) and valuation bubbles exist. Wang Mingyue believes that these issues will eventually be resolved, and the market will sort out companies with weak technologies or products. Her community aims to increase the density of talented individuals, driving industry progress.
In Conclusion
Embodied intelligence represents the future, but it is still in its early stages. It requires more talent, data, and experimentation. Platforms like the Moon Exploration Community play a vital role in supporting this emerging technology by providing a platform for collaboration and learning.