The First Global Benchmark for Legged Robots: Why China, Why Now

By Embodied AI Frontier

In the first half of 2026, the global humanoid robot market was described simultaneously by three numbers. Smart Analytics Global (SAG), in a report relayed by Bloomberg, put worldwide shipments at around 19,100 units, up roughly 272% year over year. Counterpoint Research, in an independent estimate, arrived at about 22,000 units, up nearly 300%. At the World Robot Conference (WRC) 2026, the Chinese Institute of Electronics (CIE) and the China Humanoid Robotics 100-People Committee, two industry organizations, released the "2026 Humanoid Robot Industry Development Report" claiming that China alone shipped more than 40,000 units in the half, or 97% of the global total.

The arithmetic does not add up. If 97% of global shipments were in China and China shipped more than 40,000 units, the world total would be roughly 41,000 units, more than double the SAG figure. The most likely explanation lies in the definitions. Some people count legs, meaning bipeds and quadruped robot dogs together; some count humanoids, meaning bipeds only. Some count shipments; some count production. Some count the world; some count China. More fundamental still: the International Federation of Robotics (IFR) does not maintain a separate humanoid robot category, so every "humanoid shipment" number comes from a market research firm's own definition.

Into this confusion, on July 24, 2026, a new ruler was erected. ISO 18646-5:2026, "Robotics: Performance criteria and related test methods for service robots, Part 5: Locomotion for legged robots", was published at stage 60.60 and announced through Chinese official channels on September 1, 2026. It was led by Zhejiang Lab, a public research institute backed by the Zhejiang provincial government, on behalf of China, together with experts from eight countries: the United States, the United Kingdom, France, Germany, Spain, Japan, South Korea and India. The proposal was approved for development with zero negative votes in the 12-week ballot. This article asks why the standard chose legs, why it arrived now, and why China led it.

1. Why Study Legs at All?

1.1 Bipedal Is the Hardest Branch of the Legged Family

Legged robots look like one family, but difficulty is not evenly distributed within the family. Quadruped robot dogs are already fairly stable on flat ground; bipedal humanoids are still fighting to stay upright. An August 2026 report in the China Business Journal cites an industry judgment that the difficulty of a bipedal system is double that of a wheeled platform. Bipeds require high-torque integrated joints at hip, knee and ankle, plus high-precision foot pressure sensors. Gait jitter and center-of-mass shifts move the base of the arm directly, so grasping and assembly precision must be recalibrated again and again. The hard part of a biped is never walking; it is working while walking.

1.2 But the World Is Built for Two Feet and Two Hands

The case for two legs is just as strong. Door handle heights, stair tread spans, and the dimensions of kitchen counters and tool cabinets are all designed around human body dimensions and human movement patterns. Boston Dynamics states on its Atlas product page that the goal is not to make the workspace adapt to the robot, but to let the robot use the same equipment and workstations that employees already use. Chinese industry outlet NE Times made the same judgment in October 2025: the overwhelming majority of real-world facilities are built around the human body, so humanoid robots can be integrated without large-scale renovation.

1.3 The AI-Era Logic: Physical-World Data Is the Next Bottleneck

Within the 2026 embodied AI narrative, the answer to "what is missing" is shifting from compute and models to data. Peng Zhihui, co-founder and CTO of AGIBOT, said in April 2026 that money was not the constraint; data was. Wang Xingxing, founder and CEO of Unitree, gave a similar reason for moving from quadrupeds to bipeds: humanoid robots are better suited to data collection, and quadruped technology can be reused. Xinhua News Agency has reported a telling detail: every time a robot falls, the back end receives more than a thousand sets of abnormal data, covering joint torque allocation and stride frequency adjustment. For a team training an end-to-end model, one fall is one high-quality sample. When data becomes the bottleneck, the form factor that gathers more physical-world data acquires a new source of value.

1.4 But Legs Are Already Making Money

While bipeds have not yet proven their profit model, the other branch of the legged family has already validated the business. By the numbers in Unitree's STAR Market IPO filing, the STAR Market being China's NASDAQ-style board for technology firms, its quadruped shipments totaled 2,403 in 2022, 3,121 in 2023, 7,136 in 2024 and 17,946 in the first nine months of 2025, more than 30,000 cumulatively, with a global share of roughly 60-70%. Its bipedal humanoids start from a much smaller base: more than 5,500 units shipped in 2025 and about 18,000 cumulative units off the production line through July 2026. Chinese robotics company UBTECH delivered 1,079 full-size humanoids in 2025. Quadrupeds first and bipeds later was always Unitree's route: scale the quadruped across consumer and industry scenarios, then carry the technology forward.

So the question the industry actually needs to debate is not whether the biped is the right form, but whether the ability to use legs is worth measuring. On the industry data, that answer is already clear: legs are making money, and the difference between those that earn more and those that earn less is exactly what needs a ruler to make clear.

2. Does a General-Purpose Robot Have to Be Humanoid?

2.1 What Deployment Actually Looks Like in 2026

The August 2026 China Business Journal report paints a picture at odds with the popular imagination. Judged by market demand, wheeled dual-arm robots are more widely deployed in industry than bipeds, both because industrial scenarios demand it and because mounting the robot's upper body on a wheeled base is the most pragmatic interim compromise under current hardware and algorithm constraints. AGIBOT's cumulative mass production passed 15,000 units off the line through June 2026, counting from the start of scaled production in August 2024, and its industrial deployments are mostly "semi-humanoid": a robot's upper body mounted on a fixed column. RobotEra's logistics sorting robots run in more than ten logistics centers across five provincial-level regions, and the most common configuration on the ground is also a half-body humanoid. Gao Jiyang, CEO of Galaxea, has said openly that he did not start with bipeds, because wheeled dual-arm robots suit structured scenarios; his bipedal product is prepared for unstructured ones.

2.2 The Case for the Humanoid Endgame

The pro-humanoid case is explicit. Jensen Huang, founder and CEO of Nvidia, has said publicly that the human form is the only form in which a general-purpose robot can succeed. Boston Dynamics, when relaunching the electric Atlas, emphasized that it is not humanoid for the sake of being humanoid; the humanoid appearance is a result of functional optimization. Zhao Tongyang, founder of ENGINEAI, a Shenzhen robotics startup, put it more bluntly: the lower body is the soul of the humanoid robot. Interact Analysis, the UK-based market research firm, examined more than 100 products released by over 60 commercial humanoid companies in the past three years; about 70% are bipedal. Its long-term judgment also leans bipedal, on grounds of scenario fit and technology reserves.

2.3 The Pragmatist Case

The skeptical side is equally well prepared. Zhu Xiaohu, managing partner of GSR Ventures, announced in March 2025 that he was exiting humanoid robot companies in batches, saying the customers these founders described were invented by the founders themselves, and asking who would spend more than RMB 100,000 (about $14,000) on a robot to do jobs like these. A telling detail: in June 2026 he made his fifth consecutive round of investment in an underwater robotics company, which suggests his objection is aimed at the humanoid form rather than at embodied intelligence itself. Rodney Brooks, the roboticist and iRobot co-founder, predicted in September 2025 that in 15 years the successful "humanoid" robot will be wheeled, multi-arm and fitted with specialized sensors, and will have abandoned the humanoid appearance. The most striking data come from MIT Technology Review. Figure AI, the American humanoid company, carries a post-money valuation of about $39 billion, while Goldman Sachs forecasts the entire global humanoid robot market at $38 billion by 2035. This one company sits nine years ahead of the industry's forward expectation, having shipped only about 150 units last year. The same article notes that humanoid battery life runs 90 to 120 minutes against an industrial demand of 8 to 20 hours. Wang Chenglu, CEO of Shenzhen Kaihong, an open-source HarmonyOS ecosystem company, and former head of Huawei's HarmonyOS, is blunter: the humanoid boom is 99.99% bubble.

2.4 The Observation: What Is Contested Is Not Legs

There is a mismatch at the heart of this debate. The loudest fight is humanoid versus wheeled, and yet almost nobody questions the value of legs. Wheeled chassis carry the main industrial orders; quadrupeds have built a steady, real business in power grid inspection and chemical explosion-proof operations; bipeds and quadrupeds have each won orders. Even within the consensus on legs, there is a gap between walking and working: Xiaomi's humanoid robot worked autonomously for three consecutive hours at a nut-fastening station in an electric vehicle factory, at a 90.2% task success rate, against an industrial requirement of 99.9% to 99.95%. The form debate will probably stay unresolved for a long time, but it can be dissolved at the level of measurement: stop asking whether a robot is humanoid or wheeled, and ask how high a step it can climb, how steep a slope it can handle, and how wide a gap it can cross. That is exactly the angle the national standard took.

3. What China's Standard Actually Measures: Legged, Not Just Bipedal

3.1 First, What the Standard Sets Down

Before asking why China wrote the standard, it is worth reading the standard itself. The published document is ISO 18646-5:2026, "Robotics: Performance criteria and related test methods for service robots, Part 5: Locomotion for legged robots". The "Part 5" in the name means this is not an isolated document but a member of a series. The ISO 18646 series sets out how the performance of service robots should be measured and has expanded progressively since 2016, adding wheeled locomotion, navigation, manipulation and other capabilities.

Two details matter. Part 6, on lower-limb wearable robots, was published two months before Part 5, so the accurate claim is that Part 5 is the first legged-robot locomotion performance-evaluation standard, not the first of the series' 2026 additions. Parts 7 and 8 are still under development, which shows the series keeps expanding and legged locomotion is only one piece of it.

What does Part 5 measure? The standard text defines nine test items, all of them about whether legs can walk and how well they walk: