In embodied AI, you can't jump in and do everything at the outset, nor can you do nothing at all.
Author |Qi Chengyong
Editor | Cen Feng
Dong Hao is a thoroughly atypical figure in the embodied intelligence industry.
For most people, their impression of him stops at his status as a young scholar who was already a doctoral supervisor at Peking University at 29, and his recent industry credentials as chief scientist at Qiyuan Robotics, leading the push to advance embodied intelligence for the home.
But few know that before pursuing his doctorate, he spent three years as an entrepreneur in an ultimately unsuccessful startup and lived in a factory dormitory shared by eight people; even fewer know that when AgiBot was first founded, he was among its earliest joiners, a core member who helped the fledgling startup team overcome countless hurdles.
While the industry was still debating lofty narratives about foundation models and general-purpose robots, he was focused on another decidedly unglamorous but very real problem: post-training that relies on tens of thousands of physical robots is so prohibitively expensive that it cannot be put into practice—how can this approach be broken through?
His most recent answer came in the form of three keywords: HIL-UMI, Scaling Law, and mass production.
Yet the underlying logic behind this answer had long been written into the way he had lived his life—“compressing” every step along the way.
After graduating with his master's degree, he politely declined his advisor's direct recommendation to continue on to a PhD, returned to China to found a company working on wearable devices and brain-computer interfaces, and threw himself into factory life, living in an eight-person dormitory. Three years later, after witnessing firsthand how quickly the market was squeezed by the major tech companies, he didn't linger in the fight; instead, he turned around and returned to Imperial College London to pursue a doctorate, graduating successfully four years later. The open-source framework he led in developing during that period also won the ACM MM Best Open Source Software Award.After graduating, he joined Peking University, and not long after starting his job, the pandemic hit. He spent a full year teaching robotics online to students who never turned on their cameras. In 2023, Zhiyuan Robotics was founded, and he joined as a core member, establishing the Peking University–Zhiyuan Joint Laboratory; now he has moved on to become Qiyuan’s chief scientist, plunging headfirst into the most challenging ToC segment of embodied intelligence.
In the industry, Dong Hao has always had a wide network of friends. His experience in entrepreneurship, elite academia, and leading industry also gives him an unusual sense of cross-boundary fluidity. Having experienced the entire journey of a hard-tech company, from inception to its final chapter, and then plunging back into the highly academic environment of Peking University, he now stands once again at the forefront of industry. How does he view the embodied intelligence industry today? With this question in mind, we met Dong Hao on the Peking University campus.
Below is the interview, edited by AI Tech Review without changing its original meaning.
01
A person who “runs life on fast-forward”
▎AI Tech Review: You completed your undergraduate and master’s degrees in the UK in four years and became a doctoral supervisor at Peking University at 29. That pace compresses three to five years in academia. Are you aware of that?
Dong Hao:I never deliberately calculated it. The British undergraduate and master’s system is inherently fast; finishing a master’s in four years is perfectly normal.
▎AI Tech Review: But it isn’t normal to join the faculty at Peking University before turning 30, and in your first year as a faculty member you ran into the pandemic.
Dong Hao:My first class was online. The students kept their cameras off and didn’t say a word; I did all the talking, sometimes asking whether anyone had questions, and no one answered. With robotics taught entirely online, you couldn’t touch the robots. The worst part was that the lockdown would suddenly be lifted and we could use them for a while, and then suddenly be reimposed, so everything was intermittent. Many people didn’t make it through. I was lucky enough.
▎AI Tech Review: Do you attribute it all to luck?
Dong Hao:And the fact that my work has a clear focus. Robotics is a field that both academia and industry pay attention to, so nothing is wasted on either front.
▎AI Tech Review: “Neither front is wasted”—that sounds very much like you. You seem naturally averse to working along a single dimension.
Dong Hao: Maybe. For example, the AI framework I developed previously was originally an internal lab tool. It seemed a shame to leave it unused, so I open-sourced it. As a result, it gained an increasing number of users and even won an award. One project gave rise to several things, which was highly efficient.
02
The entrepreneurial journey of a post-90s associate professor at Peking University
▎AI Tech Review: You graduated with a master’s degree in 2012. Your advisor, Guo Yike—a foreign member of the Chinese Academy of Engineering and the current Chief Vice-President of HKUST—urged you to pursue a PhD, but you did not listen and returned to China to start a business. What exactly happened?
Dong Hao: As a graduate student, I conducted research under Professor Guo. He told me to pursue a PhD directly, but I decided to return to China to start a business. He supported me and even gave me money. Later, in 2015, I went back to pursue my PhD, and he did not say anything about it. He gave me a great deal of freedom to choose my own research topic.
▎AI Tech Review: Did you start a business making wearable devices?
Dong Hao: Yes, the momentum was very strong at the time. When I was starting the business, I lived at the factory, where conditions were poor and everyone shared eight-person rooms. I was treated slightly better and had an eight-bed room to myself, so I could sleep in whichever bed I wanted.
▎AI Tech Review: Those three years must have been tough.
Dong Hao: It was not exactly that hard, but I could see the entire journey of a startup team. Wearable devices: portable brain-computer interfaces and smart wristbands—we built those first. The former targeted the future, while the latter targeted the present. Xiaomi and Huawei later entered the wearable market, but they moved incredibly fast once they got started. We could only watch as the market disappeared.
▎AI Tech Review: What is the hardest part of hard-tech entrepreneurship?
Dong Hao: Mass production.A prototype can be built in six months, but if you want to manufacture 100 products, keep costs under control, achieve a high yield, and make them durable, that takes one or two years. Even a minor change—modifying the mold, adjusting the production line, and revalidating—takes a six-month cycle. In three years, there are actually only a few rounds of trial and error. (Leiphone)
▎AI Tech Review: So, now that you’re seeing AgiBot and Qiyuan get mass production up and running, how does that feel?
Dong Hao:Mass production is a tough nut to crack. Outsiders see mass production as a number, but people who have worked with hardware see thousands of details working in concert and a set of test cases thick as a book. I used to find it difficult even when I was making something that small. Now we have to build a complete robot, passing every standard one by one. It is extremely difficult.test cases
▎AI Tech Review: You just mentioned that the smart-wearables market has been swallowed up by Xiaomi and Huawei. How does it feel to get steamrolled by the big players?
Dong Hao:It’s that you know no matter how hard you try, it won’t help. They have the supply chain, distribution channels, and brand all in place. I later summed it up this way: there is a time lag between frontier technology and the market. That’s a lesson I paid three years to learn.
▎AI Tech Review: Many entrepreneurs take a full decade to complete that cycle, but you did it in three years.
Dong Hao:When you put it that way... I guess so.
▎AI Tech Review: That’s called cutting your losses.
Dong Hao:But actually, I don’t regret those three years. Doing research and getting your hands dirty are two different things, but you have to get your hands dirty; there’s no way around it. Back in the day, among the car makers, everyone who raised money but didn’t build a car died; of those who did build, three-quarters died. You still have to build.
▎AI Tech Review: So you think it’s wrong for many embodied-AI startups today to spend no money, pursue no mass production, and adopt a wait-and-see strategy?
Dong Hao:How should I put it... Put bluntly, even if you save your money, you won’t have much saved. Either way, you’re going to have to take a gamble. Even if you have more money, can you hold your own against the giants? Building cars is a perfect example, isn’t it? If you build, you’re slightly better off, with at least a sliver of a chance of winning.
▎AI Tech Review: So what do you think embodied AI companies should spend their money on?
Dong Hao: One advantage of embodied AI is that its market potential is enormous, so there are many areas where breakthroughs are possible. For example, in many B2B scenarios, they don’t necessarily need a single model to perform many different tasks, so with this approach, short-term returns are relatively easy to achieve—for example, by developing replicable factory workstations, such as sorting stations. As for how to allocate spending, it depends on the direction each company is focusing on.
▎AI Tech Review: But doesn’t this approach seem somewhat inconsistent with what Qiyuan is doing?
Dong Hao: Qiyuan’s approach is to aim for the endgame right from the start.
▎AI Tech Review: What do you mean by the endgame?
Dong Hao: For household and consumer scenarios, general-purpose task intelligence is essential. In short, what’s needed is a generalist embodied AI large model, whereas other scenarios don’t necessarily need one.
03
From AgiBot to Qiyuan: The Hardest Choice
▎AI Tech Review: A lot of people tell us that you’re a social genius and that everyone likes you.
Dong Hao: Perhaps because I’ve always believed that the industry needs everyone to work together. We’re all friends, and there’s so much we can do in this industry. I still enjoy the process of a group of people pushing forward together.
▎AI Tech Review: You joined AgiBot as soon as it was founded in 2023; it was Zhihui and Mr. Deng (AgiBot’s Deng Taihua) who recruited you.
董豪：The division of labor is different. AgiBot focuses primarily on deployment in toB scenarios, but Qiyuan’s toC business is different. Robots for homes need a genuine brain, general-purpose navigation, the ability to interact with people, and general-purpose task intelligence. It is the hardest and most generalizable. For frontier research in embodied AI, toC is the most suitable battlefield.
▎AI Tech Review: For example?
董豪：For example, engines for civil aviation and military use clearly have different requirements. Industrial embodied AI may not need a single model to do a hundred different things, but toC really does.
▎AI Tech Review: Going from AgiBot to Qiyuan, so much has to be done in-house. Doesn’t that make things even harder?
董豪：Every pitfall has to be gone through. In short, through my previous startup, I learned that a wait-and-see strategy is wrong. You wait for others to burn money until they find the direction, while you just hunker down on the sidelines. Does hunkering down give you a higher chance of winning? No. Even if you eventually wait it out, it still won’t be yours.
▎AI Tech Review: What has it been like moving from academia to industry?
董豪：The pace and style are different, especially when it comes to engineering. For example, when I first joined Peking University, Yang Yaodong (a doctoral advisor at Peking University and chief scientist at PsiRobot) brought me and Wang He (a doctoral advisor at Peking University and CTO of Galbot) into a Beijing Association for Science and Technology project on dexterous manipulation.
Although it was a top-tier project, it didn’t have much funding. I remember that back then, if we wanted to buy hands (dexterous hands), the only one we could afford was the Shadow Hand (from a British dexterous-hand company; US$100,000 per hand). It was ridiculously expensive, so Professor Yang and I each had to buy one hand—Yang the left and I the right—and cobble them together for our research. The environment is very different now.
Funding for purely academic research is limited and cannot support research into today’s large models. Such research now requires not only frontier work in academia, but also large-scale computing power and solid engineering infrastructure. Academia also needs to work with industry.
▎AI Tech Review: How will the current environment affect young people? Students Wang Yuanfei and Zhang Jiyao are also working on embodied AI now. What do you think?
董豪：This generation has far better conditions than I did. Our technical vision, speed of hands-on execution, and financing environment are simply incomparable to what we had back then.
Embodied AI has ample room to grow. It’s a good thing for young people to venture out, and I’m also very optimistic about the directions of their startups.
▎AI Tech Review: Do you encourage entrepreneurship unconditionally?
Dong Hao:Not necessarily. Timing matters. For some students, I would advise them to first leverage a major platform to build a solid technical foundation, and then strike out on their own once they have accumulated enough experience. A supply chain, an ecosystem, and a technical team all take time and cannot be bought. Speed is not the problem; speed without a solid foundation is.
▎AI Tech Review: How much time does Qiyuan give itself?
Dong Hao:For the long term, we’re looking at three years. For task intelligence, start with scenarios where failure poses no risk, such as clothing-related tasks that have to be done every day. Whether the folding is done properly or not, it won’t cause any harm; there is a wide safety margin, and a failed attempt can be retried. We have already developed a fully autonomous laundry workflow: the robot plans its path autonomously and opens and closes the washing machine door; if the grasp fails, it proactively retries, and the entire process requires no human intervention. You can’t try to do everything from the start, and you can’t just sit around doing nothing either. First focus on getting a few capabilities fully working, then expand horizontally.
04
Technical Judgment: The Deciding Factor Is Post-Training
▎AI Tech Review: Let’s talk about technology. You have a rather contrarian view: “Pretraining and post-training aren’t a matter of proportions; they succeed one another at different stages.” Most people are still discussing how much post-training data to pair with pretraining data, while you simply reject the premise?
Dong Hao:It’s not that I’m rejecting it; the stages are different. Pretraining addresses the initial construction of the “worldview.” The path is actually clear: collect the data, train it in, and use brute force to work miracles, taking the success rate from 0 to a certain level, but there is a ceiling to the improvement. Large-scale post-training will be bigger than pretraining as currently conceived, and may even completely subsume it.
▎AI Tech Review: Where is the ceiling?
Dong Hao:Imitation learning can only serve as a cold start. All the training samples are correct trajectories; if the robot makes a mistake, it has no idea how to adjust and get back on track. The real world doesn’t follow the distribution of your demonstration data.
▎AI Tech Review: So that means we need to bring in reinforcement learning.
Dong Hao:Yes. AlphaGo followed this same path back then. It started with imitation learning from game records, but what truly cemented its legend was large-scale reinforcement learning through self-play. ChatGPT was the same later on: at its core, reinforcement was still carried out through RLHF. Today's robots are no different. For generalization and peak success rates, the entire game hinges on scalable post-training. The first half is about who can first use imitation learning to bring robots to life; the second half is about who can build the robots' error-correction networks at the lowest marginal cost and with the greatest speed.
▎AI Tech Review: But scaling up post-training is bottlenecked by real-world robots.
Dong Hao:That's the physical constraint. How could you have tens of thousands of real robots simultaneously conducting post-training in tens of thousands of real-world rooms? The cost of manufacturing even 10,000 of them would be prohibitive. AgiBot, founded only three years ago, has already mass-produced 20,000 units, which is already an impressive achievement.
▎AI Tech Review: So Qiyuan uses HIL-UMI for post-training without collecting data from the robot itself.
Dong Hao:Yes. This is one of our recent popular works. People do not need to operate the robot itself. Wearable and handheld cameras record the manipulation process and directly convert it into trainable trajectories. With the same budget, this delivers greater improvements in model capability, and it can support post-training by targeting the model's weaknesses for optimization, bringing the success rate close to 100%.
▎AI Tech Review: With tens of thousands of samples collected, can post-training be scaled up?
Dong Hao:Half of it. The other half is data efficiency. For example, right now we can use a world model/generative AI to turn a single real-robot trajectory into 50 equivalent samples with different object placements and spatial positions. This multiplies the value of real-robot data by 50. It is human demonstrations, real-robot anchors, and generative augmentation combined that make up "scalable post-training."
05
Horizontal Scaling Law: Personality Turned into Theory
▎AI Tech Review: The “Horizontal Scaling Law” you mentioned has one axis for data volume and another for task volume. Why introduce the task-volume axis?
Dong Hao:Many companies are now on an inefficient growth curve: develop one use case, overfit it, and ship it; for the next use case, start from scratch all over again. Data volume and task count are related only linearly, so marginal returns diminish. After scaling horizontally, the initial success rate for new tasks keeps rising, while the number of samples needed to achieve a high success rate keeps falling. The more you learn, the more you save.
▎AI Tech Review: Let me ask a question that may be an overinterpretation. You speed-read books, are three years into entrepreneurship, and now work on both Peking University and Qiyuan at the same time. “Not getting stuck grinding on a single axis, but accelerating on two dimensions at once”—is this way of thinking essentially your personality translated into a theory?
Dong Hao:That question... (laughs)
▎AI Tech Review: Just answer the question.
Dong Hao:It could be. To me, the work at Peking University and at Qiyuan is not two jobs but one and the same thing: pursuing technological paradigm innovation on one side and validation at scale on the other, with the two mutually supporting each other.
06
On “Pace”
▎AI Tech Review: Let’s talk about something else.In your daily life you like motorcycles and skiing. Skiers all say that when fear hits, the body's instinct is to recoil backward, and the more you recoil, the more you fall. What do you think?
Dong Hao: Just keep the pace under control. It's not that there's no fear; it's about controlling the fear.
▎AI Tech Review: People who compress time — might they also be prone to losing control, because they move fast?
Dong Hao: That's why, whether in research or in a company, it's the pace that matters. Speed isn't the problem; disorder is. How much to put in and when — the R&D pace has to be well managed.
07
Afterword
The conversation between Dong Hao and AI Tech Review took place in a traditional Chinese courtyard on the Peking University campus, called Jingyuan Courtyard No. 5, which serves as the office and teaching site of the Frontier Computing Research Center of the School of Computer Science at Peking University.
The Peking University Frontier Computing Research Center was founded by Turing Award laureate John Hopcroft and Academician Gao Wen, with the aim ofcultivating a new generation of talent in computing and producing influential research.The centerpapers and grant funding applications face no requirements whatsoever, and its review process is aligned with international standards.
Cohort after cohort of Turing Class students have gone forth from here.
Here, Dong Hao has been deeply committed to education and research, and Qiyuan's cutting-edge academic work has also taken place here. In a sense, the way he speaks, his life, and his theory are all the same thing: compression. Compressing a ten-year path into five years, squeezing fifty times the value out of a single piece of data, compressing a one-dimensional industry into a new axis.
"Horizontal Scaling Law" may be not just a technical trend; it is more like this person's self-description. (Leiphone(official account: Leiphone))
Originally published by Leiphone; unauthorized reproduction is prohibited. For details, seethe reproduction notice.
