The robot reaches into a commercial dryer, finds a pair of jeans still stuck inside, and lifts its arm again. Nothing about the motion looks remarkable. The company says it is testing a system that leaves a note of which angle failed, what changed on a retry, and whether the fix might transfer to another garment or robot.
That note is the point of the project.
On 17 November, Shanghai's Xuhui district government described a hotel-laundry robot built by Daxiao Robotics (大晓机器人) that handles the full wash-and-fold cycle and keeps adjusting its own behaviour. The same announcement named the company's chief scientist (首席科学家): Dacheng Tao (陶大程), a professor whose career until recently looked nothing like a factory floor.
The résumé behind the laundry room
Tao's official profile at Nanyang Technological University lists him as a Distinguished University Professor in the College of Computing and Data Science. It also lists a run of roles that would be a full career on its own: founding director of JD Explore Academy and senior vice president at JD.com, chief scientist of AI at UBTECH Robotics, an Australian Laureate Fellowship and a chair at the University of Sydney, and fellowships in the Australian Academy of Science, AAAS, ACM and IEEE, among others.
His degrees run from the University of Science and Technology of China to the Chinese University of Hong Kong and the University of London. NTU records more than 300 publications and over 146,000 citations, with an h-index above 180. Those figures move over time; the profile is the source.
At Daxiao Robotics, the chairman is Wang Xiaogang (王晓刚), a co-founder of SenseTime, according to the Xuhui government. The company is a spin-off culture rather than a startup in the usual sense: a group of researchers from computer vision and large models deciding that embodied intelligence (具身智能) is the next place where their skills matter.
What "dual-brain" means in practice
In a technical essay published in early November and relayed by Chinese technology media, Tao laid out a problem that sounds small and is not. A robot tries to pick up a cup, the grip slips, and it corrects. Most systems record the successful outcome and move on. Tao argues that the useful object is the whole episode — the prediction, the actual movement, the feedback, and the conditions under which the correction worked.
To capture that, Daxiao describes a two-part architecture it calls dual-brain collaboration (双脑协同):
- A cloud-side "brain" built on SenseNova, which reads the task, breaks it into goals, organises skills and reviews what went wrong after a failure.
- A local world model (世界模型), called Kaiwu (开悟), which tracks the physical state of objects, contact and motion, predicts what an action will do, and keeps an explicit measure of uncertainty.
The claim is that experience validated in one setting can be stored with its evidence and conditions, then reused by a different robot in a different place. That is a much harder claim than "the robot completed the task," and it is the one the company says it is testing.
The hotel deployment gives the idea a concrete setting. Xuhui's account says engineers adjusted the lifting angle of the robotic arm after clothes remained in the dryer, added a recognition check before button presses, and built in automatic reset, result verification and version rollback. After repeated trials on real machines, the rate of jeans left behind in the drum improved. The district did not publish a before-and-after number, which is worth noting.
Why a professor like this matters more now
Chinese robotics has no shortage of young founders. What it has lacked, until recently, is senior academic talent willing to work on the unglamorous end of the problem: reliability, error recovery and the slow accumulation of usable data.
Tao's path illustrates a broader shift. A researcher who spent years in computer vision and large models can move into robotics because the bottleneck has moved. Recognising a cup and planning a grip are increasingly software problems; making the same routine work on a Tuesday afternoon in a busy hotel is an engineering and learning problem. That is closer to the tradition of industrial research than to the image of the celebrity AI scientist.
It also shows how local government now sits inside this story. The Xuhui announcement was not a company press release; it was a district government item about a robot working in a local service scenario. For foreign readers, that is a useful signal: in China, a deployment in a hotel can be a piece of industrial policy as much as a commercial trial. The university profile, the district page and the company's technical blog together form the public record — and none of them is a peer-reviewed demonstration that the approach generalises.
Key takeaways
- Dacheng Tao (陶大程) is chief scientist (首席科学家) at Daxiao Robotics (大晓机器人); NTU lists him as a Distinguished University Professor and a fellow of several academies, with prior roles at JD and UBTECH.
- A November announcement from Shanghai's Xuhui district described a hotel-laundry robot using his team's "dual-brain" approach, with automatic error recovery and version rollback.
- The technical idea is to save not just successful outcomes but the conditions under which a correction worked, so that experience can transfer to another robot.
- The story is less about one professor than about senior AI researchers moving from model benchmarks into the physical reliability problems of service robotics.
Honest limitations
The role and résumé facts come from NTU's official staff page and a district government notice; the technical description comes from a company-authored essay relayed by Chinese tech media. No independent test of the dual-brain system has been published, and the district gave no measured improvement figure. Citation counts and h-index are snapshot values that change. This profile does not claim that the approach is superior to other robot-learning methods, nor that one hotel deployment proves generality. Chinese names and company names are given with their original characters where first used.
What readers can do now
If you are evaluating a robotics vendor that talks about "world models," ask what it does with a failed attempt — a system that only saves successes is not learning in the sense described here. If you work in facilities or hospitality, treat a hotel laundry as a reasonable test bed: it is repetitive, measurable and full of small variations that break scripted machines. And if you read a professor's new title as a career footnote, look again; in Chinese robotics, the move into a company is often where the interesting engineering now happens.
