“Staggering imagination.”
Text｜Liu Yang
Editor｜Ba Rui
In the first half of 2025, Gao Fei, the CEO of Weifen Zhifei, stood at the entrance of a mine, feeling unusually nervous. This was his first time delivering a product to a client — an autonomously flying robot.
It had to fly into an underground cavity that was60 meters high, equivalent to20 floors, without GPS, without communication signals, and without light. The mine tunnels crisscross like a maze. It could only rely on the sensors and chips on its body to map and locate its path as it flew, completing the surveying task before returning.
Ten minutes later, it brought back a map of the unfamiliar mine and found a previously undiscovered shortcut. Because it had to take a shortcut, its algorithm required it to complete the task in the shortest time possible.
The clients were veryamazing, as they had been working in this location for several years, thinking there were only two routes and completely unaware of a passage between the two mines.
From this point, the flying intelligence team in the lab entered the commercial world — their first smart flying robot was resold for several hundred thousand yuan each.
What kind of drone can sell for so much money?
Strictly speaking, it should not be called a drone. Its true definition is “flying robot” — the sensors serve as its eyes, the onboard chip as its brain, and its four propellers as its limbs. However, it looks almost identical to a regular quadcopter.
The difference lies not in appearance, but in who is flying it.
The underlying technology of the flying robot is fully autonomous flying, meaning it relies entirely on itself. There is no need for manual intervention, as it completes tasks relying on sensors, chips, and algorithms. Its greatest distinction from other forms of embodied intelligent robots is that it moves by flying — “the fastest runner” that can go anywhere, technically referred to as strong passability.
Thus, the flying robot costing several hundred thousand certainly isn’t designed for serving coffee to humans. It does not replace human capabilities, but rather complements them. It will take the place of humans in dangerous, complex, or inaccessible environments, thereby opening up more possibilities for imagination.
Hard technology often exhibits such romance — both technology and business stem from boundless imagination. Imagination is also the leverage of business, and it is expensive — it can set prices, elevate valuations, and open market space.
Weifen Zhifei recently obtained hundreds of millions of yuan in Series A2 financing, with a valuation of several billion yuan. Led by PwC Capital, followed by Honghui Fund, Yangtze River Delta Intelligent Culture Fund, and with additional investments from existing shareholders such as Wuyuan Capital, Shenzhen Capital Group, Hongtai Fund, Huaying Capital, Huakong Fund, Changshi Capital, and BV Baidu Ventures.
This magnitude is clearly not supported solely by aerial photography, selfies, or industrial drones, but rather by the valuation logic of “flying intelligence.” The “group brain” is the most imaginative part of this valuation logic: multiple drones forming a cluster in real-time without GPS or pre-arranged tactics, each one thinking independently and avoiding obstacles, working together like a tactical squad to autonomously complete tasks.
Gao Fei's goal is to create intelligence akin to the drone swarm depicted in “The Wandering Earth.”
Born in 1993, Gao Fei became a lecturer at Zhejiang University at the age of 26, became a doctoral advisor at 28, and attained the position of tenured associate professor (lifetime position) and national outstanding youth at 30, making him one of the few experts in the field at a young age — he has won the annual best paper award from the prestigious robotics journalIEEE T-RO, the first time a university in mainland China has achieved this as a leading party; the lab's cluster flying also appeared on the cover ofScience Robotics.
The founder has sufficiently impressive achievements in scientific research, which attracts hot money; the flying robot is a new category in the narrative of “AI entering the physical world,” which also appeals to investors.
However, the allure of easy money is not the reason for Gao Fei's involvement. Investors have been urging him to start a business since 2022, but he hesitated until his research results matured further, wanting to see what this technology could truly bring to the world.
Entrepreneurship was a decision that took a long time to develop, and exploring what kind of entrepreneur to become was also a long journey.
In the process, he clearly sensed the differences in judgment criteria between the two: in research, the standard for deciding whether to do something is its difficulty, while in business, it depends on whether it generates real value and gets customers to pay. He regards “entrepreneurship” as a skill to be acquired, rather than an identity.
But the challenge is how a young and famous professor becomes an entrepreneur and builds a successful company.
Below is the dialogue between“Undercurrent Waves” and Gao Fei (edited):
Part01
Generalization ability with zero samples
Undercurrent: You mentioned that the first to pay real money was the mine, just in the first half of 2025, shortly after starting your business. How was this transaction facilitated?
Gao Fei: A mining-related organization contacted me, stating that there was a well-known problem in the industry that could not be solved — there are many voids in the mine where humans cannot go down, making surveying impossible, especially in old abandoned mines, where safety risks are even greater. The only solution available at that time was a foreign drone that towed a fiber optic cable, which required a person to remote control it from outside the cave, making it impractical. They saw my videos online and realized that the drone could fly autonomously, so they approached me for a customized solution.
We customized a unit to sell to them, and later we found out we sold it too cheaply. Similar foreign products were selling for over 1 million yuan.
Undercurrent: Why was it able to sell for so much money?
Gao Fei: It depends on what problem it solves. Underground mines lack GPS signals, communication can easily be obstructed, and humans cannot go down.
The cost of traditional surveying methods in mines is very high and also very dangerous. Traditional surveying requires transporting high-precision equipment into the mine, but many places are inaccessible; for a cliff with dozens of meters of drop, people have to be tied with a rope and lowered; or drilling from the outside into the mountain, only to find after drilling for 5 meters that the thickness is5 meters. Estimating the size of the mine by drilling many holes is extremely primitive.
Flying robots are different; people give tasks from outside the cave and it explores the entire environment on its own, bringing back data. For mining companies, this is not reducing costs, but increasing efficiency.
This product has already been steadily delivered to large mining companies. In one location, we completed in six months what had previously taken over two years, covering more than 300 mining areas, saving over 10 million yuan.
Undercurrent: What other application scenarios are there?
Gao Fei: Forestry, power inspection, and emergency rescue. In forestry, previously several people would team up to walk into the woods, measuring the trunk diameter of each tree to estimate carbon content; for electricity, in indoor substations and distribution rooms, manual inspections are inefficient and robotic dogs have limited views.
These scenarios cannot all be addressed simultaneously; the choice of which to pursue first depends on two factors: whether the technology can achieve it and whether the demand is too niche or costs cannot be recovered. Mining may not be the most profitable, but it has the highest difficulty and data value; each task is conducted in a completely unknown new scenario, requiring our flying robots to genuinely possess zero-shot generalization capability. The data and models that run smoothly here can be dimensionalized to other scenarios more easily. The mining product could even be delivered for emergency rescue without modifications.
Undercurrent: Which demands are currently difficult to realize?
Gao Fei: There are some requests from firefighting, such as rescue after a fire, but high temperatures may cause component issues, plus smoke can affect sensors. Also, there are requests from tourist areas to catch thieves, but if it falls and hits a visitor, that is definitely unacceptable. Software capability has always been our advantage, but many times we cannot meet demands due to specific requirements of scenarios on the hardware itself.
Undercurrent: Embodied intelligence lacks data on human behavior; what data does the flying robot need to collect?
Gao Fei: There is a large iron ore mine that is a major client, and we stationed a flying team there. We deliver flying robots and daily operational data, and they opened up all these areas for me, allowing me to collect data anytime. We have already delivered dozens of devices, flying inside daily.
After data desensitization, they provide this data to me for free, and I also offer them very favorable prices. We have recorded over ten thousand instances of real flying data here. Once this is settled, transferring it to more structured and simpler scenarios will enhance the model’s generalization capability.
Undercurrent: What is the current delivery capability?
Gao Fei: We are in small batch deliveries, with several hundred orders on hand.
Part02
Higher-level intelligence
Undercurrent: Why do you say “no coordinates” is a higher level of intelligence?
Gao Fei: Because often there are no coordinates to provide at all. Providing coordinate points is itself a false proposition; it is hard to know where the task goal truly is. The GPS latitude and longitude coordinates must be in open outdoor locations with signals. Take a mining accident site; the original mapping data has long expired. Completing the task without coordinates is what constitutes advanced intelligence.
Undercurrent: How do you assess if the technology of fully autonomous flight is good enough?
Gao Fei: First, the data collected is used for modeling, so the modeling must be accurate with minimal error; we generally require within1 centimeter. Second, the efficiency of flight must be high; it cannot retrace its path in a maze. It must intelligently know how to return and find the way to reach its destination.
Undercurrent: What core capabilities are needed to accomplish these tasks well?
Gao Fei: Essentially, it requires a foundational model for flying intelligence, which includes agile small brains, a decision-making large brain, and collaborative group brains, as well as one for flying operations. Moreover, in scenarios requiring multi-machine collaboration, there is also a group brain model, enabling distributed real-time decision-making and autonomous operation.
The small brain is responsible for end-to-end control of the flying body, adjusting posture quickly like a bird and navigating through narrow gaps; the large brain understands commands, breaks down tasks into sub-tasks and decisions, and calls upon tools like maps and obstacle avoidance; the group brain corresponds to multi-machine collaboration, suitable for high-mobility searches in special operations and emergency rescue, with our largest distributed cluster capable of supporting1000 machines.
The most core aspect is to enable the flying machines to generate generalized intelligence aimed at their flight tasks.
Undercurrent: Is 1000 machines tested in what scenario? Is it calculated individually?
Gao Fei: No, it is a distributed cluster computed in real-time; every little black dot seen in the visuals is a drone making real-time decisions. When they fly towards the center, they may group up, but will not collide, which is quite challenging. This is the highest configuration in the industry.
The most representative real machines are the ten in Anji Bamboo Sea; each has only one camera plus onboard inertial navigation and simple inter-machine networking, which are very cheap nodes, without GPS, no manual control, and no pre-mapping.
Undercurrent: What is the hardest part of the group brain?
Gao Fei: The so-called group brain is about the tacit understanding of the group; each one can think and knows the limits of one another's abilities, just like tactical squads. It must know how to cooperate with others but cannot do so rigidly; if there’s an obstacle ahead, it must go around and cannot be foolish enough to collide.
Secondly, tasks must be self-assigned. In search and rescue scenarios, in many locations without even mobile signals, the cluster will automatically group, with the smaller ones squeezing through gaps, the larger ones carrying communication stations as relays, and the faster flying ones covering areas first to locate individuals; this is called high mobility search. Who handles which tasks is not predetermined but discussed among them.
Third, computational power is necessary; all of this must be calculated in real-time onboard, with models kept within10B.
Undercurrent: Is the group brain a real demand? What fields can it extend to in the future?
Gao Fei: Currently, the actual use of flying machines in various industries is quite limited, apart from aerial photography and marginally in pesticide spraying. This is primarily because remote control proves insufficient; numerous scenarios require intelligence, autonomy, and even must rely on clusters to complete tasks.
The most direct application now is in emergency rescue; we have already collaborated with many related domestic units, using our models for cluster collaboration in their settings. Beyond this, multiple machines can undertake joint lifting or tracking, and this collaboration is not limited to flying machines as quadrupeds and humanoids can also participate.
Have you seen the swarm in “The Wandering Earth”? That is also our future goal.
Part03
To run a company, you have to make money
Undercurrent: In 2022, it seems like you had a breakout moment on station B? Did investors start approaching you around that time?
Gao Fei: At the beginning of 2022, I compiled and posted the research成果成果 of the past year from my research group on station B. The video showcases drones flying autonomously and avoiding obstacles in unfamiliar environments like underground parking lots and forests, along with demonstrations of amphibious capabilities and wind resistance. There was no scientific explanation or packaging, and the views quickly surpassed 1 million.
Later, we let 10 drones autonomously fly through a bamboo forest without GPS, without manual control, and without pre-mapped layouts. This achievement in cluster flight made it to the cover of Science Robotics, widely reported by the media, marking a true official breakout from the academic circle.
Investors started contacting me relentlessly from that point, persuading me to start a business.
Undercurrent: Why didn’t you act at that time? Why did it take until 2024 to establish the company?
Gao Fei: I wasn't ready at that time. I had been focused on research, being a professor, and had never considered how to make products or run a company. By mid-2023, I had secured a tenure position at the age of 30, which freed up my energy significantly. I didn’t want to remain in that state, so I began to seriously contemplate industrialization.
From the second half of 2023 to the first half of 2024, I talked with many friends in the industry and spent time learning how to manage a company. By July 2024, I felt pretty prepared, so I established the company, and after the Spring Festival of 2025, I gradually made a few deals.
Undercurrent: What did you discuss with your friends in the industry? Was there any particular conversation that had a significant impact on you?
Gao Fei: Mainly learning about entrepreneurial methodologies. One conversation that left a deep impression happened after the Spring Festival of 2025, when I spoke with a senior entrepreneur for an hour and a half. I suddenly realized that from the entrepreneur's perspective, the understanding of the essence of societal operation differs from that of a scientist; it’s broader, more realistic, and more mature.
Undercurrent: What did he say that touched you?
Gao Fei: It wasn’t any specific statement, but rather his general demeanor. During our conversation, I comprehended how a mature entrepreneur communicates with others, assesses business models, and understands user needs.
For him, I was extremely naive commercially, yet highly adept technically. He maintained a very calm and objective demeanor throughout the process. He kept me updated regarding drone technology.
Undercurrent: Do you aspire to become an entrepreneur like him?
Gao Fei: Yes, in some ways, he is the person I aspire to be. He made me feel the differences in cognitive dimensions. That deeper understanding, wisdom, and higher-dimensional perspective on things.
Undercurrent: What do you think is the difference between conducting research and engaging in business?
Gao Fei: In research, evaluating a project largely revolves around its difficulty; however, in business, a more practical, harsh, and challenging assessment is whether what you produce has actual market demand.
Thus, during that stage, I understood why many people say scientists live in ivory towers, because the vast majority can't judge the righteousness or wrongness of a scientist’s work. But running a company means making money and maintaining a positive cycle.
At that moment, I developed a profound respect for entrepreneurship. I suddenly realized that despite my strengths in research, if placed in society, my ability to earn might not be stronger than that of a woman operating a breakfast shop. Because I have never completed a product in such a realistic social environment.
Undercurrent: Have you resolved that issue now?
Gao Fei: I'm working hard on it.
Undercurrent: What is your ultimate expectation for the company? An IPO?
Gao Fei: I hope this endeavor lasts long enough to allow for a lifelong career. Research results should either be placed in textbooks or on shelves.
Undercurrent: Which is more difficult?
Gao Fei: Both are difficult, but they pose challenges on different dimensions. Theorems for textbooks are harder as they require deep exploration of a problem; however, products on shelves have greater complexity, with various aspects intertwined, and must not have any weaknesses.
Undercurrent: Which one do you ideally want?
Gao Fei: I want both. But for textbooks, it may take another ten years or so.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
