What Beijing actually measured

The second World Humanoid Robot Games run from 22 to 26 August 2026 at the National Speed Skating Oval, the 2022 Winter Olympics venue in Beijing. The first edition was held over three days in August 2025. Forbes, previewing the event on 19 August, put this year's field at 2,056 robots from 666 teams (up 311 percent and 138 percent respectively on 2025), drawn from 16 countries and regions across six continents, across 1,301 individual matches. A China.org.cn release carried the same team, robot and country figures on 20 August.

Forbes also reports the number that gives those totals their scale: 96 percent of entries are Chinese. Sixteen countries is true, and it is not the same claim as an international field. For contrast, RoboCup 2026 (the established autonomous-robot-soccer benchmark, held in Incheon from 30 June to 6 July) drew 364 teams from 45 countries. A Chinese team won its Humanoid League there too: Tsinghua University's THU Huoshen, running Booster Robotics' Booster T1, defending the title it took in Brazil in 2025. That is the version of this result that needs no framing at all.

The event count does not reconcile cleanly and should not be averaged. Forbes counts 30 competitive sports events plus 20 work-scenario events, totalling 50. The China.org.cn release, English Wikipedia citing Chosun Ilbo, and RoboHorizon citing the Beijing Municipal Bureau of Economy and Information Technology all give 51: 30 competitive plus 21 scenario-based. The competitive half is stable at 30 in every account; the scenario half is where the sources disagree, by one. On matches, the Associated Press reported more than 1,000 competitions, which is a floor consistent with the organizers' 1,301, up from 487 in 2025.

Everything in this note depends on one rule, and the rule is not published. A time is a claim about a machine only if the machine produced it without a human in the loop. Otherwise it is a claim about a machine and a pilot, which is a different and much weaker thing.

The 2026 edition did tighten this. The China.org.cn release states that the 400 m, the 1,500 m and the 4×100 m relay were reclassified as fully autonomous, with no remote control or human intervention at any point in the race. That is a real, sourced rule change, and it is worth noting what it implies: the 100 m is not on that list. Beyond that the record collapses. A Nerdable report carried by Yahoo Tech on 21 August states that all track and competitive events must be run autonomously except the 100 m and the 400 m hurdles. RobotToday, writing on 31 May, states the opposite: all 100 m events are now fully autonomous with no teleoperation. A third account has it that simple locomotion demands autonomy while harder tasks still permit remote operation, with scenario events scoring teleoperated runs at half points. The organizers' own site publishes no ruleset. So the honest position on day 1 is that we cannot establish whether the 9.39 seconds was produced autonomously, and neither can anyone else reporting it. Correction, 2026-08-24: the sentence above is wrong. The organizers do publish a ruleset, it was on whrgoc.com from 22 August, and this note failed to find it for two days. It marks the 100 metres, both groups, Fully Autonomous Mode Only, which answers for the sprint the question this paragraph leaves open. Day 3 below sets out the rest.

There is one measured figure for how much of this is the robot. At the Beijing E-Town humanoid half-marathon on 19 April 2026, Honor's Lightning covered 21 kilometres in 50 minutes 26 seconds, faster than the human world record for the distance. Scientific American, reporting three days later, recorded that only 38 percent of that event's entries ran autonomously and the rest were piloted remotely, on a dedicated and rehearsed course with support crews trailing the field. Rodney Brooks, quoted there, asked how many support staff and vehicles followed the robots, and compared the exercise to staging races between horses and cars. That is a single independent source for the 38 percent, at a different event, four months earlier. But it is the only number of its kind on the public record, and it is the same robot and the same maker that appear in the sprint results below.

Opening ceremony, then a seven-a-side football exhibition, then the opening heats of the 100 m and 400 m, per RoboHorizon citing the Beijing Municipal Bureau of Economy and Information Technology.

In the 100 m, Tiangong Ultra (built by the Beijing Humanoid Robot Innovation Center, the maker the Associated Press calls X-Humanoid) ran 9.39 seconds and won its heat, overtaking Honor's Lightning near the line after falling behind early. Lightning finished in 9.47 seconds. Both are inside Usain Bolt's 9.58 seconds from Berlin in 2009. Reuters is explicit that this was a preliminary heat; Global Times, which is Chinese state media, calls it a preliminary race. We found no report of a 100 m final on day 1.

Two pieces of context change how that time reads, and both come from the same Reuters dispatch. Tiangong Ultra won the 100 m at the 2025 edition in 21.50 seconds, so the same machine name has cut its time by 56 percent in twelve months. And CCTV, the state broadcaster, reported that Honor's researchers lengthened Lightning's legs by 10 centimetres, to 1.05 metres, ahead of these Games. The fastest 100 m figure in circulation was not run in competition at all: state media reported 9.32 seconds at a peak of 14.5 metres per second from a preparatory test, and Reuters carried it as a state-media claim, which is how its own headline is worded.

In the standing high jump a humanoid, also attributed by the Associated Press to X-Humanoid, "was able to reach" 2.88 metres (the AP's wording, not a paraphrase) against 0.95 metres as the best humanoid mark at the 2025 edition. Every English-language account of the jump traces back to that same AP sentence.

Not established on day 1, and recorded here rather than dropped: whether the 100 m required full autonomy or permitted remote operation; whether the 2.88 metres was a bar cleared or a height reached, and by which part of the machine; whether the 0.95 metre and 2.88 metre figures are the same measurement at all; the start procedure, lane assignment, timing method and distance certification for the sprint; and whether a 100 m final was held. No source we could find addresses any of these.

Twelve gold medals were scheduled, second only to the closing day, on the schedule the Beijing municipal government published on 15 August. The municipality co-organizes the Games, so that is the organizers' own count. The track finals carried the day, and the 400 m ran as two separate finals split by group.

Global Times, Chinese state media, reports that Tien Kung Ultra took the large-group 400 m final in 38.15 seconds, and that Tien Kung Omni won the small-group final later the same day in 45.66 seconds. Both are X-Humanoid machines, and Tien Kung is the romanization day 1 carried as Tiangong. The Telegraph, which is not state media, gives 38.16 seconds for the same robot on the same day, but does not say where its figure came from.

The human comparison breaks again, in a new way. Wayde van Niekerk's 43.03 seconds is a single Olympic final, Rio 2016. Here there was no single final: the large group's winner beat that mark by nearly five seconds, and the small group's winner lost to it by 2.63 seconds. Whether a robot beat the human 400 m record depends on which of the two finals is quoted. Day 1's 9.39 was a heat, so 38.15 is at least a final.

On autonomy the organizers have now said more than day 1 could find. The same municipal page states that except for the 100 m and 400 m hurdle races, all track events require full autonomy. The wording will not settle whether the flat 100 m is exempt or only the two hurdles events, and it disagrees with Reuters, which attributes to the tech partner Huawei a figure of more than 40 percent of the 51 events.

Not established on day 2, and recorded here rather than dropped: which robot ran the 39.7 seconds Reuters prints for the 400 m, and in which round, since it matches neither 38.15 nor 45.66; whether either 400 m final was actually run autonomously; the 1,500 m result, given as 2:21.63 to a Team Tianzhuo machine by an aggregator carrying no agency credit and by nobody else we could find; and nine of the twelve golds, for which no result reached any outlet we searched. No ruleset with scoring has been published.

The floor exercise final, the 400 metre obstacle final, 5v5 middle-size soccer and the nail-fastening leg of the dexterous-hand challenge ran through the day. Weightlifting made its debut in the evening, set at two golds, lightweight and heavyweight, on the Beijing municipal government schedule.

The ruleset exists. Day 1 of this note said the organizers publish none, which was wrong, and the method section above is corrected. A three-page bilingual event table sits on whrgoc.com, the Games’ official site, in a directory dated 22 August, and Gizmodo linked it on 24 August. It lists 51 events, 30 competition and 21 scenario, which settles the count day 1 recorded as disputed, and its last column is headed Autonomy Weight Coefficient. The 100 metres, small group and large group alike, is marked Fully Autonomous Mode Only, as are the 400 metres, the 1,500 metres, the relay, all three soccer events and the floor exercise. Only the 100 and 400 metre obstacle events may be run either autonomously or by remote control. Scenario events weight an autonomous run at 1 and a remote-controlled run at 0.5, which Gong Xiao of the China Software Evaluation Center, an organizing committee member, gave at the official briefing on 23 August, per China News Service.

The column’s third value is a bare slash, and it falls on the day’s own headline. Weightlifting carries no autonomy coefficient in either category, and neither do the two jumps or the kickboxing. CCTV, the state broadcaster, reported that a joint team from the Beijing Humanoid Robot Innovation Center and Huazhong University of Science and Technology went up from a 15 kilogram base to lift 16 kilograms and take the first weightlifting title; that figure rests on CCTV alone. In the 400 metre obstacles, one of the two events remote control is permitted in, Tiangong Omni won for the Tianjiao team in 4 minutes 11.44 seconds, reported by Yicai and carried by the Economic Observer.

Not established on day 3, and recorded here rather than dropped: which weightlifting category the 16 kilograms took, and whether the second gold was awarded; whether that lift or the obstacle run was autonomous, which the rules require of neither; the floor exercise final, which rests on Xinhua photo captions alone giving the win to the Galaxea Starship team and third place to Chengdu Zhiyuan; and whether the 400 metre small group final ran on 23 August, as Global Times and China Daily have it, or on 24 August, as the Economic Observer does. No independent non-Chinese outlet we could find reported a day 3 result. The table gives modes and weights, not scoring, timing method or distance certification.

Seventeen competitions ran and ten golds were decided, five in the afternoon session and five in the evening, per China News Service, the Chinese state news agency. Six events made their debut: street dance, ballroom dance, tug-of-war, pitch-pot, a supermarket retail scenario and an industrial assembly loading station. The afternoon held the 4×100 metre relay, the standing high jump final and both tug-of-war finals; the evening held the two dance finals and the table tennis final, on the session list Beike Finance of Beijing News published.

Tiangong Ultra ran the 100 metres in 8.86 seconds, cutting the 9.39 it set on day 1. Reuters and the Associated Press both carry it, independently of Chinese state media. Every account places it in a semifinal, not a final: Reuters says the semifinal of the large-size division, Global Times, Chinese state media, says the first group of the large-group semifinal, and IT Home, crediting the state broadcaster CCTV, calls it the repechage. The fastest sprint at these Games has still not been run in a final. One thing is firmer than on day 1: the event table found on day 3 marks the 100 metres Fully Autonomous Mode Only, so this run was made under a published autonomy rule, and Global Times adds that the team replaced line following with map-based positioning, which is the maker’s own account of how. The Associated Press reports the robot fell past the line into a padded barrier with sparks coming off its waist, and Reuters reports that a robot came apart mid-race in the other semifinal.

Not established on day 4, and recorded here rather than dropped: nine of the ten golds, since no outlet we could find names a winner of the relay, either tug-of-war, the table tennis or either dance final, though Xinhua photographs confirm the lightweight tug-of-war was contested; whether the 8.86 seconds was autonomous in fact and not only by rule; and the standing high jump, where 2.8843 metres is attached both to a day 1 demonstration and to day 4’s final, and credited both to Tiangong Ultra and to a machine called Tianxiao. Xinhua Daily, a Jiangsu party title, calls it a height reached in the air, the first source we have seen to say either. Two Chinese outlets have also started printing the 9.32 seconds as a competition result, one of them as a world record, where Reuters on day 1 reported it as a preparatory test run outside competition.

This is the cleaner of the two errors and it is in every account. The robot competed in the standing high jump, in which the athlete takes no run-up and jumps from a standstill with both feet together. The 2.45 metres attributed to Javier Sotomayor in 1993 is the running high jump record, set with an approach and a Fosbury flop. They are different disciplines, and the standing version has its own human mark: 1.90 metres, credited to Sweden's Rune Almén, for an event last contested at the Olympics in 1912. Guinness World Records maintains a separately defined category, highest standing jump by a male, at 1.70 metres by Christopher Spell in February 2021. Take either figure and the robot's 2.88 metres is still a large number. Take 2.45 metres and you have compared a standing jump to a running one, which is the comparison that ran worldwide.

The second problem is the verb. The AP says the robot "was able to reach" 2.88 metres. Clearing a bar at 2.88 metres from a standstill and touching a point 2.88 metres above the floor at the top of a jump are wildly different feats, and only one of them is a high jump. No source states which was measured, or how. The 2025 comparison figure of 0.95 metres makes the ambiguity worse rather than better: 0.95 metres is a plausible bar height for a small humanoid and an implausibly low reach for any machine that can stand up, which is a reason to doubt that the two numbers measure the same thing before treating their ratio as a threefold gain in twelve months.

The sprint has three separate weaknesses, and none of them is that the robot did not run. First, 9.39 seconds was set in a preliminary heat, and a heat time is a real result that is not the same object as a final. Second, the fastest figure in circulation, 9.32 seconds, came from a preparatory test reported by state media: faster than the official result, produced outside competition, and therefore evidence that the number being called a record is not the fastest number anyone has claimed. Third, and unresolved, is autonomy. Bolt's 9.58 was run from blocks, in a final, in lanes, over a certified distance, under a published rulebook, with drug testing and an appeals process. The robot's 9.39 was run under conditions the organizers have not published. Both numbers are probably true. They are not measurements of the same thing, and the difference is not a technicality; it is the entire content of the comparison.

Underneath the athletics sit twenty or twenty-one work-scenario events, and they are the part of this programme that bears on anything. Forbes lists retail, food service, assembly, warehousing, hospitality and household work, plus a dexterous-hand competition of eight tasks: power-tool assembly, powder weighing, bricklaying, block stacking, nail fastening, opening caps and lids, unpacking boxes, picking beans with tweezers and connecting cables. That is not a sports programme. It is a task taxonomy, and it reads like a list of the jobs a Chinese integrator expects to sell next year. Forbes quotes the analyst Wang Peng to the effect that last year's question was whether the robots could reach the finish line and this year's is whether they can get the job done.

That reframing is the thing worth watching, and it lands exactly where the thesis says value accumulates. A sprint time is a property of a machine. Weighing powder to a tolerance, fastening a nail, and unpacking a box on a line are properties of a machine plus a fixture plus a routing decision plus a record that proves afterwards what was done, which is why the box beats the body, and why a fall cannot be certified without a framework to certify it against. A benchmark that scores autonomy differently from teleoperation is doing certification work whether it calls it that or not. If these Games publish their scenario rules and their scoring, they become a useful public benchmark for the execution layer. If they publish only the sprint times, they are a trade show. Our dated, falsifiable bets sit in the predictions register and the tracker holds the hardware; this note scores none of them.

Five limits, and the first is the largest. Nothing here says the results are fake. Tiangong Ultra ran a heat and won it, Lightning ran 9.47 behind it, and a machine that goes from 21.50 to 9.39 seconds in a year has had extraordinary engineering done to it. The objection is to the comparison, not to the achievement. Second, the autonomy question is open in both directions: the 100 m may well have been fully autonomous, and if the organizers publish a ruleset saying so, that part of this note is answered and should be corrected here. Correction, 2026-08-24: the ruleset was found on day 3 and it does say so, so this limit is closed for the sprint and the other flat track events, and stands for the jumps, the weightlifting and the kickboxing, which the same table exempts from any autonomy requirement. Third, the 38 percent autonomy figure is a single independent source, from a different event four months earlier, and is offered as the nearest available base rate rather than as a measurement of these Games. Fourth, the standing high jump human marks are thin ground of a different kind: an event abandoned in 1912 has no modern competitive record, so 1.90 metres and 1.70 metres are the best available references and not a governed record in the sense that 2.45 metres is. Fifth, this log is being written while the event is running, from wire copy and previews, with state media as the only source for several figures; a report filed on day 1 is the least reliable version of day 1 that will ever exist. Corrections are welcome and will be made in place, with the change noted.

Every figure above carries its outlet and its date. Chinese state media (Global Times, CCTV, China.org.cn) is named as state media wherever it is the source, and the two claims that rest on it alone (the 9.32 second preparatory test and Lightning's lengthened legs) are labelled in the text. Where the record is contradictory, as it is on the autonomy rules, both accounts are given rather than one being chosen. Where nothing could be sourced, the day's block says so instead of omitting it. The Games have no published ruleset that we could find, and that absence is itself reported rather than filled in.

Reuters (22 August 2026): Chinese humanoid robot Lightning beats human 100m world record, state media says ·Associated Press via ABC News (22 August 2026): Chinese humanoid robots smash human records in 100m sprint and high jump ·Associated Press via ESPN (22 August 2026): humanoid robots surpass human records in 100m, high jump ·Global Times, Chinese state media (22 August 2026): humanoid robot breaks human 100m record with a 9.39 dash ·Forbes (19 August 2026): the World Humanoid Robot Games events are a market map ·China.org.cn, Chinese state media, via PR Newswire (20 August 2026): the World Humanoid Robot Games are back ·RoboHorizon (22 August 2026), citing the Beijing Municipal Bureau of Economy and Information Technology: event counts and the opening-day programme ·Nerdable via Yahoo Tech (21 August 2026): record participation and fully autonomous events ·RobotToday (31 May 2026): 2026 event preview, stating all 100m events are fully autonomous ·Wikipedia: World Humanoid Robot Games, the 2025 and 2026 editions ·Wikipedia: standing high jump, the rules and the 1.90 m mark ·Guinness World Records: highest standing jump (male), 1.70 m, 7 February 2021 ·Scientific American (22 April 2026): a humanoid robot beat the human half-marathon record, but what did it prove ·Interesting Engineering (6 July 2026): China wins RoboCup 2026, 364 teams from 45 countries, fully autonomous play ·Beijing municipal government (8 July 2026): Tsinghua team crowned champion in the RoboCup Humanoid Soccer League ·Global Times, Chinese state media (24 August 2026): the viral face-covering sprint is spontaneous rather than designed, carrying both 400 m final times ·Beijing municipal government, a co-organizer (15 August 2026): highlights, the day-by-day schedule, the tightened track time limits and the autonomy requirement ·The Telegraph via Yahoo (23 August 2026): robot beats the human 400 m world record ·Reuters via the Taipei Times (24 August 2026): China showcases robots at the Beijing games ·Reuters via the Kathmandu Post (23 August 2026), the same dispatch as the Taipei Times entry above: robots can outrun humans, but can they plug in a cable ·Bastille Post (24 August 2026), carrying no agency credit: a Chinese robot shows agility and speed in the 1,500 m race ·World Humanoid Robot Games organizing committee (22 August 2026): the official event table, all 51 events with their Autonomy Weight Coefficient ·Gizmodo (24 August 2026): the robots beat Usain Bolt, then met their match, beans, the piece that links the ruleset ·China News Service (23 August 2026): Gong Xiao of the China Software Evaluation Center explains the 1 and 0.5 autonomy weights at the official briefing ·IT Home (24 August 2026), crediting CCTV, Chinese state broadcaster: the first weightlifting title, 16 kilograms lifted from a 15 kilogram base ·Economic Observer (24 August 2026), carrying Yicai: Tiangong Omni takes the 400 m obstacle final in 4:11.44 ·Xinhua photographs, Chinese state news agency, via NetEase (24 August 2026): day 3 captions for the floor exercise, 5v5 soccer and nail fastening ·China Daily, Chinese state media (24 August 2026): dated captions placing the 400 m small group final and the long jump on 23 August ·China News Service, Chinese state news agency (25 August 2026): seventeen events on 25 August, and the six new events making their debut ·Beike Finance of Beijing News, via NetEase (25 August 2026): the day 4 session list, ten golds, five in the afternoon and five in the evening ·Associated Press via ABC News (25 August 2026): a Chinese humanoid robot sets a new 100 metre sprint record at Beijing games ·Reuters via Al Jazeera (25 August 2026): Chinese robot Tiangong clocks a sub-9 second 100 metres in the semifinal of the large-size division ·Global Times, Chinese state media (25 August 2026): Tiangong breaks the 100 m record again with 8.86 seconds, and the switch to map-based positioning ·IT Home (25 August 2026), crediting CCTV, Chinese state broadcaster: 8.86 seconds in the large-group 100 m repechage ·Xinhua, Chinese state news agency (25 August 2026): photographs of the lightweight tug of war, four teams named, no result given ·Xinhua Daily, a Jiangsu party title (23 August 2026): Tiangong Ultra airborne 2.8843 metres in the standing high jump demonstration ·Huaxia Times, via NetEase (25 August 2026): medal tallies to 24 August, and 9.32 seconds printed as one of four world records