On 28 October, Chen Mingxiu registered Muling Technology in Hangzhou’s Yuhang district. Registered capital was about RMB 1.5625 million. The stated direction is physical recursive self-improvement, the idea that a robot should improve its own learning process through contact with the physical world rather than depend on human annotation and on-site tuning.
Chen previously headed the Leyun embodied intelligence platform at Alibaba’s DAMO Academy. His departure is the second loss to that team in just over a month, following chief scientist Zhao Deli, who left on 14 September for a destination that has not been disclosed.
The record behind the move
Chen holds 22 technology patents. Before Alibaba he led the vision algorithm team at Rokid, building the augmented reality glasses engineering group from nothing and taking the first generation of police-use AR glasses into commercial deployment.
At Alibaba he began in the security division, running AI algorithm development across several hundred thousand edge cameras, then moved into embodied intelligence and built the Leyun platform. Its defining feature was device-cloud collaboration that allowed embodied models to run on the device with no GPU, which cut both the development threshold and the deployment cycle. It is now deployed at 18 universities including Tsinghua and the University of Hong Kong, and was selected as a UN AI for Good case study.
The through-line in that career is consistent: taking expensive technology and making it cheap enough for ordinary institutions to use. Police AR glasses, edge cameras and the Leyun platform are the same move at different scales, and the startup applies it again.
Why physical RSI became the default pitch
Chen is joining a category that formed quickly. SeeAct AI was founded by Mu Yao, an assistant professor at Shanghai Jiao Tong University’s Institute for AI. Simate was founded by Zhang Ying, formerly the core lead on Horizon Robotics’ end-to-end driving system, and reached the top of the RoboDojo physical AI benchmark within three months of founding before raising several hundred million across consecutive rounds. Rimian Kaiwu was founded by Zhong Yuanxin, a former Huawei Genius Youth, and published a white paper in September proposing its own capability grading framework for physical RSI.
The shift is from building a brain, stacking models and assembling datasets, to building a system that improves itself. The commercial logic behind the shift is straightforward. Every store, every factory and every site is different, and if each deployment requires engineers to tune the model on site, the business stays a series of bespoke projects forever.
That is no longer a matter of opinion. The China Academy of Information and Communications Technology’s 2026 embodied intelligence report moved the sector’s core bottleneck from algorithms that are not intelligent enough to models that are hard to deploy.
The same week, a demonstration
AgiBot offered an illustration on 27 October when it released Genie Master, a self-evolution framework, and put a G2 unit into a working convenience store to learn on the job. The robot located its own failures, adjusted its strategy and retained the result without an engineer walking it through the task. The company describes it as the first validation of physical RSI in a live commercial setting. Six G2 units are scheduled to run a 24-hour livestream from 2 November, covering the lunch peak, the evening peak, overnight duty and the breakfast shift, as a form of probation.
The demonstration is more convincing than a benchmark because a convenience store does not care about scores. It cares whether the machine can restock a shelf at four in the morning without supervision.
What the talent flow actually changes
DAMO Academy’s shift from frontier exploration to self-financing makes attrition close to inevitable, and Alibaba is not the only large platform losing people. The consequence is that the scarcest input in Chinese embodied AI is no longer capital. Funding is abundant and, as the last month has shown, increasingly cautious. What is scarce is people who have run a platform at scale and know why deployments fail.
Those people are now inside companies small enough that the answer to “can marginal cost fall at the hundredth store and the thousandth robot” determines whether the company exists in three years. Chen’s bet is that the answer comes from a machine that learns from its own mistakes. It is a good bet to test with someone else’s capital and a hard one to prove with your own.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-10/ART-8321205-12003-30709109.html.