Post-90s Peking University associate professor Dong Hao: How Qiyuan Robot uses HIL-UMI and the horizontal Scaling Law to tackle home embodied intelligence

Most people's impression of him stops at the identity of a young scholar who became a Peking University doctoral supervisor at 29, and at the industry label of his recent appointment as chief scientist of Qiyuan Robotics, leading the team attacking home embodied intelligence. But few know that before his PhD he had a three-year entrepreneurial venture that ultimately failed, and that he once lived in an eight-person factory dormitory; even fewer know that when Zhiyuan was first founded, he was one of the earliest members to join, one of the core members who accompanied the startup team through countless hurdles.

While the industry is still hotly debating the grand narratives of large models and general-purpose robots, he is staring at another thoroughly unsexy real problem: doing post-training on tens of thousands of real machines costs so much that it is hard to deploy—how is this path to be broken?

His recent answer is three keywords: HIL-UMI, Scaling Law, and mass production.

And the underlying logic of this answer was long ago written into his life of "moving in compressed mode."

After finishing his master's, he politely declined his advisor's suggestion to go straight into a PhD, returned to China to found a company making wearable devices and brain-computer interfaces, plunged into the factory, and lived in an eight-person dormitory. Three years later, having witnessed the market being rapidly squeezed by big companies, he did not linger; he turned around and went back to Imperial College to pursue a PhD, graduated smoothly four years later, and the open-source framework he led the development of during that time won the ACM MM Best Open Source Software Award. After graduation, he joined Peking University and soon ran into the pandemic; facing students who kept their cameras off the entire time, he taught a robotics course online for a full year. When Zhiyuan Robotics was founded in 2023, he joined as a core member and established the Peking University–Zhiyuan Joint Lab; now he has moved on to serve as Qiyuan's chief scientist, plunging into the hardest ToC embodied intelligence track.

In the industry, Dong Hao has always had many friends, and his résumé of entrepreneurship, top-tier academia, and leading industry players carries a rare sense of cross-boundary mobility. A person who has personally experienced the entire process of a hard-tech company from launch to closure, then turned to immerse himself in Peking University with its extremely high academic density, and now stands at the very front of industry—how does he view today's embodied intelligence industry? With this question, we met Dong Hao on the Peking University campus.

01

A man who "moves in compressed mode"

▎AI Tech Review: You finished your British undergraduate and master's degrees in four years and became a Peking University doctoral supervisor at 29. That pace is compressed by three to five years in academia. Are you aware of it yourself?

Dong Hao: I never deliberately calculated it. The British undergraduate-plus-master's system is fast to begin with; finishing a master's in four years is normal practice.

▎AI Tech Review: But joining Peking University before 30 is not normal, and in your first year of holding a teaching post you ran into the pandemic.

Dong Hao: My first class was online, the students didn't turn on their cameras, didn't make a sound, and I lectured the whole time by myself; sometimes I asked if there were any questions, and no one replied. Doing robotics entirely online—you can't touch the robots. The worst part was suddenly being unblocked and able to use things for a bit, then suddenly locked down again, on and off. Many people didn't make it through; I was fairly lucky.

Dong Hao: Also, what I do is focused. The robotics direction draws attention from both academia and industry, so neither end is wasted.

02

The entrepreneurial past of a post-90s Peking University associate professor

▎AI Tech Review: In 2012, when you finished your master's, your advisor Guo Yike (foreign academician of the Chinese Academy of Engineering, now provost and deputy vice-chancellor of HKUST) urged you to pursue a PhD, and you didn't listen; you returned to China to start a business. What exactly was the situation?

Dong Hao: As a graduate student I did research with Professor Guo, and Professor Guo told me to go straight into a PhD, but I decided to return to China and start a company. He supported me and even gave me money; later, when I went back in 2015 to do my PhD, he didn't say anything about it and gave me great freedom to choose my own topic.

Dong Hao: Yes, the momentum was very strong at the time. While starting the company I lived in the factory; the conditions were very poor, all eight-person rooms. I had it a bit better—I had an eight-person room to myself, eight beds, and I could sleep in whichever one I wanted.

Dong Hao: It wasn't really hardship; it was just seeing the complete fate of a startup team. Wearable devices: portable brain-computer interfaces and smart wristbands—we did them first; the former facing the future, the latter facing the present. Xiaomi and Huawei later entered wearables, but once they got going they were very fast. I watched the market disappear before my eyes.

Dong Hao: Mass production. A prototype can be made in half a year, but to build a hundred products, you have to control costs, keep the yield rate high, and make them durable—that takes one to two years. A small change means changing molds, adjusting production lines, and re-validating, with cycles measured in half-years. In three years, there really aren't many rounds of trial and error. (Leiphone)

Dong Hao: Mass production is hard to chew. Laypeople see mass production as a number; people who have done hardware see mass production as thousands of details working together, with test items as thick as a book. I used to find it hard even making something that small; now we have to build complete robots, passing standard after standard—this is extremely difficult.

Dong Hao: It's just that you know no matter how hard you try, it's useless. They have the entire supply chain, channels, and brand. I later summed it up: there is a time gap between frontier technology and the market—that is the understanding I bought with three years.

Dong Hao: Now that you mention it... it seems so.

Dong Hao: But actually I don't regret those three years. Although doing research and going into the field to do it are two different things, you must go into the field and do it; you can't not do it. Back then in carmaking, those who raised money and didn't build all died; of those who did build, three-quarters died. You still have to do it.

Dong Hao: How should I put it... To put it plainly, even if you just save the money, you can't save much anyway; you definitely have to take a gamble. You have a lot of money—can you compare with the giants? Carmaking is exactly the example, isn't it? Building is still a bit better; there's still a sliver of a win rate.

Dong Hao: Embodied intelligence has one advantage: because its market space is so large, there are many points where you can break through. For example, in many toB scenarios, it doesn't necessarily need one model that can do many different tasks, so doing this kind of thing can yield relatively good short-term returns, such as replicable workstations in factories, like sorting. As for how to spend money, that depends on each company's main direction.

Dong Hao: Qiyuan's idea is to be able to aim for the endgame all at once.

Dong Hao: For homes and consumer scenarios, general-purpose manipulation intelligence must be achieved. In short, a generalist embodied intelligence large model is needed, while other scenarios do not necessarily need it.

03

From Zhiyuan to Qiyuan: The Hardest Option

Dong Hao: Maybe because I have always felt that the industry needs to work together, that we are all friends, and that there is a lot this industry can do. I still like the process of a group of people pushing forward together.

▎AI Tech Review: You joined Zhiyuan as soon as it was founded in 2023, when Zhihui and President Deng (Zhiyuan's Deng Taihua) called you in. Now Zhiyuan has over a thousand people, and its mass-production deliveries are running at the top of the industry. Why did you leave to start Qiyuan?

Dong Hao: The division of labor is different. Zhiyuan focuses mainly on landing toB scenarios, but Qiyuan's toC is different. Robots for the home must truly have a brain, must have general navigation, must interact with people, and must also have general manipulation intelligence—the hardest and most generalizable. For cutting-edge embodied research, toC is the most suitable battlefield.

Dong Hao: For example, when building the same engine, the requirements for civil aviation and military use are definitely different. Industrial embodied intelligence does not necessarily need one model to do a hundred things, but toC really does.

Dong Hao: You have to step in every pit that must be stepped in. In short, from my previous startup experience I learned that a waiting strategy is wrong. You wait for others to burn money and find a direction, while you hide on the sidelines—hide successfully and your odds of winning are greater? No. Even if you wait until it happens, it still isn't yours.

Dong Hao: The rhythm and style are different, especially in engineering. For example, when I first entered Peking University, Yang Yaodong (PhD advisor at Peking University, chief scientist at Lingchu Intelligence) pulled me and Wang He (PhD advisor at Peking University, CTO of Galaxy General) into a Beijing Association for Science and Technology project on dexterous manipulation.

Although the project was top-level, the funding was not much. I remember that at the time we wanted to buy hands (dexterous hands), and could only buy the Shadow Hand (a British dexterous hand company, 100,000 US dollars per hand). The price was outrageously expensive, so Teacher Yang and I could only each buy one left hand and one right hand, piecing them together to do research. The overall environment is different now.

Funding obtained from purely academic research is small and cannot support today's large-model research. Now we need not only cutting-edge academic research, but also large-scale compute and good engineering infra. Academia also needs to combine with industry.

Dong Hao: Their generation has far better conditions than I did back then. Their technical vision, speed of execution, and financing environment are all beyond comparison with what we had. The space for embodied intelligence is large enough, and it is a good thing for young people to venture into it. I am also very optimistic about their startup directions.

Dong Hao: Not necessarily; timing is important. For some students, I advise them to first use a big platform to make their technology solid, and come out after accumulating enough. Supply chains, ecosystems, and technical teams take time; you can't buy them with money. Being fast is not the problem; being fast without accumulation is the problem.

Dong Hao: Take the long term as three years. For manipulation intelligence, first do scenarios where failure carries no risk, such as clothing. It has to be done every day, and whether it is folded well or not will not cause an accident. The safety boundary is high, and if it fails it can retry. We have already implemented a fully autonomous laundry process: the robot autonomously plans the path, opens and closes the washing machine door, and actively retries if grasping fails. The whole process has no human intervention. You cannot do everything right from the start, nor can you do nothing and just wait there. First prioritize breaking through some capabilities, then expand horizontally.

04

Technical Judgment: The Decisive Factor Is Post-Training

Dong Hao: It is not a rejection; it is that the stage is different. Pretraining solves the initial construction of a worldview. The route is actually clear: collect data, train it in, and use scale to produce miracles, pulling the success rate from 0 to a certain success rate, but the improvement has an upper limit. Large-scale post-training will be larger than what is now called pretraining, and may even completely cover it.

Dong Hao: Imitation learning can only cold-start. The training samples are all correct trajectories, and the robot does not know how to adjust back after making a mistake. The real world does not follow the distribution of your demonstration data.

Dong Hao: Right, AlphaGo followed this path back then. It started with imitation learning from game records, but its true apotheosis came from large-scale reinforcement learning through self-play. Later ChatGPT was the same; the core was still RLHF for reinforcement. Today's robots are the same: generalization and extreme success rates—the decisive factor all lies in scalable post-training. The first half is about who can first use imitation learning to light up the robot; the second half is about who can build the robot's error-correction network with the lowest marginal cost and the fastest speed.

Dong Hao: This is the physical constraint. Do you want tens of thousands of real machines doing post-training simultaneously in tens of thousands of real rooms? The cost of building ten thousand units is already enormous. Zhiyuan has mass-produced 20,000 units in 3 years since its founding, which is already very strong.

▎AI Tech Review: So Qiyuan is doing HIL-UMI, post-training with body-free data collection.

Dong Hao: Right, this is one of our recent works that has become relatively popular. A person does not need to operate the robot body; wearable and handheld cameras record the operation process and directly convert it into trainable trajectories. With the same budget, the model capability improvement obtained is higher, and it can support post-training and targeted optimization of model deficiencies, bringing the success rate close to 100%.

Dong Hao: Half. The other half is data efficiency. For example, we can now take one real-machine trajectory and, using a world model/generative AI, generate 50 equivalent samples with different object placements and spatial positions. The value of real-machine data is amplified 50 times. Human demonstrations, real-machine anchors, and generative augmentation—put together, that is what makes "scalable post-training."

05

Horizontal Scaling Law: The Theorization of Personality

▎AI Tech Review: The "horizontal Scaling Law" you mentioned has data volume as one axis and task volume as another. Why add the task volume axis?

Dong Hao: Many companies are now on an inefficient growth curve: do one scenario, overfit, deliver; next scenario, start from zero again. Data and tasks only have a linear relationship, with diminishing marginal returns. After horizontal expansion, the initial success rate of new tasks will keep getting higher, and the sample size needed to reach a high success rate will keep dropping. The more it learns, the more it saves.

▎AI Tech Review: Let me ask a question that may be overreading things. You read books in compressed form, you've been starting a company for three years, and now you're doing both Peking University and Qiyuan at the same time. "Not grinding away on a single axis, but accelerating on two dimensions at once"—is this idea a theorization of your personality?

Dong Hao: That question (laughs)

Dong Hao: Possibly. For me, the things at Peking University and the things at Qiyuan are not two jobs, but the same thing: on one side doing technical paradigm innovation, on the other side doing scaled validation, and the two ends support each other.

06

About "Rhythm"

Dong Hao: Just control the rhythm well. It's not about not being afraid; it's about controlling fear well.

Dong Hao: So whether in scientific research or in a company, rhythm is what matters. Speed is not the problem; chaos is the problem. How much to invest and when—the R&D rhythm must be well controlled.

07

Postscript

The Peking University Center on Frontiers of Computing Studies was founded by Turing Award winner John Hopcroft and Academician Gao Wen, with the aim of cultivating a new generation of talent in the computer field and producing influential research. The center has no requirements on the number of papers or funding applications, and its review method is aligned with international standards.

One cohort after another of Turing Class students has walked out from here.

Here, Dong Hao has long been deeply engaged in education and scientific research, and Qiyuan's frontier academic research also happens here. To some extent, his way of speaking is the same thing as his life and his theory: compression. Compressing someone else's ten-year path into five years, compressing the value of one piece of data into fifty, compressing a one-dimensional industry into a new axis.

"Horizontal Scaling Law" may not just be a technological trend; it is more like this person's self-description. (Leiphone)