Dong Hao is an unusual figure in Chinese embodied intelligence. Most people know him as the Peking University doctoral supervisor who took the post at 29, or as the chief scientist now leading household robotics at Qiyuan. Fewer know that he spent three years running a hardware startup before his doctorate, living in an eight-bed factory dormitory, or that he was one of the earliest members of AgiBot when it was founded in 2023.
The interview he gave Leiphone is less about model architectures than about the part of robotics that kills companies: turning a working prototype into a product at volume.
Three years that ended in months
After finishing his master’s in 2012, Dong declined his supervisor’s suggestion to continue directly into a doctorate and returned to China to build wearable devices and portable brain-computer interfaces. The category looked promising. He lived at the factory, in a dormitory built for eight people, and had one to himself.
Then Xiaomi and Huawei entered wearables. Both had supply chain, channel and brand, and both moved fast. The market he had been building for was gone. His summary of that period is that no amount of effort was going to change the outcome.
He went back to Imperial College in 2015 for a doctorate, finishing in four years. During that period he open-sourced a framework that had started as an internal laboratory tool and it won the ACM MM best open source software award. He joined Peking University, where his first year of teaching landed inside pandemic lockdowns: a robotics course delivered entirely online, to students who kept their cameras off, in a subject where nobody could touch a robot.
The lesson he took from the startup is not that hardware is hard. It is that there is a time gap between a frontier technology and its market, and that gap is where most companies die. He estimates he bought that insight in three years, a cycle most founders take a decade to complete.
Why waiting is the expensive option
His argument against the waiting strategy now favoured by several embodied AI companies is blunt. Money kept in reserve does not grow into an advantage, and the competitor who spends it will get there first. He points at the car industry as the completed experiment: every funded startup that did not build died anyway, and roughly three quarters of the ones that did build also died. Building at least preserves a chance.
Where the money goes matters more than whether it is spent. Dong argues the embodied market is large enough that the entry points are numerous, and that many business-to-business scenarios do not need one model capable of every task. Replicable workstations in factories, and sorting in particular, generate near-term revenue without requiring general capability.
Qiyuan is taking the other road. The company is aiming directly at the endgame, which in Dong’s definition means household and consumer scenarios that require general task intelligence, in other words a generalist embodied model. The bet is that the companies which solve the easy industrial cases will build revenue but not a moat, while the company that solves the home owns the category.
His framing of the technical problem comes down to three words: HIL-UMI, scaling law and mass production. The first two concern how to generate training data at a cost that does not require tens of thousands of real robots, which is the constraint that keeps post-training budgets out of reach. The third concerns whether the resulting model can be built into something a factory can assemble at a price a household will pay.
The manufacturing point is where his startup experience shows. A prototype takes about six months. Producing a hundred units that hold tolerance, survive daily use and hit a cost target takes one to two years, and every design change restarts a cycle measured in half years. Anyone who has run that loop reads a production number as thousands of coordinated details rather than a figure on a slide.
He is also candid that he does not regret the three years. Research and building are different activities, and in his view the building is not optional.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/1mfsMzxNwaTmqZ7e.html.