Autonomous robots won most of the job events

Some of the first medals of the World Humanoid Robot Games 2026 were decided in a school library, days before the opening ceremony. The library final ran from 16 to 18 August at Beijing No. 80 Middle School, where robots sorted books, while the hotel and firefighting finals ran on 16 and 17 August in real buildings across the city. These were the World Humanoid Robot Games scenario events, 21 of them, and they are the part of the Games the organisers built to look like work. By our count of the 21 final result sheets, a robot running on its own won 13 of them. A remote-controlled robot, scored at half points, won the other eight. For a warehouse or factory buyer, those sheets are the nearest thing to a job trial on the record, and they say more than the sprint times and medal tables that made the headlines, which we check below as well.

What were the World Humanoid Robot Games scenario events?

The organisers’ scenario regulations, published on 2 July, divide the 21 events into three groups. Twelve are work posts, each modelled on a job. In retail, a robot picks goods for a customer; in catering it takes an order and heats food. Two industrial posts cover assembly and loading, and packing and warehousing, where the robot packs finished medicine boxes into cartons. Two emergency posts cover firefighting and rescue, and general emergency response. The rest are office services, housekeeping, hotel guest services, library book sorting, park management and charging electric cars.

Eight more are dexterous hand events, each a single manual task: assembling with a power tool, weighing powder, building with blocks, driving nails, opening bottles, unpacking boxes, picking up beans with tweezers and connecting cables. The last event is a combined five-task run on a 100 metre out-and-back course, with a handover, a heavy carry and a cart to push along the way.

Four work posts were held away from the arena. Housekeeping was held in a show home, firefighting at a Beijing fire service logistics centre, hotel service in a working hotel and book sorting in the school library. Park management and car charging were held outside the National Speed Skating Oval itself.

Xinhua described the choice of jobs as “designed around jobs that can be repetitive, physically demanding, dirty, dangerous or difficult”. In the hotel event, it reported, “more than 10 robot teams were given 30 minutes to tackle three tasks: moving luggage of different sizes, replenishing guest-room supplies and making beds.”

What those robots were actually built to do is the subject of physical AI explained.

What is the World Humanoid Robot Games?

It is a multi-sport event for machines, held every August since 2025 at the National Speed Skating Oval, which Beijing built for the 2022 Winter Olympics. Every competitor is a robot built in human form, though the working events allow wheels in place of legs. The RoboCup Federation’s listing for the 2026 edition describes it as the world’s first international “science and technology sports event with humanoid robots as the primary competitors”. Xinhua lists the four hosts: “the Beijing municipal government, China Media Group, the World Robot Cooperation Organization and the Asia-Pacific Robot World Cup International Council”. The last of those is RoboCup’s regional body for Asia and the Pacific. Teams enter from universities, research institutes and companies, and a team can bring a robot from any maker.

It is not the same as RoboCup, the older competition best known for robot football. RoboCup’s 2026 world championship was held in Incheon, South Korea, from 30 June to 6 July, and the Federation lists the Games separately, under its KidSize, TeenSize and AdultSize soccer leagues. A robot team can enter both.

The main competition of the second Games ran from 22 to 26 August 2026, and five of the working events held their finals in the week before. The Beijing municipal government’s official preview and Xinhua’s report on the opening set the two editions side by side:

Teams, events and matches all rose between the first and second Games.

The 51 events split into 30 competitive events and 21 scenario events. The officiating team grew to 223 referees, and around 1,000 humanoid robots took part in the 2026 opening ceremony alone. Xinhua’s Chinese report on the closing says 16 countries and regions. CGTN, the channel of China Media Group, summed up the programme at the close: “Robots competed in football, table tennis, gymnastics, weightlifting, martial arts, tug-of-war and combat events, while 21 scenario-based events included tasks related to factories, hotels, supermarkets and emergency response.”

How were the World Humanoid Robot Games scenario events scored?

Each work post is a sequence of stages completed in order inside a time limit, 20 minutes for posts such as retail, catering, assembly and packing. A team may not skip or abandon a stage. The regulations add that “if any task item does not meet the success requirement, that item scores 0; there are no negative scores”. The dexterous events are shorter: most count how many valid completions a robot makes in five minutes, while powder weighing and block building are timed. The combined event is scored on finishing time plus penalties.

Over the work posts and the dexterous events sits one multiplier. Task points are the sum of stage scores, minus deductions, times a weighting: 1 for a robot running on its own and 0.5 for one under remote control. The regulations define the two modes. Manual remote control, “including semi-automatic”, must be wireless, with the operator in a designated area. Full autonomy means “the whole run must be completed by the robot autonomously, except for sending the start command from the designated area”. A team chooses its mode at check-in and may not switch during the run. The organisers’ own tie-break puts an autonomous robot ahead of a remote-controlled one on equal points.

The weighting changes results. In the emergency response final, a remote-controlled robot scored the full 300 points on the tasks and finished with 150. A robot that ran itself scored 250 and took the gold.

The remote-controlled robot completed more of the tasks and still lost, because its score was halved.

The halving works as a price on human help. A team that trusts its robot can take the full score. A team that does not can still compete and still win, if the autonomous teams fail enough tasks or do not enter.

Which other events allowed remote control?

By our count of the rulebooks and the result notices, 23 of the 30 competitive events required the robot to run itself. The Beijing government’s English preview says: “Except for the 100m and 400m hurdle races, all track events and other competitive events require robots to complete them fully autonomously.” The athletics rulebook bears it out, and the rulebooks for football, martial arts, gymnastics, tug of war, pitch-pot, table tennis and the dance events also require full autonomy.

Two rulebooks say something different, and the summary leaves both out. The kickboxing regulations state: “The robot’s control mode may be manual remote control (including semi-automatic), full-body teleoperation or fully autonomous.” The rulebook does not define full-body teleoperation; the term usually describes a robot copying the movements of a person. Kickboxing, which ran in three weight classes, is the one discipline where a person may drive a robot through a fight. The weightlifting regulations let a robot be driven by remote control, but “after reaching the competition area, the whole competition process must be completed autonomously by the robot”. Both weightlifting classes, 40 kilograms and under and above 40 up to 80, run under that rule. The quotations are our translation of the Chinese.

How did the organisers check that robots ran on their own?

A rule that pays double for autonomy invites teams to declare it. Teams “must declare control mode truthfully”, and the organising committee may inspect “run data, control command records, operation logs, video” before, during and after the event. If a team declared autonomous mode but used remote control or “hidden intervention”, the committee may cancel its result.

The committee did cancel results, under a separate notice on fair competition for autonomous teams, which its announcements cite but which it has not published. Its notices, issued between 20 and 26 August by a fair-competition technical review centre, cancelled four results. Two teams, in the hotel and firefighting events, did not attend the post-event review in time. Two more, in the housekeeping and assembly events, were asked to reproduce their runs on site and did not reach the required task completion rate. None of the notices says a team was caught steering its robot.

Who won the World Humanoid Robot Games scenario events?

The organisers post the scenario finals on the announcements page of the official site, and we read all 21 scenario sheets, using the corrected versions where one was issued. Split by group, a robot running on its own won nine of the 12 work posts and four of the eight dexterous hand events. Remote-controlled robots won the other three work posts and the other four dexterous events. The combined five-task event went to a robot scored at the remote-control weighting, and every entry on that sheet carries the same 0.5, though four of the 17 withdrew and two did not finish.

By our count of the corrected sheets, autonomous robots won 13 of the 21 scenario finals and remote-controlled robots won eight.

Teams carrying AgiBot’s Chinese name, Zhiyuan, won seven of the eight dexterous hand finals and five work posts: firefighting, emergency response, packing, assembly and library sorting, the last as a joint team with Tsinghua University and Shanghai Jiao Tong University. Teams linked to X-Humanoid, the Beijing Humanoid Robot Innovation Center, won park management, car charging and the combined event. A team carrying the Chinese name of UniX AI won the hotel post, a team named after Spirit AI won office services, and a team called Beijing Yinhe Xingshu won both housekeeping and catering. The sheets name teams, not robot models.

Four of those results were decided after the event, by the cancellations above. In all four events the cancelled team had been ahead of the eventual winner. In hotel service and firefighting, the gold passed to a remote-controlled robot scored at half points. In effect, a checked half score beat an unchecked higher one.

Who topped the World Humanoid Robot Games medal table?

The organisers publish result sheets for most finals and no medal table. Their announcements page carried 84 notices when we checked it on 27 September: result sheets for finals and preliminary rounds, corrected sheets, cancellations and administrative notices. None is a medal count, and a few finals, football among them, have no result sheet there at all. Three tables are in circulation, with three different winners.

The closest thing to an official table is the one the Beijing Daily’s app published on 27 August under the heading “Final version!”, carrying the Games’ logo. It ranks teams by golds, then silvers, then bronzes. First is Tiangong, X-Humanoid’s own team, with 5 gold, 0 silver and 4 bronze. Second is the Shanghai AgiBot team with 4 gold, 4 silver and 1 bronze.

The company counts are a different kind of table. AgiBot’s announcement of 26 August claims 18 gold, 16 silver and 12 bronze and says it “ranked first on both the gold medal table and the overall medal table”. X-Humanoid, in a post of 8 September, claims “the Tiangong family’s 15 gold, 12 silver and 18 bronze, first on the medal table”, produced with “secondary development joint teams”: 49 joint teams using its Tiangong robots, drawn from 31 universities and from companies including UBTech and BrainCo. Its total is a count by robot brand. AgiBot does not say how it counted; teams carrying its name were entered from Shanghai, Shenzhen, Ningxia, Chengdu, Wuxi and Beijing, and others are joint teams, such as the Wuhan University and AgiBot team that won powder weighing. Company tables compiled by Chinese news sites follow the companies’ figures, with AgiBot first, X-Humanoid second and Unitree third.

AgiBot has more golds and more medals on its own count, X-Humanoid more bronzes. Both count several teams under one brand.

Some of AgiBot’s claims can be checked against the sheets, and most hold. AgiBot says OmniHand “reached the finals in all eight events and won seven gold medals” in the dexterous hand competition; the eighth, power tool assembly, went to a team called Lingxin Qiaoshou. AgiBot says its X2 won the 100 metres obstacle race, and the sheet agrees that an AgiBot team, from Shenzhen, won in 1 minute 28.32 seconds, though it does not name the robot. Two claims do not match. AgiBot says X2 took “silver and bronze medals in the 400m obstacle race”, but the official sheet gives silver to the Shanghai AgiBot team and bronze to Tiangong. And AgiBot lists hotel services among its golds, while the corrected hotel sheet puts the UniX AI team first and two AgiBot teams second and third; the first version of that sheet had no AgiBot team ahead of UniX AI’s either.

Unitree’s own team appears on the podium twice in the organisers’ sheets, with silver in Tai Chi, behind the Shanghai AgiBot team, and silver in the 40 kilogram free combat, behind a team from North China Electric Power University. It is not in the top three of any athletics final. One media table credits Unitree with 10 medals, a total that can only come from counting other teams that used its robots. Neither company table can be rebuilt from the official sheets, because the sheets do not say which robot each team used.

Did humanoid robots break human world records at the 2026 Games?

On the track, yes, in three events. The fourth claim does not measure what a human record measures. X-Humanoid made the claims in its own posts. Its post of 23 August leads with “9.32s!! Faster than Bolt’s 9.58s”, says its Tiangong Ultra, “performing as the opening-ceremony curtain-raiser demonstration, leapt to 2.88 m, beating the human high jump record”, and says it won a 400 metres preliminary in 39.70 seconds. Later posts added the 1,500 metres.

The 9.32 seconds has no official sheet behind it. The organisers’ announcements page carries result sheets for the 100 metres finals only, so no heat time is published. Reuters and the Associated Press attribute a 9.32-second run to a different robot, from the phone maker Honor, in a test before the Games.

The organisers’ result sheet for the large-class 100 metres final on 26 August lists five finishers. The Tianjiao team won in 8.64 seconds and the Tianxiao team was second in 8.65, both joint laboratories that X-Humanoid runs with Peking University and Hebei University of Technology. Two teams entered by Honor companies came third and fourth in 8.83 and 8.85, and the Jinghong Dongli team was fifth in 8.87. The sheet’s wind and reaction-time columns are empty, and it names teams, not models; X-Humanoid names the winning robot as its Tiangong Ultra.

All five finishers in the large-class final ran inside Usain Bolt’s 9.58 seconds. The result sheet names teams, not robot models.

The comparison is fair as far as a clock can make it. The athletics regulations list the 100, 400 and 1,500 metres and the 4 by 100 metres relay as fully autonomous: a team may send only a start command and a stop command. The preview also tightened time limits, to one minute for the 100 metres instead of three, five minutes for the 400 metres instead of 15, and 15 minutes for the 1,500 metres instead of 40. For the 100 and 400 metres, robots are split by height into a small class, 1.4 metres and under, and a large class above that.

CGTN gives the comparison with 2025: “The large-size robot 400m record fell from 1:28.03 last year to 38.15 seconds, while the best 1,500m time improved from 6:34.40 to 2:21.64. The standing high jump record also rose from 0.95641 meters to 3.40 meters.”

The human records are Usain Bolt’s, Wayde van Niekerk’s, Hicham El Guerrouj’s and Javier Sotomayor’s, as the World Athletics world records list gives them today. The high jump row compares different events. Each robot takes off from both feet inside a 3 metre by 3 metre area, with “run-ups, stutter steps and repeated jumps” banned, and “the high jump result is the vertical distance from the lowest point of any part of the robot’s body to the ground when the robot reaches the highest point of its jump.” There is no bar, and World Athletics lists no standing high jump at all. The 2.88 metres was an opening-ceremony demonstration, not on any result sheet. The final, on 25 August, was won with 3.40 metres.

Four numbers, three ways of measuring: the robot marks are the height of the robot’s lowest point at the top of a standing jump, and the human record is a bar cleared after a run-up.

No body ratifies humanoid robot world records, so a Games record is the most any robot time can claim. The times are real, set by machines in a class of their own, the way a cycling record over 1,500 metres is real and is not a running record.

What the World Humanoid Robot Games scenario events mean for buyers

The work posts are the closest thing the Games have to a product test, and they are harder than a spec sheet because the task order is fixed and a failed stage scores nothing. A robot that packs medicine boxes into a carton, boxes blister packs and folds flat cartons, all in 20 minutes without help, has done something a buyer can picture in a warehouse. No robot managed that on its own this year: the three packing medallists all scored the full 300 under remote control, and the best robot running itself scored 54. The same robot finishing under remote control has shown what a person and a robot can do together, which is a different product. Spec sheets describe the same abilities one at a time, as we found when we compared humanoid robot lift figures, and a timed sequence of stages tests them together.

That is the question to carry from Beijing into any sales conversation. AgiBot says the G2 it used at the Games already works in factories run by Longcheer and SAIC, and that its 15,000th robot rolled off its production line in June 2026; those are company statements, not results. What the scenario sheets add is the mode each run was scored in. Our piece on when you can buy a humanoid robot reads the delivery dates the makers have promised, and what a humanoid robot costs sets out their prices. A buyer can put the referees’ question to each of them: which parts of the job does the robot do on its own?

How to read a maker’s medals and records

A medal total is the number most likely to appear in a sales deck, and it is the hardest to check. Both company totals gather several teams under one brand, so they measure reach as much as performance; with 666 teams registered, a maker missing from every podium may simply have entered few. Ask a maker for the list of teams behind its total, then read those teams’ sheets. Weight the events by what the business needs: a gold in Tai Chi and a gold in packing count the same on every table. Where a claim and a sheet disagree, as with AgiBot’s gold in hotel services, treat the sheet as the record.

For a record claim, ask whether there is an official result sheet, whether it was a competition or a demonstration, how it was measured, and which model and configuration ran it under whose control. The 9.32 seconds fails the first test and the 2.88 metres the second.

How to check a robot’s autonomy in a pilot

The scenario regulations give a pattern a buyer can copy into a pilot agreement. Ask for the control mode to be declared before each run and fixed for its length. Ask for the records the committee reserved the right to inspect: run data, control command records, operation logs and video. Ask for the right to have a run repeated on site, the test two teams failed in Beijing. All of it has to be agreed before the pilot starts, because a maker cannot hand over a control log that was never kept. Then cost any run that needed an operator with that operator’s time included. In the emergency response final, the remote-controlled robot completed more of the tasks, 300 points against 250, and a buyer counting only tasks done would have picked it.

Which World Humanoid Robot Games scenario events matter to a warehouse, and how to simulate them

Two of the twelve work posts map onto a warehouse or factory floor: assembly and loading, and packing and warehousing. The retail post, where a robot picks goods for a customer, is close to order picking. In packing, the three medallists all needed a person, each scoring the full 300 under remote control, and the best robot running itself scored 54, which is 18 percent of the points on offer.

Assembly is harder to read. The leading team’s result in that final was cancelled after the event, so the standing sheet is not the run spectators saw. The dexterous hand events are closer to single factory motions, such as driving nails, connecting cables and unpacking boxes, and there autonomous and remote-controlled robots split the eight finals four each. Autonomous robots won nine of the 12 work posts, but a win is relative to the field.

Physics simulation is how a post like packing could be rehearsed before any robot arrives. In our MuJoCo on Mac guide we measured Unitree’s G1 at 38,596 steps a second on one core, 77 times real time, and its model holds its pose because its joints are position servos. That speed is for the standing model with no objects. We have not built the packing post in simulation, so we have no speed with contacts and no completion rate. What would settle it is a MuJoCo scene with a G1, a table, a carton and boxes, the regulations’ 20-minute limit and fixed stage order, measuring simulation speed with contacts and the share of attempts that finish every stage, and that is on our list.

At the closing ceremony, according to CGTN, organisers “released a dataset containing more than 2,500 hours of real-world operational data gathered from training and competition”, covering 12 application scenarios, more than 100 skills and over 10,000 tasks. When we looked a month later we could not find it anywhere to download.

CGTN also reported that “the third World Humanoid Robot Games will be held in Beijing in August 2027, with scenario-based events expected to make up a larger part of the next program.” As of 27 September 2026 the organisers had not published exact dates, a venue or a list of events; they should appear on the official site, whrgoc.com. The 2026 rulebooks were published on 2 July, 51 days before the main competition, so the 2027 rules should be public well before any result. This year, eight of 21 scenario golds went to robots operated by a person. The count to watch next August is whether that number falls, and a result sheet with one more column, the robot model, would make every claim that follows checkable on the night.