Summary of Key Points
Embodied intelligence (in simple terms, robots that can perceive, act, and understand the world like humans) is not a scam created by capital speculation, but a genuine “paradigm shift” in the robotics field that has not yet reached a stage of widespread implementation. It is supported by three major technologies: traditional control systems, reinforcement learning, and large-scale models (LLMs), as well as a strong industrial foundation in China (a large robotics market, comprehensive supply chain, and supportive policies). However, in the short term, there are many aspects that need improvement (many demonstration scenarios are not practical), and long-term challenges such as cost and reliability must be addressed.
I. Technological Aspects: Three Technologies Combined to Create a New Approach
The advancement of embodied intelligence does not rely on a single technology; rather, it is the integration of traditional control systems, reinforcement learning, and LLMs:
- Traditional control systems are not obsolete, but they are no longer sufficient: Traditional robots (e.g., industrial robotic arms) were “fixed” in their functionality—engineers would create a model and write fixed action sequences to perform precise tasks. However, they struggled in complex environments (e.g., legged robots walking on gravel or dexterous hands handling soft objects) because these scenarios could not be accurately modeled.
- Reinforcement learning enables robots to learn skills on their own: Instead of engineers specifying action steps, robots are allowed to trial and error in virtual environments (e.g., learning to stand up after falling thousands of times) and develop their own action strategies through reward mechanisms (receiving “rewards” for correct actions). This approach solves the problem of rapid iteration in learning complex skills, which is impossible with manual programming for quadruped robots.
- Large-scale models enhance the ability to understand tasks: Previously, robots could only “execute commands” without understanding their purpose. For example, when instructed to “get something that can drive a nail,” a robot might not just find a hammer but understand the need for a screwdriver or even a stone. LLMs (such as GPTs) enable robots to understand human language and task context, evolving from being able to perform a single action to completing a task as a whole.
II. China’s Industrial Foundation: A Solid Base, but Not Yet Universal
China has a strong foundation in the robotics industry, but this does not equate to the maturity of embodied intelligence:
- Industrial robots are already in use: In 2024, China accounted for 54% of global industrial robot installations, with local manufacturers exceeding foreign ones for the first time (57%), and production increased by 28% in 2025. These robots are essential for factory automation (e.g., automotive welding) and are a critical part of manufacturing upgrades.
- Humanoid robots are in the pilot phase: In 2025, 18,000 humanoid robots were shipped globally, generating revenue of $440 million. Although the numbers are small, it indicates progress from laboratory prototypes to commercial applications (e.g., performances, education, guided tours).
- Government support is there, but challenges exist: The government aims to double the density of robots in manufacturing by 2025, and humanoid robots need to overcome technical hurdles in brain, cerebellum, and limb technologies. However, among the more than 150 humanoid robot companies, half are startups or cross-industry players, which may lead to duplicate research and product saturation.
III. Short-Term Challenges: Many Demos Are Just For Show, Far from Profitability
Embodied intelligence is not yet at a profit-making stage, and the main issues include:
- Limited application scenarios: 85% of humanoid robots are used in demonstrative purposes (performances, education, guided tours), with few actually working reliably in practical applications such as factories, warehouses, and elderly care.
- Technical hurdles: To be profitable in the long term, these questions must be addressed: How long can robots operate continuously? What is the failure rate? How expensive is maintenance? How are safety responsibilities defined? These cannot be resolved by demonstration videos alone.
- Bubble risk: Many companies rely on financing and promotional materials; their products often have high duplication, leading to waste of research resources. The National Development and Reform Commission has warned about this risk.
IV. Long-Term Opportunities: It’s About Systematic Capabilities, Not Just Dancing Robots
The future of embodied intelligence lies in the ability to integrate technology, hardware, and applications effectively:
- A complete technological ecosystem: Traditional control systems for reliable execution, reinforcement learning for skill acquisition, and LLMs for task understanding are all essential. For example, a robot that can transport goods must first understand the task (“move a box to the warehouse”), plan the route (reinforcement learning), and then accurately grasp the box (traditional control).
- Practical application scenarios: Realistic use cases that generate revenue are needed, such as material handling in factories and companion care in elderly care facilities, where higher costs are acceptable and there is a clear demand.
- Opportunities for professionals: Those with knowledge of traditional control systems, combined with reinforcement learning and LLMs, can quickly master the core capabilities of embodied intelligence. After all, control is the foundation of robotics; without it, even the most intelligent robots cannot function effectively.
Conclusion: Neither Myth nor Denial—It’s a Slow Revolution
Embodied intelligence is not a scam because the underlying technologies and industrial foundation are real. However, it will not change the world immediately due to current shortcomings and technical challenges. It is more like a 5-10-year engineering revolution, and the winners will be companies that can integrate these technologies and applications seamlessly. For the general public, there’s no need to rush to embrace or dismiss it; it is a future trend that requires time. For professionals, this is an exciting era with opportunities for transformation and innovation, as the convergence of these three technologies creates new possibilities.