In China’s embodied AI sector, 20 billion yuan has become the new threshold for admission to the top tier. At least five companies, ZhiPingFang, ZiBianLiang, XingHaiTu, QianXun Intelligence and Galaxea, have crossed that valuation line. The number is striking because private-market funding across China fell 32.65 per cent in the first half of 2026, and government-backed limited partners cut their commitments by roughly half.
The article, published by Guangzhui Intelligence, argues that investors are paying for brains, not bodies. The five companies have relatively low hardware shipment volumes compared with Unitree, which priced its IPO at a 60.9 billion yuan valuation, or Zhiyuan, which is targeting 36.3 to 43.3 billion yuan. Yet the 20 billion yuan club is valued at 35 to 40 times annual sales, implying each needs roughly 500 million yuan of revenue to avoid looking expensive.
Five companies, five answers to the same question
The technical approaches differ. ZhiPingFang and XingHaiTu are strengthening the vision-language-action, or VLA, architecture. ZhiPingFang’s FiS-VLA uses a fast-and-slow dual system: a slow module handles instruction parsing and high-level reasoning, while a fast module generates actions. It later evolved into NeuroVLA, adding a cerebellar module for stability.
XingHaiTu’s G0.5 removes the encoders that usually translate between vision, language and action, instead turning all three into a single internal token sequence. QianXun Intelligence bets on pure scaling of VLA, using large volumes of unfiltered “dirty” data to improve generalisation across scenes.
ZiBianLiang takes a different path with WALL-B, which jointly trains physical prediction, vision, language and action so the robot can forecast outcomes before it moves. Galaxea, the most architecture-heavy of the group, stacks multiple VLA models for upper-body grasping, lower-body movement and human interaction, now unified under a base called AstraBrain.
The real battle is over data
The article divides data strategies into two camps. Real-world data is scarce and costly but transfers cleanly to deployment. Simulation data is cheap and scalable but suffers from the sim-to-real gap. XingHaiTu and ZhiPingFang favour deployment feedback from factory floors, where robots perform inspection, assembly and logistics tasks. ZiBianLiang focuses on the harder home environment, with its messy lighting, variable layouts and human movement. Galaxea built its GraspVLA on ten billion simulated grasp frames, but is now adding real-robot teleoperation and deployment data.
The commercial logic points toward industry first. Galaxea and QianXun already deploy robots in CATL factories. ZhiPingFang signed a three-year, 1,000-unit order with HKC, a semiconductor display maker, last year. Retail is the next target. Galaxea’s Galbot G1 has served thousands of cups of coffee at a Beijing forum with a 99.97 per cent task success rate.
Home deployment remains the most distant. The article notes that even an 80 per cent success rate is not good enough when the remaining 20 per cent of failures create more work for humans than they save. For now, 20 billion yuan buys a credible data engine and a plausible path to factory revenue, not a household robot.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience.