A pharmacy robot in Shenyang moves boxes down an aisle nobody rebuilt for it. It takes up roughly 2.5 square metres of shop floor, plugs into the store's existing order system, and is expected to do one unfashionable thing: keep working.
The team that built it is run by a professor who spent years telling the robotics field that it was counting the wrong thing.
Key takeaways
- Cewu Lu (卢策吾) is co-founder (联合创始人) and CEO of Noematrix (穹彻智能), founded in late 2023 and incubated out of Flexiv (非夕科技), a robotics group he did not run.
- He is vice dean (副院长) of Shanghai Jiao Tong University's School of Artificial Intelligence and vice dean of the Shanghai Chuangzhi Institute, and in 2026 received the ACM China Distinguished Researcher honour.
- He co-authored the Chinese standard term for embodied intelligence (具身智能) with Wang He (王鹤), published by the China Computer Federation (中国计算机学会) in July 2023.
- Noematrix raised another several-hundred-million-yuan round in June 2026 from investors including Shanghai Jiao Tong University's AI Future Fund, and released a technical preview of its Noe-0 world model in August 2026.
- Lu's core argument: what limits embodied AI is not how many hours of robot data exist, but how diverse that data is and how fast a model can absorb it.
The name came before the business
Lu did not arrive at robotics through industry. His published record runs through Hong Kong, Beijing and the United States: a bachelor's at Chongqing University of Posts and Telecommunications (重庆邮电大学), a master's at the Institute of Electronics of the Chinese Academy of Sciences, a doctorate at the Chinese University of Hong Kong, postdoctoral (博士后) stints at Hong Kong University of Science and Technology and at Stanford's AI Lab under Fei-Fei Li and Leo Guibas, and a professorship at Shanghai Jiao Tong from 2021.
His university page lists more than 100 papers as first or corresponding author in Nature, Nature Machine Intelligence, TPAMI and CVPR, along with the IROS Best Paper award and the Wu Wenjun Natural Science Award. Two open-source projects from his lab became standard tools: AlphaPose for human pose estimation, whose GitHub repository had passed 8,600 stars when checked for this profile, and GraspNet, whose GraspNet-1Billion benchmark holds 97,280 images of 190 cluttered scenes with more than 1.1 billion annotated grasp poses.
In 2023 he and Wang He wrote the definition of embodied intelligence (具身智能) that the China Computer Federation published — the first time the field had a settled Chinese term, which mattered once the phrase started appearing in industrial policy documents. In 2026 the ACM China Distinguished Researcher award came his way, presented at the Chinese Turing conference.
A lab that became a company
Noematrix was incubated out of Flexiv, the force-control robotics company, in November 2023, with Lu as co-founder and chief executive. The funding sequence has been fast: an angel plus Pre-A round in 2024 (announced as several hundred million yuan combined, roughly US$30M–US$125M depending on where in that band it landed), a Pre-A+ round led by Sequoia China in December 2024, Pre-A++ in April 2025, an investment from Alibaba in October 2025, a Series A led by C Capital with Sea Limited and Puhua Capital in February 2026, and another several-hundred-million-yuan round in June 2026 from Shanghai Jiao Tong University's AI Future Fund, the Shanghai Chuangzhi Institute's wholly owned vehicle, the Wuxi Data Group and Yicun Capital. The company lists Forbes Asia's companies-to-watch among its honours.
In August 2026 Noematrix published a technical preview of Noe-0, a world model built on what it calls full-chain ontology-free data (全链路无本体数据) — a term for collecting without hand-labelling objects into fixed categories first.
The bet: data slope, not data hours
The industry's default metric for embodied AI is hours of robot data. Lu has argued publicly, in an interview with Huxiu, that raw hours are a crude metric: a million hours of one fixed workstation carries less information than far less data covering many situations. What matters, in his framing, is the slope — how fast a model converts new experience into reliable behaviour.
That argument maps onto how Noematrix collects. Its RoboPocket system pairs a wearable end-effector with a phone app so ordinary people can record manipulation tasks without standing near a robot; the collection network now spans more than 50 cities and has peaked at roughly 1,000 simultaneous collectors, according to the company's own figures presented at a 2026 Alibaba Cloud conference. The company's stated principle is that real data and simulated data complement rather than replace each other, with real scenes supplying stability and simulation supplying scale.
Lu is also sceptical of explicit world reconstruction. In the same interview he argued that what matters is a model's ability to predict and act robustly, not whether it builds a high-fidelity 3D model of the room.
Where the money has to work
The commercial test is unglamorous. Noematrix says its robots are deployed in chain pharmacies, including sites in Shenyang, where the "embedded upgrade" approach needs about 2.5 square metres and no shelf reconstruction, and it has shown multi-robot laundry handling in hotels.
Lu also pushes a definition of commercialization: the share of tasks a system can perform at 99% robustness, and the post-training cost of getting there. That is a deliberately unglamorous bar, and it is the one he has set his own company against.
Honest limitations
This profile rests on official sources: Shanghai Jiao Tong University's faculty and leadership pages and its news office, the China Computer Federation's terminology release, Noematrix's own announcement archive, and reporting from The Paper, Xinhua's Shanghai channel and Xinmin Evening News. The revenue, scale and deployment figures quoted here are company statements repeated by those outlets, not audited disclosures — Noematrix is private, publishes no financial statements, and I could not independently verify the collection-network size, the number of cities, the 2.5-square-metre footprint or the pharmacy deployment counts. I have therefore reported them as claims with attribution rather than as established fact, and I have not attempted to convert the vague "several hundred million yuan" round sizes into single point estimates. I did not find a verified primary transcript of Lu's public remarks, so his positions here are paraphrased from a Huxiu interview rather than quoted verbatim, and I have kept my own reading separate from his. The Reuters-style rule of double-sourcing a number applies throughout: where I could only find one source, I dropped the number.
What readers can do now
- Judge embodied-AI companies on how fast they can turn collected data into reliable behaviour, not on dataset size in hours.
- Watch Noe-0 in production settings such as pharmacy logistics; deployment evidence in unglamorous scenes matters more than demo videos.
- If you are building on Chinese embodied-AI stacks, compare vendors by breadth of hardware support and by measured cost per completed task, not by benchmark leaderboards.