When Robot Brains Take to the Skies: Taking on Dangerous Jobs in Our Place—A Conversation with Xiyu Technology
2026-08-22 16:53:13
Source: QbitAI
Joe Buchi, reporting from Aofeisi
QbitAI | WeChat Official Account: QbitAI
The embodied intelligence startup scene can be described as “blazing hot.”
New startups and innovative technologies are emerging almost every day. The number of approaches and players keeps growing, while financing, models, data, hardware, and more have all reached fever pitch.
Amid all this excitement, another professor with an intriguing background has entered the arena.
His name is Zhang Fu.
Each of the labels attached to his name could tell a story on its own: a student of Li Zexiang, an eight-year industry adviser to DJI, and an associate professor at the University of Hong Kong.
While studying drones under Li Zexiang, he learned to stay grounded and communicate directly with supply-chain teams and users—not just act as a scholar or observer. During his eight years at DJI, he learned how to turn laboratory technologies into products and gained firsthand insight into the real challenges of industrialization. At the University of Hong Kong, he and his team continued advancing cutting-edge research into autonomous flight and related fields.
So when Zhang Fu launched his company, he did not join the crowd competing over the “ground.”
Instead, he combined these three experiences to build aerial intelligent agents that can perceive their surroundings, determine their own positions, plan their next actions, and ultimately complete missions autonomously—without relying on GPS positioning or remote control by a pilot.
Image
This also creates another possibility for embodied intelligence to enter the human world:
Embodied intelligence does not necessarily have to grow human hands and feet.
It can first grow a pair of wings, entering on our behalf the dozens-of-meters-high curtain walls, steep hillsides, and hazardous industrial sites that are difficult or dangerous for humans to reach.
Another Professor-Turned-Entrepreneur Enters the Embodied Intelligence Arena
Xiyu Technology has attracted attention not only because it has brought embodied intelligence into the sky, but also because of its founder, Zhang Fu.
Zhang Fu is a student of Hong Kong’s “godfather of entrepreneurship,” Li Zexiang. He earned his undergraduate degree from the University of Science and Technology of China and later pursued a PhD at the University of California, Berkeley.
Around 2015, Zhang Fu joined Li Zexiang at the Hong Kong University of Science and Technology, where he began systematic research into robots and drones.
Starting in 2016, he also became deeply involved in DJI’s core technology R&D and productization efforts, working on flight control, multi-sensor fusion, LiDAR, and other areas.
After years of working on the front lines of industry, he participated in the full process spanning R&D, product definition, and mass production.
Image
Zhang Fu later joined the University of Hong Kong as a tenured associate professor in the Department of Mechanical Engineering. During this period, he founded the MaRS Lab (the HKU Mechatronics and Robotic Systems Laboratory) and continued to focus his research on one goal: enabling flying robots to find their way autonomously in unfamiliar environments.
According to publicly available information, MaRS Lab has long researched UAVs, autonomous navigation, LiDAR SLAM, and swarm coordination, while emphasizing the transfer of laboratory成果 into the real world.
One representative achievement is the FAST-LIVO series.
Put simply, after GPS fails, FAST-LIVO2 enables a robot to “find its way” on its own: observing its surroundings, determining where it is, and piecing together the world it sees along the way into a map.
A paper related to this work received the IEEE Transactions on Robotics (TRO) Jing-Sun Fu Memorial Best Paper Award.
This technical approach has become the foundation of Xiyu’s current “aerial intelligence”: first enabling machines to know where they are, then helping them understand what surrounds them, and finally allowing them to determine “what I need to do and how to do it next.”
Zhang Fu is now leading his team in commercializing the laboratory’s research technologies through the company’s products.
The company has built a technology stack encompassing multimodal perception, end-to-end motion control, world navigation models, dexterous manipulation, and swarm coordination. Its goal is to free aerial vehicles from dependence on GPS, prior maps, and continuous human control.
So what Xiyu really wants to sell is not merely smarter drones, but a general-purpose brain for drones.
The aim is to give drones autonomous capabilities so they no longer require real-time manual operation, while also enabling the same intelligence to be deployed across different aerial vehicles, allowing them to perceive, make decisions, and complete tasks autonomously.
Capital has already begun placing its bets. Xiyu completed four rounds of financing within six months, raising several hundred million yuan in total. Investors include Yao Capital, Jinqiu Fund, Alibaba, and others.
Image
For Zhang Fu, what truly interests investors is not simply a smarter drone, but a potential shift in the capability paradigm of aerial vehicles:
When tens of thousands of different flying platforms enter the real world, can they share a common set of continuously evolving intelligent capabilities?
Giving Drones a Brain
QbitAI: Embodied intelligence is so hot, yet most attention is focused on ground-based humanoid embodiments. Why did you choose aerial embodiment, which is even more difficult?
Zhang Fu: I studied control theory and artificial intelligence during my PhD. At the time, I saw some rudimentary quadrotor drones in our university laboratory and found them fascinating.
Beginning in 2015, I formally began working with drones in Professor Li Zexiang’s laboratory at HKUST. I carried out a great deal of foundational work, ranging from novel drone configurations and hardware to positioning, perception, and navigation.
Over the past two years, embodied intelligence has developed rapidly, and we began combining embodied intelligence with drones at an early stage.
As our research progressed, we needed more computing power, data, and patents.
At the same time, we received a large number of emails from potential customers hoping to use more intelligent aerial vehicles to solve problems that traditional methods could not adequately address.
Together, these two factors motivated me to start a company: first, to meet market demand; and second, to use a corporate platform to advance technological breakthroughs more effectively and take general-purpose aerial intelligence to the next stage.
Image
The places where drones can operate are often locations that are difficult or expensive for people to reach, or simply too dangerous. This gives drones an inherently irreplaceable role.
I believe the commercial development of aerial embodied intelligence will move faster.
If the first half of the aerial vehicle industry was about engineering maturity, then autonomous capability will be the key variable in the second half. This has already become a major industry pain point. Everyone is investing technology and capital in this direction, so the value of this sector will become evident sooner.
QbitAI: After the financing, there were reports that you were entering mass production. What tasks can your current products perform?
Zhang Fu: Because of limitations in technology and intelligence, manually operated drones still cannot reach many places. Those are precisely our target scenarios.
By the end of this year, we will mass-produce our first-phase products. Drones will no longer need to be remotely controlled by pilots and will be able to fly into more difficult and complex environments to complete specific tasks autonomously. Examples include utility corridors and tunnels in underground spaces; warehouses and factories indoors; and areas beneath bridges and in forests, where drones currently have difficulty operating and flying.
Going one step further, we want drones to possess autonomous reasoning capabilities—to reason about the world and make decisions like humans, determine how to fly based on their surroundings, and even rely on past experience to make judgments and complete tasks in completely unknown environments. It would be like putting a human brain into a drone.
Our technology is still undergoing R&D and iteration, and its stability and maturity need to be improved further.
Image
Third-phase products will be capable of interacting with the world—for example, by grasping objects. Commercialization and deployment in this area will require time to develop, including applications such as high-altitude operations and cleaning glass curtain walls.
At present, these tasks require people to be sent to elevated locations, which is extremely dangerous, costly, and inefficient. In the future, they could all be performed by intelligent aerial vehicles.
QbitAI: What are Xiyu Technology’s core competitive strengths?
Zhang Fu: Our team performs particularly well in two dimensions.
The first is original technological innovation. Over the past decade, the technologies we developed have been widely adopted by companies.
For example, while at DJI, I led the development of Livox LiDAR, which addressed the high cost and weight of LiDAR systems. Today, from lawn mowers, mobile robots, logistics robots, and robot dogs to the aerial embodied intelligence industry, a large number of drones use Livox.
For aerial equipment, the impact of high cost and weight is even more pronounced. Every additional bit of weight results in energy and flight-time losses. Our product should be among the lightest in its category, weighing less than one kilogram and measuring roughly the size of a single sheet of A4 paper.
Its flight time is similar to that of common drones, at 20–30 minutes. But it is important to note that absolute flight time is only one aspect. More important is effective flight time—the number of tasks that can be completed per unit of time.
When a drone is operated by a person, the operator must first assess the environment and then control the drone, constantly intervening and correcting its flight. All of this reduces the drone’s operational efficiency in ways that may not be immediately visible.
Once an aerial intelligent agent can solve problems autonomously, manual control is no longer necessary. It can fly autonomously at high speed, reach more places within a given period, and complete far more tasks than a manually operated drone—potentially several times as many.
A task that used to take 20 minutes might take only five minutes. In my view, this is actually the best way to address the flight-time problem.
Image
The second is our experience with mass production. We know how to turn technology into products that are reliable, consistent, and capable of addressing users’ pain points.
QbitAI: You spent eight years as a scientific adviser working on the front lines of industry. How has that experience helped you as an entrepreneur?
Zhang Fu: The greatest benefit was learning the entire process of turning technology into products.
I used to have a fairly typical academic mindset. I only wanted to pursue interesting, cutting-edge research that could be published in papers.
But I had little understanding of how the technologies we developed could be applied to products and the real world.
It was only after joining a company that I encountered the problems that arise during industrialization, the resources required, how to solve those problems, and how to bring products to market. Only by truly immersing yourself in the production process—communicating with R&D personnel at factories, observing workers, and even performing the operations yourself several times—can you understand the many difficulties and challenges involved in combining software and hardware and turning technology into a finished product.
This experience taught me to think beyond the academic perspective. I also began approaching and solving real industrial problems from the standpoint of commercialization.
QbitAI: The company attracted participation from investors such as Alibaba soon after its founding. People say they invested because of you personally. Do you think your status as a professor was the main reason?
Zhang Fu: Being a professor is one factor, and certainly a highlight. It signals our ability to carry out original technological innovation.
But ultimately, investors are not looking at a single identity label. They want to know whether the team can truly deliver on something difficult and long-term.
Reconstructing Digital-Twin Simulation Environments
QbitAI: The embodied intelligence industry is facing a data shortage. Does aerial embodiment suffer from an even greater lack of data? Is data collection also more expensive?
Zhang Fu: Yes, and the cost of acquiring it is higher. A great deal of ground-based embodied intelligence data can be collected directly by people, but aerial data requires an aerial vehicle to operate in three-dimensional space. It also involves pilots, equipment, safety, and site conditions. As a result, high-quality real-world aerial data is inherently scarce.
If we relied entirely on real flights to cover every environment, the cost would be extremely high, and many long-tail scenarios would be difficult to capture comprehensively.
That is why we have been exploring how to use a small amount of real-world data to build larger and richer training environments.
Image
QbitAI: How does Xiyu collect data?
Zhang Fu: If we relied solely on real aircraft to collect real-world data, even 10 years of accumulation might not be enough, and the cost would be incalculable.
We have built a Real-to-Sim data collection system. With only a small amount of real flight data, it can reconstruct highly realistic simulation environments and generate training data for different scenarios in batches, greatly reducing our dependence on real-aircraft data collection.
Real-to-Sim functions like an amplifier. A single piece of real-world data can be replicated and expanded thousands or even tens of thousands of times, enabling thousands or tens of thousands of repeated flights and addressing the scarcity of real-world data to some extent.
QbitAI: Would this create a sim-to-real gap?
Zhang Fu: That is precisely why we use Real-to-Sim rather than pure simulation.
Pure simulation can produce a substantial sim-to-real gap. Our gap is much smaller because the environment is not generated directly from noise (errors); it is reconstructed from real data and therefore closely matches the real world.
In architecture, this is called a digital twin: whatever the real world looks like is what the digital world built on a computer through Real-to-Sim looks like.
The better the Real-to-Sim reconstruction, the smaller the gap. We have accumulated a great deal of foundational technology, enabling us to achieve high fidelity at high speed. A scene covering dozens of square kilometers can be generated in roughly a dozen minutes.
QbitAI: Could the scenes generated by Real-to-Sim be too uniform and limited to a particular space?
Zhang Fu: A simulation is not static. We can further modify parameters such as object positions, environmental structures, viewpoints, and task conditions, expanding limited data into a richer training distribution.
Real-world data collection also does not involve collecting just one or two samples. We gather tens of thousands of real-world data samples, but do not need to collect too many for each category.
Real-world data only needs to provide sufficiently comprehensive coverage. Diversity can then be created in the simulation environment. Combining the two is more effective for supporting model training.
Image
QbitAI: What else can this environment be used for?
Zhang Fu: It can do more than solve the data shortage problem; it can also support reinforcement learning.
At present, most people are focusing on supervised learning. Supervised learning is largely imitation, whereas reinforcement learning involves the model iterating on its own.
After seeing a large number of scenarios, a model can remember distributional characteristics, generate behavioral reasoning based on memory and data, and continuously interact, self-play, and iterate in a simulation environment. Its level of intelligence could ultimately surpass that of humans.
This is one aspect that excites us greatly. We have even begun considering how to regulate the technology to prevent misuse.
QbitAI: Are the models already better than humans?
Zhang Fu: Not yet. But as data, models, and computing power continue to accumulate, machines will become increasingly capable. The next stage is to enable machines to handle real-world environments that are more complex and open-ended.
Aerial Intelligence Gradually Enters High-Risk, High-Cost Operations
QbitAI: What stage do you think China’s aerial embodied intelligence industry is at?
Zhang Fu: It is just getting started. No product has yet truly reached mass production, let alone achieved meaningful sales. Everyone is still exploring technical approaches and application scenarios.
QbitAI: How far are we from large-scale commercial deployment and mass production?
Zhang Fu: Not far—within two years. But there will be significant differences between different scenarios.
Image
QbitAI: What application scenarios are there for aerial intelligent agents?
Zhang Fu: We have worked with a customer that wanted to use our products to inspect infrastructure on hillsides.
There is infrastructure on hillsides that requires regular inspection and maintenance, such as wire mesh designed to prevent landslides and debris flows.
The mesh itself is only two meters high, but it is surrounded by trees. The aerial vehicle must perceive its surroundings autonomously, avoid obstacles such as branches and leaves, fly to key locations to take photographs, build a 3D model of the mesh, and check whether any areas need reinforcement or repair.
This work previously had to be done by people. Workers had to climb steep hillsides and make their way through weeds and loose rocks. The sites could be extremely hot, with mosquitoes and other insects, creating a dangerous and inefficient working environment. Safety incidents had also occurred in the past.
Now, aerial vehicles can gradually replace people in these high-risk, high-cost environments. As their manipulation capabilities mature further, they can also be used for more aerial operations, such as infrastructure inspections in mining and electric power, as well as warehouse inspections in the warehousing and logistics industries.
QbitAI: Beyond B2B scenarios such as municipal services, mining, and power, where could ordinary consumers use this technology?
Zhang Fu: In the future, as intelligent capabilities improve, the way people interact with aerial vehicles will change.
Take photography, for example. An aerial vehicle could have an aesthetic sense and capture emotions like a photographer, autonomously selecting angles, camera movements, and shooting methods. Consumers would effectively carry a photographer with them.
But all of this depends on one common prerequisite: aerial vehicles must first move from being “remotely controlled” to being “autonomous.”
Once this capability truly matures, aerial embodied intelligence will gradually expand from a handful of professional scenarios into a broader range of real-world applications.
Image