The second World Humanoid Robot Games closed in Beijing on 26 August 2026 with 2,056 robots from 666 teams and 16 countries. A year earlier the same event drew 280 teams and roughly 500 machines. The headline is not the crowd, it is the rate of change: entries grew 138 percent in twelve months, the programme doubled from 26 events to 51, and the winning time in the 100 metres fell from 21.50 seconds to under 9.
For anyone evaluating humanoid robots for a European deployment, a sports meet in an ice rink can look like theatre. It is worth a closer read. Competition does something that vendor demos cannot: it puts every platform under the same rules, on the same day, in front of referees, with no second take. The results published this week are the closest thing the sector has to an independent benchmark, and they name the companies whose hardware is actually finished enough to be measured.
What the World Humanoid Robot Games 2026 actually was
The Games ran from 22 to 26 August at the National Speed Skating Oval in Beijing, the venue built for the 2022 Winter Olympics. The organising committee put the scale at 51 events across nine disciplines and 1,301 competition sessions, with preliminary rounds dropped so that every registered team competed. Thirty of the events were athletic, 21 were scenario based, meaning industrial and service tasks rather than sport. The judging panel grew from 123 referees to 223.
The composition of the field matters as much as its size. Of the 666 teams, 641 were Chinese, drawn from 157 companies and around 200 universities and research institutes. Foreign entries came from 16 countries in total, including the United States, Germany, Japan and Brazil, several of them RoboCup squads. Anyone still hoping that the humanoid supply base will turn out to be evenly distributed should read that entry list carefully, because it is the same picture our ranking of Chinese humanoid platforms has been tracking all year.
Worth knowing: the opening ceremony was staged by the competitors. Eighty Booster Robotics T2 units performed together, 36 Linkerbot machines played instruments, and 25 wheeled Galbot G1 robots worked the floor.
The scoreboard names the companies to watch
AGIBOT entered the Games for the first time and finished top of both medal tables with 18 gold, 16 silver and 12 bronze. The distribution is the interesting part: seven golds came from dexterous hand events and six from operational task events, against five from sport. That is a manipulation company winning a manipulation contest, and it is consistent with what the A2 platform was built to do.
The speed records went to the Tiangong Ultra line from the Beijing Innovation Centre, which took both the large and small categories in the 100 metres. The official record stood at 9.39 seconds early in the week and was broken again at 8.86 in the semi finals. The 400 metres fell to 38.15 seconds from a previous best of 1:28.03, and the 1500 metres to 2:21.64 from 6:34.40. In weightlifting, a first ever champion was crowned, the Beijing Humanoid and Huazhong University alliance, while the tug of war went to Yubot’s Walker Taishan.
Not everything went the way the form book suggested. Unitree, the most recognisable name in the field and the most exported, failed to finish the 100 metre obstacle race. In a market where reputation is currently built on video, an event where the clock is public and the run is single take is a healthy corrective.
Worth knowing: dancesport entries went from 29 teams to 108 in one year, and the Games added tug of war, long jump, table tennis and heavy weightlifting. Robots also played tennis against Zheng Jie and table tennis against Ding Ning, both former professionals.
Why a contest tells you more than a demo video
A humanoid robot demonstration is a marketing artefact. It is filmed until it works, cut to length and published without the failures. We have written a whole guide to reading demo videos critically, and every rule in it exists because the sector has trained buyers to expect edited footage.
A timed event removes most of those defences. The robot starts when the official says so, on a surface it did not choose, with a battery that has already done a morning of heats. Teams have to repair between rounds. Failures happen in front of 223 referees and a live audience. None of that makes the Games a certification, and a stadium is not a warehouse, but it produces something the industry has been missing: comparable numbers, generated on the same day, under rules everyone agreed to in advance.
The human comparison that dominated the coverage deserves less weight than the internal one. Robot sprint categories are split by machine size, the start procedure is not an athletics start, and the conditions are not those of a sanctioned track meet, so a robot time next to Usain Bolt’s 9.58 seconds is a headline rather than a measurement. The number that carries information is the one measured against the same event a year earlier: 21.50 seconds to 8.86 seconds in twelve months, on the organiser’s own clock.
The 21 events that should interest buyers most
Scenario events grew from 6 to 21 and this is where the Games stop being sport. The programme included factory assembly, hotel housekeeping, medicine sorting, library shelving, logistics handling and emergency response, scored on completion and accuracy rather than speed alone. Winning entries in the dexterous hand events worked to 0.1 millimetre positioning accuracy, which is the class of precision that separates a demonstration from a task a plant manager would pay for. Our analysis of where dexterous manipulation stands in 2026 sets out why the hand, not the legs, remains the bottleneck.
Jiang Guangzhi of the Beijing Municipal Bureau of Economy and Information Technology put the intent plainly: once robots master these last metre skills, medals turn into orders. That is a policy statement as much as an observation, and it explains why the scenario half of the programme is growing faster than the athletic half. The same tasks are already being run on real production lines, which is the context in which platforms like UBTech’s Walker S in automotive plants should be read.
The failures are the most useful data in the event
Fifty teams entered the 100 metre obstacle race and 13 finished within the time limit. Thirty seven did not. In weightlifting, no entry lifted a 15 kilogram barbell autonomously, so the discipline ran under teleoperation. Those two facts are worth more to a buyer than any medal, because they describe the gap between a robot that can perform a skill and a robot that can perform it reliably, unattended, at the end of a long day.
For buyers: treat the athletic results as evidence of hardware capability and the completion rates as evidence of maturity. A platform that sets a record and does not finish an obstacle course is telling you that its actuators are ready and its autonomy stack is not. That distinction should shape a pilot’s scope, its service level agreement and its spare parts budget, not just its purchase price.
Worth knowing: the organisers released more than 2,500 hours of real operating data free of charge, covering 12 application scenarios, 44 task types and over 100 skills, and they deliberately included failure cases and edge conditions. Failure data is the scarce commodity in embodied AI, and it has just been given away.
What this means for a European decision made this quarter
Three things follow from the week in Beijing. The first is that the capability curve is steep and public, so a platform assessment made a year ago is out of date, and a procurement process that takes eighteen months will specify hardware that no longer represents the market. The second is that the depth of the field is Chinese, which means European buyers are choosing among Chinese platforms whether or not they intended to, and the real variable is not the robot but who services it locally, a point covered in our guide to spare parts and service coverage in Europe.
The third is that events like this accelerate the sector in a way marketing cannot. Public rules, public clocks, published failure rates and a free dataset compress the learning cycle for every team in the hall, including the ones that lost. The third edition is already confirmed for August 2027 in Beijing, with a larger scenario programme. Expect the athletic records to keep falling, and watch the completion rates instead. That is the number that will tell you when these machines are ready to work.
Frequently asked questions
When and where were the second World Humanoid Robot Games held?
They ran from 22 to 26 August 2026 at the National Speed Skating Oval in Beijing, the arena built for the 2022 Winter Olympics. The event drew 666 teams and 2,056 robots from 16 countries across 51 events and 1,301 competition sessions.
Which company won the most medals at the 2026 Games?
AGIBOT topped both medal tables in its first appearance, with 18 gold, 16 silver and 12 bronze. Most of its golds came from dexterous hand and operational task events rather than athletics, which reflects a manipulation focused platform rather than a sprinting one.
Did humanoid robots really break human athletic records?
Robot times in Beijing were faster than the human 100 metre record, but the comparison is not like for like: categories are split by machine size and the start procedure and conditions differ from sanctioned athletics. The meaningful figure is the improvement against the same event in 2025, from 21.50 seconds to 8.86 seconds in one year.
What did the Games reveal about robot reliability?
Only 13 of 50 teams finished the 100 metre obstacle race within the time limit, and the weightlifting events still required teleoperation rather than autonomous control. Completion rates like these are a better guide to deployment readiness than medal counts.
Is the competition data available to companies outside China?
Yes. The organising committee released more than 2,500 hours of real operating data free of charge, covering 12 application scenarios, 44 task types and over 100 skills, including failure cases and edge conditions that are normally absent from published robotics datasets.