Large models add to the fire, but "worker" robots are still in the incubation stage
2024-08-27
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Before 2020, an investor asked Wang Xingxing, the founder of Yushu Technology, whether he would make humanoid robots. The latter said categorically that he would not do so. But at the beginning of 2023, Yushu Technology entered the humanoid robot track. Wang Xingxing believed that the core reason for the change was mainly around 2022, under the leadership of Tesla founder Elon Musk, global technology paid great attention to humanoid robots. With the launch of the large language model at the end of 2022, AI technology has undergone a qualitative change, and the industry has seen the potential brought by AI-enabled robots.
With the support of the trend of large models, the popularity of the 2024 World Robot Conference has reached a record high, with 27 robots on display. Through conversations with many industry insiders, the reporter of Yicai.com found that the current consensus in the robot industry is that the track is still in a very early stage, and the combination of large models and robots is the general trend. However, there are still differences in different dimensions such as the robot technology route, the specific forms of humanoid and non-humanoid, and the direction of commercialization.
Gao Jiyang, founder of Xinghaitu, told reporters that the robotics industry is in its early stages of development, practitioners still have different opinions on the technology path, new companies are still emerging, the technology path has not yet converged, and the commercialization cycle is longer. All these factors make the robotics track in the first half of this year more like the Spring and Autumn Period. In the second half of this year or next year, the top players will gradually surface and gather more talents and funds.
Robotic products are flourishing
At the World Robot Conference held from the 21st to the 25th, many new robots were unveiled. The "Tiangong" independently developed by the Beijing Embodied Intelligent Robot Innovation Center realized the world's first full-size pure electric humanoid robot to run anthropomorphically and then show new skills. The world's first orthopedic surgical robot equipped with artificial intelligence deep learning technology was released for the first time, and soft robots broke through the bottleneck in the field of industrial applications.
Take "Tiangong" as an example. When it was first launched in April this year, the robot achieved the world's first full-size, all-electric humanoid robot running in an anthropomorphic manner with its stable trot of 6 km/h. Four months later, "Tiangong" was upgraded again. With the support of the embodied intelligence large model, it has mastered the ability of voice interactive grasping based on the predictive reinforcement imitation learning method based on state memory. When a person issues a voice command, the embodied intelligent robot can complete a set of grasping and releasing actions based on the "open vocabulary target detection and arbitrary object segmentation multimodal model".
Hao Baoyu, vice president of UBTECH Robotics, said that the company has combined end-to-end imitation learning, visual precision recognition, whole-body fine motion control and other technologies, and the industrial version of the humanoid robot Walker S Lite has "joined" a number of automobile factories, such as at the CTU warehousing and loading station in the Zeekr factory, collaborating with employees to perform handling tasks.
Yi Gang, the founder of Titanium Tiger Robotics, which has just released the humanoid robot T230, said that T230 is the first 2.3-meter-tall humanoid robot in China, and is mainly used in heavy-duty transport scenarios. Relying on core components such as the independently developed lightweight reducer, the robot is light but powerful, with a strength equivalent to three times that of a normal adult.
The robotics industry actually has a history of development of several decades. In the view of Cao Wei, partner of BlueRun Ventures, the Chinese market is a potential huge robotics market. Whether from the perspective of robot consumption or robot export, China is a huge pillar market in the global field. In addition, China has an aging population problem. In the future, robots are likely to be the key technical node to solve China's future aging population and labor supply. At the same time, in terms of talent supply and community ecology, China also has very good talent soil. More than 400 universities in China have set up robotics majors. In the past five years, more than 100,000 robotics companies have been registered, with an average of 10,000 to 20,000 companies each year.
In terms of capital, the total amount of investment in the robotics field by the entire investment community, whether in US dollars or RMB funds, has exceeded 100 billion in the past decade. Local governments and guidance funds are also actively setting up large industrial funds related to robots to support and promote the development of the robotics industry. Coupled with industrial clusters and long-term policy support, Cao Wei believes that in the next 5-10 years, or even 15-20 years, the robotics track will be an important and pillar-type basic innovation track.
In addition to industry logic, many industry insiders told reporters that the biggest factor affecting the popularity of the robot track this year is the upgrade and iteration of AI technology represented by large models. In the interview, Wang Xingxing said that the biggest factor affecting the current wave of robot development is the AI wave. At present, people believe more in AI and believe that humanoid robots can create more value, which was completely unimaginable ten years ago.
Cao Wei holds the same view. He said that the big model has enabled the industry to see a significant improvement in the connection and execution of robots in complex tasks. The success rate of traditional robot algorithms used to be around 50%, which is the primary level in the laboratory. However, with the support of the big model, the success rate of the same algorithm combined with the big model has increased by more than 50%, and some have even increased by 100%, which is gradually approaching the commercial level.
Be cautious about technology enthusiasm
Large model technology is not a panacea, not to mention that the technology itself is not yet fully mature.
Wang Xingxing told reporters that to truly unleash AI capabilities, physical robots must be able to actually work. “Working” is the industry’s primary expectation for robots, which also increases the industry’s imagination of robots. For this reason, Wang Xingxing believes that the biggest limitation of current technology on the robotics industry is that AI is “not enough” - AI models, AI training data sets, and AI scene deployment are far from enough.
Although the hardware is not fully mature, there is no theoretical threshold for hardware at present. The main problem is engineering, that is, to make the robot lower in cost and better in engineering, make the appearance more extreme, and make the hardware functions richer, but these problems are predictable in time. In comparison, the fact that robot AI technology has not yet made a breakthrough is more challenging.
According to Cao Wei's observations in the industry over the years, the robot has made great progress in both motion control and refined operations performed at the end. The algorithm has gradually moved from model base to learning base. This trend also allows the industry to see that the growth space of robots in the future can be combined with data. The more data, the better the learning performance.
But it should be noted that Cao Wei said that the robot is composed of thousands of parts, and the robot body is also a very important part. In the past 2-3 years, the preliminary hardware architecture of the robot humanoid has been established, but its key modules and technical paths are still being iterated and explored.
Different from the general belief in the industry that large language models are the core variable in the robot track, Gao Jiyang, founder of Xinghaitu, believes that embodied intelligence really solves the problem of the robot's ability to perform in the physical world. In the future, the key bottleneck that hinders the large-scale integration of intelligent robots into human society is the intelligent system, not the electromechanical system. The key factors are computing power, sensor systems, and the algorithm itself.
In the multi-technology path, the industry has increasingly stringent requirements for technical safety and stability. Juha Röning, vice president of the European Robotics Association, said that from the perspective of mechanical engineering, the field of robotics has achieved a high degree of standardization. But for modern systems and software architectures, the industry is still a long way from achieving the level of "plug and play". Compared with mechanical engineering, the field of computer science has a lower degree of standardization, and simply combining functional components together is far from enough. In the field of AI, this standardization is even more scarce.
Shigeki Sugano, president of the Robotics Society of Japan, said that the industry combines AI with hardware and other types of public information to create a hardware entity with intelligent capabilities, thereby achieving symbiotic interconnection between humans and machines. Since humans are always the ultimate service objects, safety and high power output become key issues that need to be solved urgently. Although humanoid robots are already on the market, most of these robots cannot effectively support human activities because their power output is insufficient. How to achieve high power output while ensuring safety is what Shigeki Sugano believes is the most core issue.
The survival direction behind the differences
It should be clear that the robotics industry has a history of several decades, and industrial technology is changing at the micro level every year. The general trend is moving towards automation.
In the view of Ren C. Luo, President of the IEEE Industrial Electronics Society (2000) and He Yi-ci Chair Professor at National Taiwan University, the robot 1.0 era two or three decades ago had some motors and controllers; the 2.0 era added visual inspection and became more intelligent; the 3.0 era saw the emergence of humanoid robots and collaborative robots; and now the industry has entered the 4.0 intelligent era.
Wang Xingxing believes that the current robotics industry has different ideas for each company. For example, what should be installed on the robot's camera? Where should it be installed? How many cameras should be installed? How should sensor data be collected? Should tactile sensors be installed? Tactile is a broad topic. Some schools do not want to use super tactile sensors or data sensors. Some schools do not even want to install dexterous hands, but only want to use claws instead. Some schools want more flexible hands, which means more fingers, and each finger should have enough sensors.
Therefore, in summary, Wang Xingxing said that everyone in the industry has different ideas, and the current technical route of the entire AI model is not so unified. This makes it difficult to determine which technical route is correct, which route is wrong, and which route has been advanced to which pace. These are all difficult problems that have no definitive answer.
If we compare it to a large model, Wang Xingxing said that after the emergence of the large language model, everyone has forgotten other models before GPT, but in fact, in the circle of large language models, before the emergence of ChatGPT, there were many language model structures in the industry. But after the GPT model proved the value of its own architecture, other architectures were also eliminated. The current humanoid robot and embodied intelligence track is somewhat like the one or two years before the emergence of ChatGPT. The industry is consciously working in a certain direction, but no one dares to guarantee that this direction is absolutely correct.
With the popularity of large models, hot money is pouring into the robotics industry. According to third-party data statistics, as of June 30, 2024, there have been 69 financings in the domestic robotics industry this year, 12 of which have disclosed financing amounts of 100 million yuan, and the total disclosed financing amount in the robotics field is about 7.5 billion yuan. Examples of high financing amounts include Yushu Technology's nearly 1 billion yuan B2 round financing, Galaxy General's 700 million yuan angel round financing, etc. In addition, there are star start-ups such as Zhiyuan Robotics, Xingdong Jiyuan, Xinghaitu, Pasini Perception, and Stardust Intelligence.
Jacqueline Du, a research analyst at Goldman Sachs and head of China industrial technology research, previously predicted that by 2035, the global humanoid robot market will reach $38 billion, more than six times the previous forecast of $6 billion.
Cao Wei told reporters that the technology in the field of embodied intelligence has greatly improved in the past one or two years. There must be bubbles, but good companies should "enjoy" the bubbles and create real value from them. Under the heat, value judgment should be based on the startup team itself: whether the team's ability in the process of technology implementation and productization is strong enough, and whether the team has thought about the real commercialization path. "If you have the ability to build 1,000 robots in a year and 10,000 machines in two years, this is doing business." Cao Wei said.
The implementation of robots is a gradual process. Cao Wei said that in the B-end market, enterprises' demand for robots is mainly focused on clear cost-effectiveness and stability. Enterprises hope that robots can complete simple but efficient tasks in specific scenarios. At the same time, the C-end market is also facing new challenges. With the popularization of consumer electronics, many standardized tasks have been replaced by mature products, leaving robot startups with basically complex tasks, such as household cleaning, caring for the elderly and other non-standard tasks. These tasks not only involve spatial interaction, but also require the collaboration of robots with other intelligent bodies.
In the early stage of the industry, Cao Wei's advice to start-ups is still to "stock up". Before the industry has formed mature commercialization, first find ways to raise more money and reserve funds. The second is to improve commercialization capabilities. "When the market is good, get more money. When you can't get money, think clearly about how to make money."
(This article comes from China Business Network)
