NextFin News -- The focus of competition in the humanoid-robot industry is shifting from athletic performance to job execution in real-world settings. What will determine where a company ultimately lands is no longer how fast it can run or how high it can jump in an arena, but whether it can reliably complete tasks in factories, hotels, homes, and other environments—and accumulate enough real operational data to support model training. Investment aimed at data acquisition and standards building is becoming the key variable reshaping the competitive landscape.

A telling example emerged in late August. On August 26, the Second World Humanoid Robot Games concluded at the National Speed Skating Oval in Beijing. A total of 666 teams and 2,056 robots took part; Tiangong Ultra finished the 100-meter sprint in 8.64 seconds, and its standing high jump reached 3.40 meters. The results drew plenty of attention, but the industrial significance of the event was not the scores themselves.

On August 28, at a press briefing held by the National Development and Reform Commission, Li Chao—Deputy Director of the Policy Research Office and the commission’s spokesperson—set boundaries for the industry’s development. The robotics sector must guard against blind bandwagoning and a rushed, herd-like pile-in.

Hosting a competition tests the upper limits of athletic capability; setting the tone constrains the pace of industrial expansion. The two moves do not target the same thing. The industry is in the middle of a shift in evaluation standards: athletic performance is becoming less important, while task execution in real scenarios—and the accumulation of data—are starting to become the primary yardsticks for valuing companies.

The Real Center of Gravity Beyond the Arena

Alongside the record-chasing results, another release drew far less attention. At the closing ceremony, the CCID Research Institute and the Beijing Aoyun Group, together with companies including AgiBot, Galbot, Wan Jing Qian Xun, and Xinghaitu, released a real-world operational dataset collected during training and competition. With a total duration of more than 2,500 hours, it covers 12 categories of scenarios—industrial, retail, food service, home, firefighting and rescue, hotels, and more—spanning 44 types of tasks and over 100 skills, and is being made freely available to the public.

This data is fundamentally different from competition scores: it records the operating processes robots went through while performing everyday jobs such as folding quilts, restocking shelves, and loading materials. What the industry is racing to secure is precisely this kind of real-scenario data that captures operator judgment.

This move is not unique to China. Major companies worldwide are all investing resources to obtain data, though they are taking different routes. According to media reports, Tesla has set up a dedicated teleoperation team for Optimus: operators use motion-capture and VR equipment to simulate the robot performing assigned actions, with requirements even extending to parameters such as height. As for Figure AI’s partnership with BMW, the company has officially characterized it as an experimental deployment and a data-collection effort.

"When the pricing of a service can’t even cover its marginal cost, that price is no longer a market price—it’s an investment."

This assessment came from reporting on Tau Robotics, a robot cleaning company in San Francisco. The company prices its humanoid robot cleaning service at $30 per hour—below the going rate for human cleaners locally—but at this stage it still requires human operators behind the scenes to assist with control. After factoring in operators’ hourly wages and robot depreciation, costs and fees essentially break even. The founder admitted that AI still struggles to handle the complexity of home environments; human involvement is, on the one hand, to keep the service running, and on the other, to accumulate movement data for the AI.

From Corporate Experimentation to Top-level Design

A similar approach can be seen with 1X’s NEO home robot. The company calls the current stage “Expert Mode,” meaning an operator provides real-time assistance while the AI gradually builds up experience during the process.

Comparable practices have also emerged in China. Reports indicate that a job titled “embodied intelligence data collector” has already appeared, with practitioners wearing VR or exoskeleton devices and repeatedly performing actions such as folding clothes and grasping objects. Other reports mention that JD.com’s home services business has cleaners wear data-capture devices while on the job, turning each housekeeping visit into a synchronized data record.

In real-world settings, operational data that incorporates human judgment is scarce enough to justify sustained corporate investment—even without pursuing immediate returns. Once this logic expands from a choice made by a handful of companies to the industry and policy level, its impact is no longer confined to individual firms.

The sports-meet dataset is a clear signal of this escalation. The publisher, CCID Research Institute (full name: China Center for Information Industry Development), is a public institution directly under the Ministry of Industry and Information Technology. The decision to make the dataset freely available was led by local governments and research bodies with official backing, rather than being an independent move by any single participating company. What began as company-level experimentation—trading investment for data and accepting costs in exchange for real-world scenarios—has been consolidated by official actors into a public, industry-wide piece of foundational infrastructure.

These policy moves were not isolated cases. On February 28, the Ministry of Industry and Information Technology’s Standardization Technical Committee for Humanoid Robots and Embodied Intelligence released the Humanoid Robots and Embodied Intelligence Standards System (2026 Edition). Jointly compiled by multiple research institutes and companies, it was the country’s first top-level standard framework covering the entire industry chain and the full product lifecycle. At the launch event, the committee’s vice chairman and Unitree Robotics founder Wang Xingxing delivered a speech titled “From ‘Kung Fu Mode’ to ‘Work Mode,’” arguing that humanoid robots should not remain mere performance props, but should become tools capable of taking on real work.

On June 1, the Embodied Intelligence Benchmark Testing Method officially took effect, marking the first time the industry established a unified approach—across both simulated and real-world environments—for task-set construction, testing, and metrics calculation.

Different players were pushing in the same direction over the same period. Mobility was no longer the main battleground; scenario capability took its place. This shift had previously been an optional move chosen by individual companies. Now it has been folded into the industry’s design framework—and has begun to influence how the capital markets make their judgments.

The Temperature Gap Between Primary and Secondary Markets

Moves and judgments in the capital markets were broadly in sync. According to ITJuzi data, in the first half of 2026, total financing in China’s embodied-intelligence track reached roughly RMB 93.5 billion—about five times the level of the same period a year earlier—with 322 disclosed funding deals. As of June 12, financing in the track totaled about RMB 43.8 billion, with more than half flowing to “brain-focused” companies that emphasize algorithms and do not engage in hardware manufacturing; “body-focused” companies accounted for the smallest share of financing among all categories. The two sets of statistics use different methodologies, so the figures are not directly comparable, but the direction implied by the flow of funds was broadly consistent.

Beyond the purely algorithm-driven “brain-focused” route, companies such as Galaxea General and Stardance Era, which emphasize full-stack in-house R&D, also drew capital’s attention. With both routes attracting investment, the market was signaling that intelligence capability is the decisive factor—but where algorithmic capability will ultimately settle within the stack remains an open question.

No matter which technical path they bet on, companies still face the same test: can the robots actually be deployed for real operations and generate returns? In the primary market, multiple funding rounds can help cushion valuation pressure; in the secondary market, the test comes faster and more directly.

On August 19, that test arrived earlier than expected. Unitree Robotics, billed as the first humanoid-robot stock, listed on Shanghai’s STAR Market. Its competitiveness came from mass-production capability for complete machines, rather than an algorithm narrative. Even so, the market reaction was intense. The offering price was RMB 150.8, the opening price RMB 1,100—up 629.44%—and its market capitalization once reached RMB 444.9 billion. According to Wind data, as of August 17, the average first-day gain for the 94 new A-share listings so far this year was 279.22%; Unitree’s opening jump was far above that level.

Within four trading days after the listing, the share price pulled back sharply. It closed at 845 yuan on the first day; fell 18.70% on the second trading day; dropped 2.12% on the third; and slid another 10.31% on the fourth, ending at 603 yuan. Market capitalization shrank to 243.9 billion yuan, wiping out nearly 200 billion yuan from its intraday peak.

In the view of Hong Hao, managing partner and CIO of Lotus Asset Management, humanoid robots have not yet truly been deployed in practical scenarios such as household chores or industrial use, and their commercial viability remains to be proven. This assessment is not aimed at any particular technical route, but at whether the industry has genuinely completed this round of transition. Data are being accumulated, standards are being established, and capital is being invested; once these conditions are in place, whether commercial practicality can keep pace is a question the industry cannot avoid.

The NDRC’s remarks about preventing blind bandwagoning and “everyone rushing in at once” came between the opening up of datasets and the pullback in Unitree’s share price—highlighting the pace of industrial development. Preparations in data, standards, and capital are already in place; the industry’s next task is to turn scenario capabilities into real returns.