Executive Summary
The second World Humanoid Robot Games opened in Beijing on 22 August 2026 and run to the 26th. The field is 666 teams and 2,056 robots from 16 countries — up from 280 teams and roughly 500 robots at the inaugural event twelve months ago. Chinese entrants account for 641 of the teams and 1,975 of the machines, about 96 percent of the field.
Read as sport, this is a curiosity. Read as an industrial document, it is more informative than most corporate disclosure in the sector. The organisers publish which events require full autonomy and which permit a human at the controls — an explicit, public statement of which capabilities are considered solved and which are not. They also publish 21 work-scenario events covering factories, hotels, logistics, retail and a live fire brigade, which amounts to a specification of the jobs China expects these machines to do.
The analytical point. This report treats the Games as evidence about industrial strategy rather than about robots. On that reading, the event is not primarily a demonstration of capability. It is an attempt to solve the problem this publication identified in June: China built humanoid manufacturing capacity faster than it built demand for the output. The Games are, among other things, demand creation.
Note on timing: the Games conclude on 26 August 2026. Results cited here are those announced during the competition and are necessarily partial. Final medal tables were not published at the time of writing.
What Is Actually Happening in Beijing
The venue is the National Speed Skating Oval, the “Ice Ribbon” built for the 2022 Winter Olympics. The programme comprises 51 events: 30 competitive and 21 scenario-based. Organisers include China Media Group, the Beijing Municipal People’s Government and the World Robot Cooperation Organization.
Among results announced during the week, Unitree Robotics took gold in the 1,500 metres, 400 metres, 4×100 metre relay and 100-metre hurdles. X-Humanoid’s Tiangong Ultra recorded 9.39 seconds over 100 metres and set a standing high jump mark of 2.88 metres. A long jump of 7.97 metres won that event’s final. For scale, the winning 100-metre time in 2025 was 21.50 seconds; the 2026 mark is roughly 2.3 times faster in a single year.
A comparison worth handling carefully. Coverage has widely noted that 9.39 seconds is faster than Usain Bolt’s 9.58-second world record. The two numbers are not comparable in any meaningful sense: a machine engineered for a single repeatable task on a prepared surface is not competing under the constraints that make the human record meaningful. Separately — and often conflated in reporting — a robot named Lightning won the Beijing E-Town Half-Marathon in 50 minutes 26 seconds in April 2026, a different event some four months earlier, during which it struck a barrier, fell, and was helped upright by humans. The year-on-year improvement is the real signal. The Bolt comparison is not.
The Most Useful Document Is the Rulebook
Humanoid robotics has a persistent credibility problem: many public demonstrations are teleoperated, and the distinction is not always disclosed. Teleoperation is a legitimate engineering tool. It becomes a problem when a machine is presented as autonomous labour while a human off-camera supplies the judgment.
The Games handle this in an unusually transparent way. Certain events require full autonomy: flat-ground running races, football, gymnastics, dancing, boxing and Tai Chi, table tennis. Others permit teleoperation: hurdles, jumping events, weightlifting, kickboxing and the pentathlon. In the 21 work-scenario events, an autonomously completed run scores full points; if a human intervenes at any stage, the run is classified as teleoperated and its score weighted at 0.5.
That division is a capability disclosure. It says closed-loop perception at speed on flat ground is considered solved, and that explosive dynamics with a hard landing, heavy load-bearing and adversarial physical contact are not. No vendor press release conveys that. An engineer reading only the rulebook learns more about the state of the field than one reading a year of launch videos.
The scenario events are equally legible as a market map. They cover retail, food service, assembly, warehousing, hospitality and firefighting, and the dexterous-hand category includes power-tool assembly, bricklaying, nail fastening, picking beans with tweezers and connecting cables. That is a list of target occupations, published a year or more before the machines can reliably perform them.
The Production Ledger
Morgan Stanley began 2026 forecasting 14,000 Chinese humanoid shipments for the year. In January it raised that to 28,000. In June it raised it again, to 50,000. The same bank projects 446,000 units annually by 2030 and a Chinese market growing from roughly $2 billion this year to $15 billion by then. China’s Ministry of Industry and Information Technology has set a national target of 100,000 units for 2026.
Underneath the forecasts, the company-level picture is genuinely difficult to compare, because firms disclose different things — cumulative production, monthly capacity, annual targets, or units deployed. The table below preserves those distinctions rather than flattening them into a single misleading number.
| Company | Base | Reported figure | What kind of number |
|---|---|---|---|
| Unitree | China | ~12,500 produced; 190,000/yr capacity goal | Cumulative + future capacity |
| BYD | China | 20,000 in 2026, from 1,500 in 2025 | Company target |
| AgiBot | China | ~15,000 | Cumulative |
| UBTech | China | ~150/month; 3,000 target for 2026 | Run rate + target |
| Tesla (Optimus) | US | 50,000–100,000 stated target; internal use only | Target, not shipments |
| Figure AI | US | 10,000+ deployments; 100,000 over four years | Deployments + plan |
Two observations follow. First, Chinese firms accounted for more than 80 percent of global humanoid installations in 2025, and nothing in the 2026 data suggests that share has materially reversed. Second, and more important for anyone trying to keep score: Tesla’s headline number is a production target for machines it says will be used internally rather than sold, which is a different category of claim from Unitree’s cumulative output or Figure’s deployment count. Comparing them directly is the most common error in coverage of this race.
Where the United States Is Ahead
A report that counted only units would mislead. On several dimensions the American position is strong, and on one it is dominant.
Capital. Figure AI raised more than $1 billion in a Series C at a reported $39 billion post-money valuation, with over $1.9 billion in total funding. Apptronik closed $935 million at more than $5.5 billion. These are extraordinary sums for companies with limited commercial revenue, and they buy years of runway independent of near-term demand.
Frontier models. The hardest unsolved problem in humanoid robotics is not locomotion; it is general manipulation and reasoning about unfamiliar physical situations. That problem is closer to frontier AI than to mechatronics, and the leading laboratories remain overwhelmingly American. This publication’s earlier analysis of the open-weight gap argued that America’s disadvantage lies in diffusion rather than capability; humanoids are the clearest physical instance of that pattern.
What the US lacks is coordination. A coalition including Tesla, Boston Dynamics and Apptronik has publicly urged Washington to adopt a national robotics strategy, including a dedicated federal office, arguing that China has made intelligent robots a pillar of national planning while the United States has no equivalent. That the industry is asking for it is itself evidence of the asymmetry.
The Chokepoint Under Everything
A humanoid robot contains dozens of precision actuators, and high-torque actuators depend on neodymium-iron-boron permanent magnets, which depend on neodymium, praseodymium, dysprosium and terbium. China holds roughly 63 percent of humanoid component supply and a far larger share of rare-earth processing specifically.
Reporting indicates that the China Rare Earth Industry Association and the Beijing Humanoid Robot Innovation Centre have met to coordinate on robotic joint modules and high-torque magnet motors — deliberate alignment of an upstream chokepoint with a downstream industrial target. Some American manufacturers have begun engineering to reduce Chinese content, which raises cost and lengthens timelines.
This is the same structure this publication examined in its analysis of the FCC’s Covered List and market access for robots: the contest is being fought less over who can build the best machine than over who controls the inputs and the terms of sale.
The Demand Problem the Games Are Designed to Solve
In June this publication argued that China had won the humanoid factory race and acquired a demand problem: manufacturing capacity was scaling faster than any identified buyer base. That analysis holds up, and the Games are best understood as a response to it.
Consider what the event actually produces. It generates a national spectacle that makes humanoids familiar to consumers and procurement officers. It publishes a list of 21 workplace tasks, effectively telling domestic industry what to start specifying. It creates a competitive ranking that gives municipal governments and state-owned enterprises a defensible basis for purchasing decisions. And it supplies a stream of benchmark results that private capital can underwrite — Unitree listed on Shanghai’s STAR Market in 2026, raising roughly 6.1 billion yuan and rising sharply on debut.
Whether that succeeds is a separate question. Manufacturing 50,000 units is not the same as finding 50,000 economically productive positions for them, and the gap between capacity and paying demand remains the central risk in the Chinese humanoid sector.
What the Games Do Not Prove
Participation is not capability, and 96 percent is not a world ranking. The Games are held in Beijing, organised by Chinese state and municipal bodies, and scheduled around a domestic industry. A Chinese share of 96 percent of entrants reflects proximity, cost and policy alignment at least as much as technical superiority. Notably, most leading American humanoid firms did not enter — an absence that says little about their machines.
Athletic events are weak proxies for economic value. Sprinting on a prepared track is a narrow, repeatable, well-instrumented task. Unstructured manipulation in a cluttered environment is the capability that determines commercial usefulness, and it is precisely the one the organisers still permit humans to assist with.
The unit numbers are not audited. Company production figures in this sector are self-reported, definitions vary, and independent verification is largely unavailable. The forecast revisions cited here are a bank’s estimates, not observations, and that bank has revised twice in six months — which should be read as a signal about uncertainty as much as about growth.
Results are incomplete. The Games conclude on 26 August 2026. Medal totals, scenario-event outcomes and the crucial question of how many scenario runs were completed autonomously rather than at 0.5 weighting were not available at the time of writing. That last figure is the single most informative statistic the event will produce.
What to Watch
The autonomous share of scenario events. When final scoring is published, the proportion of work-task runs completed without human intervention is the closest thing to an objective capability benchmark the industry has produced. If that share is low, the manufacturing numbers are running well ahead of usefulness.
Whether Morgan Stanley revises a third time. Two doublings in six months suggests the analysts are tracking announcements rather than deliveries. A downward revision would be more informative than another increase.
Whether Tesla ships Optimus externally. Optimus units built in 2026 are, by the company’s own account, for internal use. The moment they are sold to third parties, the American position changes from capital and models to capital, models and volume.
Rare-earth export policy. If magnet supply becomes an explicit instrument of trade policy for robotics as it has for other sectors, cost curves in Western humanoid manufacturing move immediately, and the unit-count gap widens for reasons that have nothing to do with engineering.
Conclusion. The Beijing Games are not a measurement of who is winning a robotics race, and treating them as one would be a mistake in both directions — they neither prove Chinese superiority nor deserve dismissal as theatre. What they are is an unusually candid industrial document: a published boundary between what humanoids can do alone and what they still cannot, attached to a published list of the jobs a state intends them to fill, staged for an audience of domestic buyers. The United States retains the capital and the frontier models. China has the volume, the components and now a national mechanism for manufacturing demand. Those are different advantages, and on current evidence neither side has the other’s.
Sources
World Humanoid Robot Games, 2025 and 2026 editions — participation, venue, event structure and announced results. https://en.wikipedia.org/wiki/World_Humanoid_Robot_Games
Koetsier, J. The World Humanoid Robot Games Events Are A Market Map. Forbes, 19 August 2026 — autonomy requirements, teleoperation rules and scenario-event scoring. https://www.forbes.com/sites/johnkoetsier/2026/08/19/the-world-humanoid-robot-games-events-are-a-market-map/
CNBC and South China Morning Post, June 2026 — Morgan Stanley raising its 2026 Chinese humanoid shipment forecast to 50,000 units. https://www.cnbc.com/2026/06/24/morgan-stanley-china-humanoid-robot-market-forecast.html
Rest of World and Shanghai Stock Exchange — Unitree Robotics’ 2026 STAR Market listing and stated capacity plans. https://restofworld.org/2026/unitree-china-humanoid-robot-shanghai-ipo/
The Robot Report and Forge Global — Figure AI Series C, valuation and deployment figures. https://www.therobotreport.com/figure-ai-raises-1b-in-series-c-funding-toward-humanoid-robot-development/
Koetsier, J. Making Humanoid Robots Without A Chinese Supply Chain. Forbes, June 2026 — component dependence and reshoring efforts. https://www.forbes.com/sites/johnkoetsier/2026/06/17/why-some-us-robot-makers-are-reducing-dependence-on-china/
Reporting on the US robotics industry coalition’s call for a national strategy, including Tesla, Boston Dynamics and Apptronik. https://www.scmp.com/tech/policy/article/3304088/us-robotics-firms-push-national-strategy-compete-china
Smithsonian Magazine and NBC News, April 2026 — the Beijing E-Town Half-Marathon result, cited here to distinguish it from the August Games. https://www.smithsonianmag.com/smart-news/a-humanoid-robot-just-beat-the-human-world-record-for-the-fastest-half-marathon-during-a-race-in-china-180988599/
