机器人何时迎来“ChatGPT时刻”？具身智能爆发前夜，行业人士的几点思考

From technological breakthroughs to commercial implementation, the robotics industry is standing at the starting point of a new round of explosion. Driven by capital, policies, and application scenarios, has the robotics industry finally reached its "ChatGPT moment"? Is a humanoid form a must or an option?

Behind these questions, macro data provides some references for the industry.

IDC predicted in July this year that the global robotics market will exceed $400 billion by 2029, with the Chinese market accounting for nearly half. Meanwhile, the global shipment of commercial service robots exceeded 100,000 units in 2024. Delivery robots and cleaning robots led the market with shares of 38.4% and 33.3% respectively, and Chinese manufacturers accounted for a staggering 84.7% of the shipments in this field.

Regarding the humanoid robot market, IDC forecasts that the shipment of commercial humanoid robots in China will reach around 5,000 units in 2025 and increase to nearly 60,000 units by 2030, with a compound annual growth rate of over 95%. All signs indicate that the robotics market is expanding rapidly, driven by labor shortages and technological innovations.

During ROSCon China 2025, Jiemian News reporters interviewed several industry insiders to discuss some of the most pressing issues in the robotics field.

Topic 1: When will embodied intelligence have its ChatGPT moment?

Industry insiders have differing views on when embodied intelligence will reach its ChatGPT moment.

Hu Chunxu, the vice president of the developer ecosystem at Sweet Potato Robotics, told Jiemian News that he is highly optimistic about the future of embodied intelligence and the robotics industry.

"With the development of large models and AI, we are entering an era of intelligence, and robots will inevitably be reshaped by AI. I'm very bullish on the development of embodied intelligence and firmly believe that robots will be widely adopted in the future." Although he admitted that there are still issues with the lack of generalizability, "a robot may perform well in one scenario but fail miserably in another." In his view, this is an inevitable stage in the development process.

Tan Weijia, the secretary-general of the Shenzhen Robotics Association, also pointed out that the penetration rate of robots has been extremely low in the past decade, "only in the single digits." The reason is that high costs are incurred for secondary development every time robots enter a new scenario, which is often unaffordable for enterprises. However, embodied intelligence has breathed new life into the industry by shortening the development and implementation cycle and enabling more basic performance improvements through AI algorithms.

She believes that embodied intelligence may lead to an emergence phenomenon similar to ChatGPT. It could also create commercial opportunities by accumulating data in specific scenarios first.

However, there are also some skeptical voices.

Shi Fengming, the technical leader of the innovative business at Fourier Intelligence, warned Jiemian News that while embodied intelligence is a potential path to achieving general artificial intelligence, there are real technological bottlenecks and commercial challenges. "We should be cautious about excessive short - term hype and maintain a rational and optimistic attitude in the long run."

He emphasized that the most important thing is to solve the fundamental problem of how "intelligence" can effectively and reliably interact with the real physical world.

So, does the attention from policies and capital mean that we are "waiting for an inflection point"? Most industry insiders interviewed by Jiemian News believe that the development will be more of a "step - by - step" process.

Yao Jiajun, a visiting scholar at the School of Information of the Greater Bay Area University, believes that we should combine "long - term optimism" with "short - term pragmatism." On one hand, a real breakthrough in embodied intelligence requires a reconstruction of the underlying architecture. The current mainstream VLA has a strong coupling between information flow and control flow, with a relatively simple design. Coupled with limitations in ontology communication and computing power, it is difficult to maintain stable generalization in non - standard environments. On the other hand, obtaining real - world data is also hampered by human factors.

"Therefore, instead of aiming for general - purpose robots right away, we should first focus on workstations with labor shortages and high risks to achieve'scenario - specific generalization,' accumulate high - value data and process know - how, and then expand gradually," said Yao Jiajun.

Topic 2: Is the humanoid form a necessity or an option?

The future of humanoid robots has always been a topic of debate in the industry. A McKinsey analysis released in June this year pointed out that general - purpose robots come in various forms and do not necessarily have to be humanoid. However, the humanoid shape does have an advantage in adapting to the existing environment. They can move in spaces designed for humans without major modifications to the working environment, which is a unique selling point of humanoid robots.

However, from an industrial application perspective, the current commercial implementation path is more flexible. Liu Yizhang, the head of the Embodied Tian Gong Division at the Beijing Humanoid Robot Innovation Center, mentioned that the domestic humanoid robot market is just in its infancy. Only a few hundred units were sold last year, and the number is expected to increase to about 20,000 units this year. "Most of these robots are used in scientific research and education, and their application in industrial or service scenarios is still being verified."

Echo, a technology expert in the perception and autonomous system at the National and Local Joint Humanoid Robot Innovation Center, also suggested that instead of trying to cover all scenarios at once, we should follow the development path of the Internet and aerospace technology, starting with some special scenarios supported at the national level, accumulating experience, and then expanding.

Yanyan from ZNZX analyzed from the perspective of application structuring. Non - structured scenarios such as home care are technically challenging. In the short term, we should start with semi - structured scenarios and gradually transition. She also mentioned the robot - as - a - service (RaaS) model, which can lower the initial investment threshold and allow enterprises to test the waters before expanding.

Overall, the industry generally agrees that the choice of robot form should be scenario - specific. Humanoid robots are not a necessity for all applications, but they do have a natural advantage in seamlessly integrating with the human living environment. Alternative solutions could involve environmental modifications or the use of other platform solutions.

Topic 3: How to achieve ROI in terms of cost and scenarios?

The matching of cost and application scenarios is crucial for robots to enter the market. Despite the broad market prospects, the actual penetration rate of robots is still very low.

Tan Weijia pointed out that the penetration rate of robots in the manufacturing industry is only in the single digits, and it is difficult to make significant breakthroughs even in intelligent assistance. The reason is that expensive secondary development and deployment are required every time robots enter a new scenario.

In fact, enterprises need a clear return on investment (ROI) before large - scale adoption. Otherwise, even if the equipment can operate 24/7, it may be difficult to recoup the cost due to sub - optimal efficiency. This requires manufacturers to optimize the configuration according to scenario requirements.

Yao Jiajun also added that in non - standard scenarios such as welding, workers are often reluctant to cooperate with data collection for fear of being replaced. He believes that instead of aiming for general - purpose robots immediately, we should first achieve scenario - specific generalization in high - risk or labor - short scenarios and gradually promote technology implementation and return on investment.

Gu Qiang, the co - founder of Guyueju, drew an analogy with the mobile phone industry. He believes that as mass production and technology mature, the cost of robots will eventually decrease. For now, the key is to focus on effective scenarios.

Liu Yizhang emphasized that the real value of humanoid robots lies in the added value of "emotion and service," not just the hardware cost. He pointed out that many enterprises are engaging in price wars, with prices close to or below the cost, which is not conducive to the healthy development of the industry.

Most industry insiders agree that before price cuts, robots need to prove their ability to solve real - world problems. Only then does it make sense to talk about price reduction.

Topic 4: How to solve the problems of data and standardization?

The bottlenecks in data collection and standardization have long restricted the development of the robotics industry.

Hu Chunxu admitted that there is currently no unified standard for data collection in the industry. Different companies have different standards for collecting multi - modal data such as vision, language, and force feedback. The lack of a unified standard means that most of the existing data is of poor quality and cannot be directly used for models such as VLA.

He pointed out that compared with autonomous driving in the automotive industry, robots lack a large number of data samples. "Millions of cars on the road can generate a vast amount of real - world data, but there are not enough samples in the robot scenario. Data is the biggest pain point."

Similarly, Tan Weijia mentioned that the traditional method of collecting data from a single robot configuration is inefficient, and migrating to other configurations requires a large amount of repetitive work. A general method or a world model needs to be established to enable cross - platform migration.

In terms of standardization, the industry is still in the early stages of development. Liu Yizhang revealed that there are currently no established standards for humanoid robots, ranging from the production process to testing standards, performance indicators, and even the interfaces of key components. For example, there are no unified specifications for what constitutes qualified motion safety or how to evaluate the reliability and durability of robots. The lack of standards means that each company is working in isolation, making it difficult to achieve large - scale promotion.

Moreover, enterprises are also cautious about data sharing. Sensor manufacturers and algorithm companies are reluctant to share their core data, fearing it will become their trade secrets. Several industry insiders told Jiemian News that it is difficult for a single institution or country to solve these problems. A more open open - source platform and ecosystem are needed to collaborate on standard - setting.

According to several industry insiders, although general artificial intelligence is the ultimate goal, the path to industrial implementation may be more like "laying eggs along the way" in real - world scenarios. By gradually accumulating data, optimizing models, and reducing costs, we can finally unleash the potential in both productivity and daily life.

This article is from Jiemian News, author: Xu Meihui. Republished by 36Kr with permission.