A carmaker, a search company and a humanoid maker backing one data supplier suggests robot developers expect to buy training data from shared vendors instead of collecting it themselves.
GenRobot AI, a Beijing company that turns recordings of people doing physical tasks into training data for robots, has raised US$100 million. GenRobot AI has completed a US$100 million Series A+ round backed by SAIC Motor, Baidu, Malaysia’s Eastworlds, Shoucheng Holdings and Galaxy General Robotics. The financing closed in the third quarter and was disclosed on November 2, TechNode reported, citing Sina Finance.
Financial investors GL Ventures, CITIC Goldstone, CCB International and Huatai Zijin also took part. TechNode did not report a valuation or name a lead investor.
The round puts a large sum behind one of the least visible parts of China’s robot industry. Humanoid and other general-purpose robots learn tasks from examples, and good examples of humans folding laundry, packing boxes or sorting parts are scarce and expensive to collect.
GenRobot sells the recording rigs and the cleaned data
GenRobot makes equipment and software for capturing human demonstrations, according to TechNode. Its recording systems capture what a person sees together with hand and body movements across household, manufacturing and logistics tasks. Its Matrix platform then synchronizes, cleans and labels the recordings so robot developers can use them for training and evaluation.
The company described its hardware in more detail in June, when it announced earlier rounds totaling hundreds of millions of yuan led by Ant Group, Didi and Delian Capital, Gasgoo reported. Its kit includes a first-person vision module that tracks head movement, a bionic hand module that captures grasping, pinching and twisting, and tactile sensors. A multi-camera array uses six 200-megapixel cameras, and other sensors record sound, touch, fabric strain and force feedback, with multi-device synchronization delays held under one millisecond.
At that point, GenRobot said it had collected more than 1 million hours of real-world data from more than 3,000 contributors in over 10,000 settings, including homes, factories, logistics hubs, labs and medical facilities, covering more than 2,000 everyday skills. It said its first-person and hand-capture system produced more than 100,000 hours of high-precision data a month and that it worked with more than 30 AI companies in China and abroad. Existing investors at the time included Shunwei Capital and Baidu Ventures.
The investor list maps the buyers
The strategic investors read like a customer list. SAIC is a carmaker with factories to automate; Baidu develops AI models and has backed GenRobot since earlier rounds through Baidu Ventures; and Galaxy General Robotics, known as Galbot, builds humanoid robots that need exactly the kind of manipulation data GenRobot records. EastFrontier has covered Galbot’s push into retail through Hong Kong.
GenRobot’s approach is often described as "body-agnostic" data: recordings of humans rather than of a particular robot, which can in principle train many different machines. Gasgoo described the June financing as the largest so far in that segment. The bet is that one supplier recording human skills at scale can sell to many robot makers, instead of each one running its own data operation.
That model depends on the recordings being useful across different robot bodies. Human hands and robot grippers differ, and data captured from people has to be translated into instructions a specific machine can follow. GenRobot has not published evidence of how much its data improves customers’ robots.
Customer results will decide the data market
The company has not said how it will spend the new money. In June, it named three priorities: expanding into more kinds of human behavior data, building a model that improves the precision and consistency of its data, and connecting data production, training and evaluation into one closed loop. It also said it planned to expand abroad.
The next signal will come from GenRobot’s customers. Robot makers that report specific gains, such as higher task success rates after training on its recordings, would support the case for a shared data supplier. A move by Galbot or SAIC to sign long-term data contracts would show the strategic investors are also paying customers.