The Busier the Robotics Race Gets, the More We Need to Respect the Fundamentals | Li Feng Column
The Hype, Patterns, and Non-Consensus of Robotics Investing
The robotics market continues to heat up. On August 8, the World Robot Conference opened in Beijing. According to Xinhua News Agency, the conference sold a cumulative total of 19,000 robots and related products, generating over 200 million yuan in sales. Then in mid-August, the World Humanoid Robot Games kicked off, featuring events from the 100-meter sprint to high jump, 1,500-meter race, freestyle combat, and 5v5 soccer. The robots alternately "played dead" and charged ahead, giving us plenty of entertaining and inspiring moments.
Behind these viral moments lies a key data point worth noting: from 2015 to 2024, China was the world's largest industrial robot market for 12 consecutive years. In 2024, China's industrial robot sales reached 302,000 units. The hardware manufacturing supply chain, represented by industrial robots, is converging with new technologies like artificial intelligence to form a combined force that will advance the development of embodied intelligence robots.
Before this wave of embodied intelligence enthusiasm, FreeS Fund had already made systematic investments in the robotics sector. We have invested in nearly 10 companies related to embodied intelligent robots, including LimX Dynamics, Yuanluo Technology, Yifei Technology, Dongyi Technology, Yinshi Robot, and Hangkai Microelectronics, covering areas from body manufacturing and core components to sensors.
Recently, Feng Li, founding partner of FreeS Fund, shared his investment thinking on the robotics sector in the "TAIXUE" program under the New Economist Think Tank, addressing current hot topics and controversies surrounding robots, including:
- Why did we choose to invest in robots?
- Why are we optimistic about embodied intelligence today?
- Do robots have to enter factories, or could services be the bigger market?
- Are robots actually capable right now? What challenges do they face?
Here are some of his views:
We see enormous potential for robots in addressing future labor shortages and service economy demands. As China's service sector develops, labor supply shortages are becoming increasingly prominent. As a substitute or supplement to human labor, robots have broad market prospects.
Embodied intelligence involves not only complex hardware manufacturing supply chains but also combines advanced software and artificial intelligence. It could become a key tool for solving labor shortages in the service sector. For example, through automation and intelligent technologies, robots can provide efficient services in healthcare, dining, logistics, and other fields.
Although robots have made significant progress in certain areas, such as mobility and specific task manipulation, they still face numerous challenges in upper limb operations. These include the lack of accumulated physical world data and the absence of standardized sensors.
We hope this offers fresh perspectives. You can listen to this episode by searching for "High Energy" on the Xiaoyuzhou App and Apple Podcast, or watch the video version by searching for "Feng Li" and "Investment Logic for Robots" on video platforms, Bilibili, and Douyin.
Engagement Giveaway
What unexpected changes do you think robots will bring to our lives in the future? Share your thoughts in the comments. By 17:00 on August 21, 2025, the three most thoughtful commenters will receive a copy of Mind and Body: The Philosophical Challenge of Cognitive Science and Artificial Intelligence.
01
Why did we choose to invest in robots?
Today, from the venture capital industry to ordinary users, opinions on "robots" have largely split into two camps.
One view holds that robots have tremendous potential. Especially this year, "embodied intelligence" was included in the government work report as one of the "future industries" in national planning. The fact that it was written into a national strategic document indicates this is a promising, forward-looking direction.
The other view argues that robots are still not mature enough. For instance, at the world's first humanoid robot half-marathon held in Beijing in April, people saw some robots not only running slower than humans but also stumbling and falling, with some even needing two people holding them from behind to finish the race. Can robots in such a technical state really enter our lives and serve us?
Before 2021, the robotics field was far less hot than it is today. We chose to invest in robots based on several considerations:
The first is technological change. Although large language models were not yet as popular as they are now, some early AI technologies, particularly reinforcement learning, had already begun to profoundly impact robotics. Reinforcement learning enables robots to improve their behavior through trial and error, providing crucial support for robot intelligence.
The second consideration is China's unique advantages in the global industrial chain.
A robot consists of two core components: the software technology part, namely artificial intelligence, responsible for decision-making and control — essentially the robot's "brain"; and the hardware body part, including various precision mechanical structures, motors, sensors, chips, and so on, forming the robot's complex and highly integrated physical system.
In China, this combination of "software + complex hardware" like robots happens to be an area where we have significant advantages.
Over the past decade, China has proven that if we take a high-value-added technology and combine it with a long, complex hardware manufacturing supply chain to create an entirely new product form or category, once that product is widely accepted by society, it can become a notable Chinese advantage on the global stage.
In almost every industry following this "soft tech + hard manufacturing" model, China has demonstrated this capability. From smartphones to electric vehicles to action cameras, these products all exemplify the combination of software technology and complex hardware. For example, Huawei's phones now hold a place in the high-end phone market. Another example is Insta360's action cameras, which have surpassed GoPro to take a leading position in the global market.
Just as smartphones rely on the synergy between precision hardware manufacturing and powerful operating systems, robots similarly require deep integration of artificial intelligence with complex mechanical systems. China not only has a vast consumer market but also a complete hardware manufacturing supply chain. These combined advantages are gradually translating into competitive strengths for China in the global robotics market.
02
Why are we optimistic about embodied intelligence?
FreeS Fund has currently invested in nearly ten robotics-related companies. Why are we optimistic about embodied intelligence?
First, any major industry typically needs to go through a bubble phase
For an industry to develop, it may be difficult to avoid experiencing a bubble burst. During the peak of development, the industry attracts massive resource investment, driving technology and supply chain advancement; after the bubble bursts, the market eliminates some enterprises, leaving truly competitive "survivors" that lay the groundwork for the next round of industry development.
This up-and-down process may seem chaotic, but it is actually an inevitable path for industry development. In China, industries like the internet, e-commerce, and new energy vehicles have all experienced such cyclical fluctuations.
Specifically in the new energy vehicle industry, around 2014, China's "internet car manufacturing" concept emerged, with brands like NIO, Li Auto, and Xpeng Motors beginning to appear. Since then, the new energy vehicle industry has experienced many highs and lows, but China's new energy vehicle industry has still developed. China's new energy vehicle production and sales have ranked first globally for 10 consecutive years, with products exported to over 70 countries and regions.
Second, national strategy has enormous driving force for industry development
In China, sometimes an industry's development is not just due to commercial logic or market demand, but also because national strategy is pushing it forward.
Let me share a small investment lesson. When "new car makers" first emerged, I wondered whether it was too risky for people who had never built cars to lead an industry so heavily dependent on manufacturing capabilities. So we invested heavily in batteries, sensors, controllers, and other key components, but hesitated on investing in complete vehicle brands.
I summarized one lesson: under the push of national strategy, an industry's development speed, scale, and growth space often far exceed predictions based purely on market logic.
Looking at it today, what drove the rise of the new energy vehicle industry was far more than market preference for new products. The fundamental driving force lay in China's urgent need to address energy structure transformation.
Starting from 2017, China surpassed the United States to become the world's largest oil importer. Currently, over 70% of China's annual oil consumption relies on imports. This external dependence means that in extreme situations, energy security would be at risk.
Developing new energy vehicles became one solution to this energy dependence.
Third, what opportunities exist in the embodied intelligence field?
Referencing the new energy vehicle industry, what kind of opportunities are we actually encountering in the robotics field?
1. Replacement demand driven by service sector labor shortages
Many policies issued by the state in recent years have pointed toward the service sector. These policies cover not only traditional service industries like dining, tourism, and cultural entertainment, but also multiple sub-sectors including finance, education, and healthcare. For example, the 2023 Guiding Opinions on Accelerating Digital Empowerment of Life Services, the Implementation Plan for Full-Chain Support of Innovative Drug Development that began implementation in 2024, and the 2025 Implementation Plan for Loan Interest Subsidy Policies for Service Sector Business Entities, among others.
International experience shows that when a country's per capita GDP exceeds $10,000, it typically enters a stage of rapid service sector development. In 2024, China's per capita GDP exceeded $13,000, with the service sector accounting for 56.7% of GDP. By comparison, the U.S. service sector accounts for about 80% of GDP. This means China's service sector still has considerable room for growth.
But problems follow: the service sector is extremely labor-dependent. With changing population structure, particularly accelerating aging, we may face a serious contradiction in the future: more and more people wanting to buy services, but fewer and fewer people able to provide them.
How do we solve this?
As mentioned above, robots are a combination of "soft technology" (such as artificial intelligence) and "complex hardware manufacturing supply chains." This "complex hardware" refers not just to shells and structural parts, but also includes numerous core components like motors, chips, and sensors.
When we integrate these technologies and capabilities to address the structural contradictions faced during service sector development, this could mean an entirely new industrial development opportunity for China.
Let's understand innovation opportunities in the service sector driven by technology, using e-commerce and ride-hailing as examples.
First, look at e-commerce. Around 2000, China entered a stage of rapid growth in commodity economy. During this period, the real estate market developed explosively, similar to the U.S. post-WWII situation. Rapid real estate growth not only drove housing demand but also stimulated related consumption needs, such as renovation, home appliances, and furniture purchases.
However, China's offline retail system was not yet mature enough to fully meet consumers' growing demands. In 2003, Taobao launched, giving birth to a business model fundamentally different from abroad — e-commerce. E-commerce used digitalization plus technology to solve the supply-demand contradiction at that time. China's e-commerce outpaced offline retail.
Now look at ride-hailing. Although premium car services originated with Uber in the U.S., the ride-hailing experience enjoyed by Chinese users may be the best in the world — abundant vehicle supply, good vehicle configuration, and plentiful additional services like drinking water and charging cables.
Before ride-hailing emerged, the enormous challenge facing China's transportation market was that demand for mobility services far exceeded the supply and dispatch capacity of the taxi system. So we used digitalization plus technology to vigorously develop the ride-hailing model.
These two examples illustrate that when economic development reaches a certain stage, we can try using technological means to solve supply-demand imbalances — "structural contradictions." This is also our main reason for investing in robots: using technology to address potential contradictions that the future service sector may face.
2. Why do robots need to be humanoid?
Whether robots should be humanoid is one of the hot topics in the industry today.
The reason robots need to be humanoid is because they are entering the service sector, appearing in daily life, and interacting with you and me.
Unlike manufacturing, the service sector requires robots capable of interacting with people, moving freely in human daily life spaces like homes, shopping malls, and hospitals, and completing tasks. If robots were not humanoid, we would have to redesign all infrastructure, including theater seats, corridor widths, table heights, and so on. Otherwise, robots could not adapt to existing environments.
In other words, rather than having society change existing infrastructure for robots, it's better to have robots adapt to human spaces as they currently exist.
Returning to the original question: why are we still optimistic about robots, still optimistic about embodied intelligence?
First, any industry may experience bubbles and bubble bursts — this is an inevitable path for industry development.
Second, the underlying logic behind robots is not just technology itself, but also a series of macro factors including national development strategy, economic structure adjustment, and demographic changes.
Finally, what embodied intelligence faces is a huge, long-term structural opportunity — the severe future shortage of labor supply in the service sector. And robots are precisely the key tool for solving this problem.
Are robots actually capable right now?
What challenges do they face?
First, robot upper limbs are far from mature
If we divide robots into upper and lower halves, the upper limbs are responsible for manipulation, while the lower limbs are responsible for mobility.
Currently, robot lower limbs (mobility capability) have made significant progress. Thanks to advances in AI algorithms like reinforcement learning, robots' mobility has become quite strong — they can walk and run on complex terrain, and even get back up after falling.
However, robot upper limbs (manipulation capability) are far from mature. Upper limb manipulation involves complex physical interaction, which is far more complicated than simple locomotion.
Although many flashy demo videos show robots folding clothes, chopping vegetables, or pouring coffee, these tasks mostly involve manipulation of fixed objects. What is truly difficult is having robots interact with people or other dynamic objects, such as giving massages, cutting hair, or dressing people.
We can use autonomous driving as an analogy to understand where the difficulty lies in robot upper limbs.
Current autonomous driving technology is mostly at L2 or L3 level, still requiring driver monitoring. Today's robots can already set their own routes to climb mountains, run, and get back up after falling — equivalent to achieving L4 level.
The goal of autonomous driving is relatively clear: from point A to point B, without colliding with anything. By comparison, robot manipulation requires considering not avoiding contact with everything, but touching "the right things" — involving far more variables and uncertainties, such as hand-environment, hand-object, object-environment, and upper-lower limb coordination.
For example, simply pouring water requires the robot to consider how to pour water from a kettle, how to recognize a cup changing from empty to full, and how to grip the cup tightly enough to not drop it on the floor.
Second, robots lack "common sense" data about the physical world
The robotics industry currently faces another huge challenge — lack of accumulated data about the physical world.
Let's look at a comparison. As early as the internal combustion engine era, the automotive industry attempted autonomous driving. For instance, in 1958, General Electric conducted unmanned acceleration and braking tests on a highway. In 2009, Google launched its self-driving project, later spun off as the independent company Waymo. In 2020, Tesla released the FSD beta, and autonomous driving gradually entered a stage of technical攻坚.
These breakthroughs mean that even under relatively simple control modes (like steering wheel, throttle, and brake), the automotive industry had already accumulated decades of data. Moreover, this data could be continuously reused to lay the foundation for L4/L5 autonomous driving. Additionally, nearly every new energy vehicle today is equipped with numerous sensors, and the data they collect can still be used to train high-level intelligent driving systems.
Robots, however, face a highly complex physical interaction environment. For example, how much water is in a bottle? Is this bottle soft or hard? Is it a paper cup or metal cup? Is it cool or hot? Can it be touched? What's the temperature? Can it be tilted? Will it spill?
This information about object material, weight, state, and force feedback constitutes the "physical laws" of the real world. But historically, such data has hardly been systematically collected, or only a small portion is available.
This also explains why current video large models often exhibit "clipping" phenomena — such as arms passing through walls or movements violating gravity. This is because when generating video, the model only matches based on text and existing video content, without truly understanding physical laws.
Humans are different. Through bumps, falls, and trial and error during growth, we gradually build an intuitive system about the physical world. For example, if you see a stool that looks like it's about to fall apart, you definitely won't sit on it. This judgment isn't innate but the result of accumulated experience.
But robots don't have this "life experience." They cannot accurately judge how much force to apply or at what angle to grasp or manipulate an object. Therefore, we rarely see demos of robots doing complex interactions with people — because the risk is too high. It might accidentally injure you.
Some have suggested this problem might be solvable through large models. With massive video data from the internet, plus some training and virtually generated data, robots might be able to "learn" this physical common sense.
Large models may help robots learn about the physical world, but relying solely on them is far from enough. Similar to humans, robots also need to go through processes of "falling, colliding, and trial and error" to establish true physical intuition. This process cannot rely solely on virtual data.
Third, in areas like sensors, robots have not yet formed standardized technical approaches
Beyond the lack of physical common sense, robots face another major technical challenge — sensor configurations and technical approaches have not yet become standardized.
We can use the development process of new energy vehicles as an analogy to understand this.
For intelligent driving, basically only electrified architectures (like motors, electronic control, and batteries) can achieve true digital control. Instructions sent by chips can drive motors in real-time with precision, achieving "equal speed, equal effect, equal time" control methods. Internal combustion engines are completely different — they need to inject fuel first, then combust, and finally convert to power. The entire process has high latency and significant energy loss, with control that is neither real-time nor precise.
When the new energy vehicle industry first began exploring electrification, its sensor system was already gradually becoming fixed. For example, the number and position of cameras, whether to use LiDAR, and so on.
Although there are slight differences between manufacturers, overall, sensor combinations have basically stabilized, and training data has continuity and reuse value.
Robot technical solutions currently are far from reaching this level of standardization.
Robots must face a highly dynamic, complex world. They need to interact with the world, requiring extremely rich perception and judgment capabilities, which means they need to collect large amounts of multi-dimensional data.
The problem is: what sensors should we install on robots today? How many? How should they be installed?
There is no standard answer to this question yet. Different research teams and companies may adopt completely different sensor combinations, precision levels, and installation positions. Once these configurations change, previously accumulated training data may become unusable.
To give an exaggerated example. For a robot's hand, is two fingers better, or three? Or five? If you initially trained your model with two fingers and suddenly decide to switch to three, all previous training data may be invalidated, and you have to start over. If someday you want to try five fingers, you have to do it all over again.
These challenges show that robot technology development has a long road ahead — this is what an emerging industry always goes through.
Looking back at the development history of new energy vehicles, investors in new energy vehicles probably didn't anticipate such enormous changes occurring over just 10-plus years. In the end, those who dared to invest received rich returns. They experienced multiple cycles of bubbles and bursts, waiting for the market to finally explode.
Currently, although robot technology faces many challenges, it is indeed a future industry.
With our limited imagination today, we may not be able to envision what the future will look like 10-plus years from now — how China will resolve contradictions in economic development through technological innovation: developing the service economy while addressing labor shortages, while also hoping service prices become more affordable.
Going back to 2012, when DiDi's predecessor Xiaoju Technology was just founded, who could have imagined it would change our way of getting around? Similarly, future robots may, like today's BYD and CATL, bring unexpected changes to our lives.
Engagement Giveaway
What unexpected changes do you think robots will bring to our lives in the future? Share your thoughts in the comments. By 17:00 on August 21, 2025, the three most thoughtful commenters will receive a copy of Mind and Body: The Philosophical Challenge of Cognitive Science and Artificial Intelligence.
Will 2025 Be the "ChatGPT Moment" for Embodied Intelligence? | FreeS Fund Research
The Road to Embodied Intelligence | FreeS Fund Report 37
