AI models are awakening the "financial prowess" of robots.
2025 is known as the "Year of Commercialization of Robots," and some also call it the "Year of Cash Cows for Robots."
The term "cash cow" is not without reason. At the beginning of 2025, a wave of "money - grabbing" occurred in China's embodied intelligence track: in the first half of the year, there were 91 investment cases in the domestic embodied intelligence industry, with a total investment of 11.037 billion yuan, exceeding the total for the whole of 2024. The turnover on the application side increased by 17 times year - on - year, showing an explosive trend.
Behind the figures is a transformation of the industrial logic—the robot track has crossed the "science - fiction concept" stage and entered the period of large - scale commercial value realization. Its core driving force lies in the accelerating formation of a closed - loop for technology monetization of "data collection - model training - commercial transformation."
The Awakening of "Humanity"
Since the term "robota" was coined by Czech writer Karel Čapek in 1920, the leap - forward development of large models has given robots a "human touch" for the first time. The embodied intelligence model based on the Transformer architecture has increased the accuracy of robots' natural language understanding to 92.3%, approaching the human level. Robots are undergoing a qualitative change from "mechanical arms" to "thinkers."
Traditional industrial robots execute fixed tasks using pre - set programs, and they have significant limitations in environmental adaptability and task flexibility. The current technological breakthrough lies in the embodied intelligence system based on deep reinforcement learning. Through neural network training with a parameter scale in the hundreds of billions, robots have gained human - like environmental understanding capabilities, including multi - dimensional cognitive functions such as context recognition, analysis of the relationships between interaction objects, and perception of emotional states. Through the integrated processing of language, vision, and action signals by the multi - modal large model, robots have achieved the ability to generalize and execute tasks across different scenarios. It is this technological progress that has accelerated the current transformation of robots from being "program - driven" to "cognition - driven."
It should be noted that the key to this "humanity" awakening lies in the support of high - quality and large - scale data sets. For robots to accurately understand the world, they need a vast amount of data covering various scenarios and situations for model training. From data parameters in daily life to those in industrial production, the more and more detailed the data, the more accurate the world cognitive model constructed by robots. In fact, robots can be regarded as intelligent agent entities that make independent decisions based on built - in algorithms and knowledge bases and execute multi - modal tasks.
However, establishing a high - quality data model requires high training costs. To ensure the authenticity, accuracy, and comprehensiveness of the data, a large amount of human and material resources need to be invested to collect data in various scenarios. Data annotation requires professional personnel to carefully label the collected data and tell the model what each piece of data represents, which is a time - consuming and labor - intensive process. According to the research and observation of Sice Think Tank, many domestic enterprises are currently involved in the data annotation industry, including large listed companies. In addition to technology companies, there are also some traditional industry enterprises, and a number of outsourcing enterprises focusing on data annotation services have emerged. From data annotation in machinery industry, education, medical care, finance, literature, to autonomous driving scenarios, audio - video, etc., they provide customized data annotation services for various artificial intelligence and robot R & D enterprises. It can be said that the "China Data Dividend" is accelerating the intelligent iteration in the global robot field.
The Model Begins to "Introspect"
For robots to achieve a higher level of intelligence, self - learning and self - correction are crucial.
Traditionally, when robots learn new skills, it often requires manual writing of a large amount of code and readjustment of the algorithm structure, which is a complex and inefficient process. Now, when faced with a new task, the model can quickly understand the task requirements based on existing learning experience and try to combine and optimize strategies from the past skill library to execute the new task, achieving autonomous learning of new skills. With the support of a good model, robots can monitor the gap between their own behavior and the task goal in real - time. Once a deviation is found, they can quickly trace back and analyze the cause and adjust their behavior strategies independently to ensure that the task stays on the right track.
The transparency and openness of the model play a "catalyst" role in this process. A transparent model allows developers to clearly understand the internal operating mechanism and decision - making logic of the model, facilitating problem - solving and performance optimization. Open - source models bring together the wisdom of global developers. Based on the same open - source model, developers can improve and innovate from different perspectives, accelerating the iterative upgrade of the model. Huang Renxun of NVIDIA recently demonstrated a robot at the 3rd China Supply Chain Expo. It is reported that 80% of the training data for its GR00T large model comes from AgiBot World, and this data set was launched on a domestic open - source platform at the end of July. The opening of data and models has brought more technological inclusiveness to the development of embodied intelligence.
The application of edge - side large models is also of great significance. In the past, robots relied on the powerful computing power of the cloud for model calculation and decision - making analysis. Problems such as data transmission delay and network stability restricted the real - time response ability and application scenario expansion of robots. The edge - side large model moves some key computing capabilities to the local devices of robots. Even in the case of poor network or no network, robots can quickly make decisions and process tasks in real - time with the help of the local edge - side large model. At the same time, the continuous emergence of new architectures has also opened up new paths for the development of robots.
The "Dealer" in the Trillion - Dollar Game
Last year, I mentioned in the article "Industry Observation: Foreseeing the Future, the New Track Arrives as Scheduled" that "whether it is humanoid robots, robot dogs, or robot cats, the earliest and also the largest value - added application scenarios will be in industrial production and the military field. This is destined to make it a new arena for great - power competition." A recent research report by Morgan Stanley believes that by 2050, the global humanoid robot market will exceed the $5 trillion mark. However, due to cost factors, the vast majority of humanoid robots will be applied in industrial and commercial fields, and only about 10% of robots are expected to enter the home environment.
Interestingly, the research report also predicts that the scale of the robot market in 2050 will be almost twice the total revenue ($2.488 trillion) of the world's top 20 automotive OEMs in 2024, and the latter may continue to shrink in the next 25 years. The glory of the traditional automotive industry is being surpassed by the exponential growth of the robot industry.
The change in this market pattern from technology verification to a commercial closed - loop reveals that the inflection point of the industrialization of China's embodied intelligence has arrived. Marina Bill, the President of the International Federation of Robotics (IFR), once said that China will not only maintain an average annual growth rate of 5% - 10% but also take the lead in forming a cluster effect in the service robot field.
When machines learn to think and industries start to take root, whoever masters the continuously updated world model will be able to continue to "be the dealer" in the robot business world.
This article is from the WeChat official account "Jiong Jiong Time," author: Jiong Jiong Time. It is published by 36Kr with permission.