Maxinsights, a physical AI company headquartered in Silicon Valley with its Chinese entity in Shanghai, has closed a USD 30 million seed round led by Peak XV, with Tola Capital, South Park Commons, Yingce Capital and Alumni Ventures participating. None of the money goes to compute or robot hardware. All of it goes to the least glamorous input in the industry: real-world training data.
The two founders came out of Shanghai Jiao Tong University and Peking University. Chief executive Yuan Yuguang co-founded an industrial robotics company before a management stint at Eaton. Chief technology officer Wang Kun worked as an engineer on Google’s self-driving programme, then returned to China to co-found the autonomous driving startup Qingzhou Zhihang, and saw that technology through to deployment at the million-vehicle scale.
The missing textbook
The internet gave language models an ocean of text. It never gave robots a corpus of humans washing dishes, prepping vegetables or tidying a room. There is no ready-made textbook of physical action for a machine to read.
Maxinsights went to the kitchens instead. It recruited more than 5,000 contract workers across six countries, who wear the company’s head, wrist and chest cameras and record first-person footage of their own daily work in home kitchens, restaurant back kitchens and hotel rooms. The company has now collected more than two million hours of video, of which 1.5 million hours are directly usable for model training. More than 30 frontier AI labs and robotics companies are paying customers.
Quality control as engineering, not labour
Much of the data collection sector is still competing on hours and headcount. Vendors advertise total recorded hours, the footage contains staged and unusable segments, and a large manual team is needed afterwards to clean it.
Maxinsights separates on the cleaning step. Its MaxLoop agent-in-the-loop quality system evaluates and feeds back on incoming data, compressing a review cycle that used to take one to two weeks into under a day. The human review team shrank by 90 per cent and review cost fell by 80 per cent.
Vision alone has a known blind spot. A camera cannot capture grip force or the deformation of a slipping contact surface, and those signals are exactly what a dexterous hand needs for fine manipulation. Maxinsights is developing a sensor glove called MaxGlove with Genesis AI, targeting one million hours of force data by 2027.
Peak XV’s investment lead argued that turning quality control into an engineering capability is the part later entrants cannot easily copy, and that this is why many labs chose the company after comparing suppliers.
What the round actually prices
The strategic signal is larger than one startup. Discussion of physical AI has centred on compute, model architecture and humanoid hardware specification, on the assumption that compute is the binding constraint.
At this stage of the cycle, high-quality real physical interaction data is scarcer than compute. A compute gap can be closed by building fabs and buying cards. Ten million hours of genuine household and service interaction cannot be generated by purchasing GPUs.
That changes the yardstick. Comparing total recorded hours is ending. Effective hours, meaning footage that survives quality control and carries the right sensing modalities, is what determines whether a dataset has training value. A million hours passed through automated review and multimodal capture is worth far more than a larger pile of low-grade recording.
It also implies a division of labour. A robotics company that builds its own collection chain carries hardware development, workforce organisation and quality auditing at the same time, and dilutes the team that should be building the model and the body. A mature third-party data layer absorbs that heavy work.
The constraints have not disappeared. Scene diversity, privacy compliance and the mass production of force-sensing gloves are all unsolved at scale.
Industrialisation, in the end, is a data problem before it is an intelligence problem. The robots that reach a kitchen will have watched thousands of them first.
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-8321205-12005-30709479.html.