Humanoid Robot Industry: Observations and Reflections
The 7th World Artificial Intelligence Conference (WAIC)

Introduction
Recently, Shanghai hosted the 7th World Artificial Intelligence Conference (WAIC). This year's event was particularly high-profile, with Premier Li Qiang in attendance. Many humanoid robot companies participated, and a special humanoid robot display was set up featuring eighteen humanoid robots, dubbed the "Eighteen Arhats".

We are based in Shenzhen, and for the month leading up to the exhibition, our colleagues were on business trips in Shanghai, helping our clients prepare for the event. Of the 18 companies in this display, 5 used our control systems. Thank you to every client for your trust! Although our own booth was quite modest, enabling our clients' success is our greatest achievement.
Scarcity of Motion Control Talent in the Humanoid Robot Industry
On the one hand, with the surge of embodied AI and humanoid robots, more and more companies and teams are entering the legged robotics industry, but the growth of talent lags far behind the growth in market demand.
On the other hand, legged robot motion control technology has entered an entirely new phase since 2023. After decades of ZMP and several years of MPC, the new paradigm is reinforcement learning.
In the legged motion control industry, talent has always been extremely scarce, and now, with the dual factors of paradigm shift and explosive market demand, both leading research institutions and leading robotics companies find it difficult to hire talent.
Anyone who has tuned a new physical robot from scratch knows that debugging is an extremely painful process, especially without prior success. Only those with immense mental fortitude can navigate the fog, guided by a faint glimmer of light, to find the right direction.
A friend from Tsinghua University summed it up well: "Most Tsinghua students aren't suited to tuning robots because their lives have been too smooth so far."
Our Solution
Since late 2023, with our strong R&D capabilities, we rapidly completed the paradigm shift from MPC to reinforcement learning, and through extensive experience with numerous clients, we invented an efficient system adaptation process. For mature hardware, algorithm adaptation can be completed in as fast as three and a half days. To date, we have accumulated hardware debugging experience with over a dozen legged robots of various sizes, a team with this level of experience likely has no equal.
According to current market conditions, hiring an experienced and capable RL motion control lead would cost at least 2 million yuan per year to start, and with supporting staff and backup (10+ people needed for competitiveness), the annual investment would be close to 10 million yuan; even a less robust setup would cost 5 million.
However, a company's robot models are limited, and after the algorithm team tunes one model, they may be idle most of the time, leading to slower experience accumulation and growth, which is also a waste of talent.
Rather than investing energy in building your own team, it's better to choose a third-party team like ours. We can deliver complete training and deployment solutions plus source code, so you needn't worry about technology being controlled by others.
This model places greater demands on our team's innovation speed, otherwise we would have nothing to sell. But from another perspective, this is good for us, as the pressure forces us to innovate at the fastest possible pace, rather than resting on past laurels and gradually losing vitality.
Historical Patterns
The reason there are currently almost no other teams in the domestic market specializing in motion control algorithms like us is that teams with our technical capabilities face too many short-term temptations—they have ample opportunities to join a robot company or start one for short-term excess returns. Our path, however, is exceedingly difficult, requiring us to face many doubts and challenges along the way, with little short-term gain to expect. So why do we believe this path can succeed?
The personal computer revolution in the United States in the 1980s, as depicted in the book "Fire in the Valley," bears a striking resemblance to the changes happening in China's robotics industry today, though no one has yet called it a "revolution." Years from now, looking back, people will think: the sparks have just ignited. From the history of the personal computer, we can learn profound industrial patterns.
In a fiercely competitive, vast market, it is not impossible for a startup to achieve first-class standards in every aspect, but it is indeed extremely difficult. In the personal computer industry, Apple represents vertical integration, while Microsoft represents horizontal integration. Even mighty Apple saw its market share fall from a peak of 16% in the late 1980s to 4% by 1996. In the PC era, Apple lost to Microsoft.
Many say Microsoft is unrepeatable, but industrial patterns are hard to break. One of the founders of modern economics, Adam Smith, pointed out in his 1776 work "The Wealth of Nations" that specialization of labor brings efficiency gains, especially in sufficiently free and competitive broad markets. The automobile, personal computer, and mobile phone industries of the past were all such markets—fully competitive and vast.
Intel CEO Andy Grove, drawing on his own experiences in "Only the Paranoid Survive," also emphasized the importance of industrial patterns. If you cannot adapt, you face elimination; even the blue giant IBM was powerless.

The quest for general-purpose robots is not a short-term endeavor. Building a decent company may take less than ten years, but building a great company takes decades. We have enough time and patience to think and act with a long-term perspective.
Technical Outlook
We firmly believe in the efficiency of industrial specialization, a confidence rooted not only in historical industrial patterns but also in a profound technical principle. This principle also governs the development of today's artificial intelligence technology.
Computer scientist Richard Sutton, in his essay "The Bitter Lesson," pointed out that AI researchers often inject human knowledge into algorithms, which helps in the short term but in the long run, large-scale "search and learning" is what leads to breakthroughs. This pattern has been fully validated in CV and NLP: first from rule-based methods to small-scale learning algorithms, then to large-scale learning general algorithms. In future AI, the proportion of human knowledge in systems will decrease, while capabilities and generality will increase.
Today's legged motion control technology has reached the second stage. Currently, training still requires injecting human knowledge such as robot properties represented by URDF, and the models used are relatively small, mostly targeting a single robot model. This is like the past in NLP, where models were trained for only one language, whereas today's general-purpose language models can work well across languages. Future general-purpose controllers will also achieve cross-robot morphology, and this is what will happen in the third stage, a future that we believe is not far off.
Once such a general-purpose controller is developed, it will have a huge impact on the humanoid robot industry. The barrier to assembling a humanoid robot will be greatly lowered, and robot capabilities will be significantly enhanced. At that point, more participants will enter the industry, and control systems will no longer be the core focus. Teams that can better integrate resources, define products, and handle sales will have the advantage, while the R&D technologies accumulated by traditional robotics companies will be replaced, even becoming worthless.
And we are the most eager pursuers of such a general-purpose controller, and the team best positioned to develop it. We have encountered the greatest variety of hardware configurations, and control system R&D cannot be divorced from hardware. No single robot company can easily develop controllers for other companies' hardware; without a sufficient variety of hardware configurations, there is neither the conditions nor the motivation to develop general-purpose controllers.
Therefore, from the perspectives of both capability and willingness, we are the team most likely to develop a general-purpose controller. If you also believe that general-purpose controllers will become a reality, please contact us (WeChat: bridgedp) to invest or join us.





界面新闻
深视新闻
深圳特区报
财新
36氪
钛媒体
中国新闻周刊
蓝鲸新闻
Z Finance
投资家网
雷峰网
亿欧
环球网
南方日报
极客公园



