Summary of Key Points
At this year's World Artificial Intelligence Conference (WAIC), embodied intelligence—robots that can perceive the physical world and perform tasks just like humans—became the focus. The number of exhibitors doubled, shifting from merely showcasing technological capabilities to practical applications in real-world scenarios such as agricultural harvesting, automobile assembly, and authenticating sneakers. Additionally, the National Development and Reform Commission has identified "high-quality data supply" as a top priority for the development of artificial intelligence, emphasizing that robots need to be able to understand the real world accurately. This requires high-quality data and models. The current core challenges in the industry include the return on investment (ROI) of robots and system stability. In the future, collaboration among data, computing power, and related ecosystems will be essential, with the ultimate goal of enabling robots to perform tasks that humans are unwilling or unable to do, embodying the principle of "human-centeredness."
I. Embodied Intelligence: From Showoff to Practical Use
In the past, robot exhibitions might have consisted of simple actions like walking a few steps or posing. This year at WAIC, there was a greater emphasis on the practical effectiveness of robots. For example:
- Agricultural Applications: Maimei Technology has developed rubber tapping robots for Southeast Asia, where humans wear protective gear and work in environments full of snakes and insects for six hours; these robots can take over that task. Robots for harvesting durians and coconuts are also on the way.
- Industrial Applications: On automobile assembly lines, three robots collaborate to install speaker components—one grabs the parts, and two assemble them, making the process more efficient than human labor. The AI-powered robots from Qianjue Robot and Dewu can inspect details by flipping objects and use spectroscopy to measure changes in shoe inserts, identifying genuine sneakers with greater accuracy than the human eye.
These examples demonstrate that embodied intelligence is no longer just for show; it targets tasks that are tedious, dangerous, or require precision.
II. Data as the Foundation of Embodied Intelligence
For robots to function effectively, they need data to understand the world. For instance, to pick apples, a robot must learn through training that these are fruits, which ones are ripe, and how to sort them by size. Li Nan from Maimei Technology pointed out that many robot manufacturers lack suitable datasets to optimize their systems, preventing them from accurately recognizing objects.
- Industry Challenges: Companies are finding various ways to obtain data: Maimei uses national census data, research collaborations, and project outcomes; Shizhihang collects data from automobile wiring assembly processes and applies it to other flexible assembly scenarios; Youaizhihe trains models with real industrial data to reduce errors between simulation and reality.
The government's emphasis on high-quality data supply aims to address this issue of lacking the necessary learning materials for robots.
III. Industry Barriers: Slow ROI and System Instability
Companies using robots are concerned about the time it takes to recoup their investment (ROI) and the reliability of the systems.
- Agricultural Challenges: Robots can be easily introduced to farms, but maintaining them is challenging. Low-value crops (like wheat) are not cost-effective for robot use, while high-value crops (such as rubber and durians) are more profitable. Additionally, the lack of standardized farming practices (e.g., unevenly grown trees) makes it difficult for robots to work efficiently.
- Industrial Challenges: Even in industrial settings, robots must be highly reliable. In automobile assembly, any mistake can be critical, so models and data need continuous refinement to ensure accuracy.
These factors contribute to the slow widespread adoption of robots in many industries—while the technology is feasible, it may not be economically viable or stable enough.
IV. Underlying Technologies: Computing Power and Ecosystems
Robots rely on chips for processing data and running models. To cover a broader range of applications, a collaborative ecosystem is necessary.
- Role of Chips: Chip manufacturers need to develop foundational technologies that support advanced models, such as those that integrate vision, hearing, and movement (e.g., VLA), to make robots more intelligent.
- Ecosystem Collaboration: No single company can achieve intelligent manufacturing on its own; global partners are needed. This includes chip providers offering computing power, data companies supplying relevant data, and robot manufacturers developing hardware, creating a mutually beneficial ecosystem.
V. Human-Centeredness: The Warmest Essence of Embodied Intelligence
The ultimate goal of embodied intelligence is not to replace humans but to liberate them. For example, robots can perform tedious or dangerous tasks like rubber tapping throughout the night, freeing humans from such labor. Li Nan emphasized that this reflects the true meaning of human-centeredness—using technology to free people from burdensome and hazardous work.
Conclusion
Embodied intelligence has moved beyond being a concept to becoming a practical solution. However, to truly become intelligent partners for humans, challenges in data availability, cost, and ecosystem integration must be addressed. When robots can perceive the world like humans and perform tasks reliably, the day of machines helping us reduce our workload will be near. All these efforts should revolve around human needs, as this is the core direction that embodied intelligence should pursue.