News
What's happening in robotics and AI — curated by the wujec.ai editors.
A robot ran 100 m in 8.64 s, then hit the wall — the record is in the legs, the gap is in the brakes
Tiangong Ultra won the large-machine 100 metres at the second World Humanoid Robot Games in Beijing on 26 August 2026 in 8.64 seconds, a Games record and almost a full second inside Usain Bolt's human mark of 9.58 s. The two numbers are not comparable — the robot runs on a prepared track, on a machine tuned for one straight line — but the sprint itself is a real engineering result, and it is the third record set by the same machine, built by the Beijing Innovation Center of Humanoid Robotics, founded in 2023. What happened next is the part worth recording. Several finishers, the champion among them, ran straight into a padded barrier past the line. Sparks came off some of the machines, staff moved in with fire extinguishers, and according to the Associated Press report from the venue nearly all of them were carried off on stretchers. That is not a footnote to the record; it is the other half of it. Acceleration and gait are now solved well enough to beat a human time over a fixed distance. Controlled deceleration — knowing where the finish is, shedding speed without falling, stopping a 100-kilogram machine before it meets an obstacle — is not. In a stadium the failure is padded and photogenic. In a warehouse aisle or a hospital corridor, the same missing capability is the whole safety case. The Games were held in an Olympic format across football, boxing and athletics, with more than 2,000 humanoids entered. Jonathan Hurst of Oregon State University, quoted by AP, put the caveat plainly: real progress will be measured in warehouses and factories, where humanoid robots work autonomously for hours and produce actual economic value. For a catalogue, that is the useful reading of a sprint record. The event proves a specific, narrow thing about what these machines can do, and — by what happened twenty metres past the tape — an equally specific thing about what they cannot.
Tiangong Ultra →The humanoid games were won by a hand, not a leg: seven of AgiBot's 18 golds came from a gripper
AgiBot closed the second World Humanoid Robot Games in Beijing at the top of both medal tables — 18 gold, 16 silver and 12 bronze, 46 medals in total — in what the company describes as its debut at the event. The distribution of those medals says more than the total. Seven golds were won by something that does not walk at all: OmniHand, the company's dexterous hand, reached the finals of all eight manipulation events and won seven of them, including block building, powder weighing and picking beans with tweezers. Five more golds came from scenario-based events — hotel service, library operations and emergency response. Only two golds came from the kind of contest that dominated the opening days: X2 won the 100 m obstacle race and A3 won the tai chi event. AgiBot entered four machines in total — OmniHand, the Genie G2 already working on factory floors, A3 and X2. That is a shift in what the games measure. The first days of the tournament were about sprint times and records; the overall table was taken by a company that won only two events on the track. The points sat where a robot has to grip something, measure it out, or complete a job on the spot — the work that decides whether a humanoid is useful in a warehouse or a hospital corridor. Two caveats belong with these numbers. The tally comes from the winner's own press release, which does not say who finished second or third, and the organisers have not published a full table in text form. The release also refers simply to "A3"; our catalogue documents the A3 Ultra shown in July 2026, and the company does not specify whether the same variant competed.
AgiBot A3 Ultra →Last year's champion cut its own entry list — and the times Unitree won with in 2025 were beaten by 50 seconds and four minutes
The second World Humanoid Robot Games close in Beijing today, and the clearest measure of how fast this field moves is not the winner but the machine that stopped competing. At the first edition a year ago, Unitree won the 400 m in 1:28.03 and its H1 won the 1500 m in 6:34.40. At this edition, the 400 m large class went to Tiangong Ultra in 38.15 s and the 1500 m was won in 2:21.64. The old marks were beaten by roughly fifty seconds and by more than four minutes. Both of this year's times are inside the human world records for the distances — 43.03 s and 3:26.00 — and in the 1500 m the entire top six finished ahead of the human mark. Unitree itself did not defend either title. Its robot finished at the back of its 100 m qualifying group, and the company issued a statement the same day: constraints on time, on the number of new machines available and on testing had led it to withdraw from some events it had already entered, and its effort remains on products that are in volume production. Founder Wang Xingxing watched from the stands. The company did compete in taijiquan, where it scored 21.21 points; the gold in that event went to AgiBot. The scale of the event grew with it. Organisers counted 666 teams from 16 countries across six continents, 2,056 robots and 51 event categories over 1,301 contests at the National Speed Skating Oval — 138% more teams than last year and four times as many machines. The rules changed in a direction that makes the times harder to dismiss. Several competitions, the 400 m among them, dropped remote control entirely: an operator may give the start command and nothing else. In the scenario contests — pharmaceutical packing, office work, EV charging, catering, emergency response — a fully autonomous run is scored at weight 1.0 and a teleoperated one at 0.5. Two robots now share a lane instead of running alone, and the obstacle courses added box climbing, net crawls, tyre arrays and stepping stones. Han Gang, a motion-control specialist at the Beijing Innovation Center of Humanoid Robotics, points at navigation as the quiet reason for the jump: in 2025 the robots ran by looking at the painted lane markings, while this year they localise and map, and hold a lane at speed without seeing the line at all. For a company already shipping robots by the thousand, a games medal is not the same prize it is for a research centre. But the gap is now measured in minutes, and it was opened in twelve months.
Tiangong Ultra →Robots now outrun humans — but they still jump a metre short
The second World Humanoid Robot Games in Beijing produced a result that reads oddly when you put two of its events side by side. On the track, Tiangong Ultra took the 1500 m gold in 2 minutes 21 seconds. The human world record over that distance is 3:26.00, set by Hicham El Guerrouj in 1998 — and the machine's builder, the Beijing Innovation Center of Humanoid Robotics, headlined the result as beating the human mark by one minute and six seconds. In the long jump the same machine won as well, but with a very different picture: 7.97 m for gold, then 7.24 m and 7.19 m for the other two places on the podium. Mike Powell's human record has stood at 8.95 m since 1991. The robots beat us at running by a wide margin and lose the long jump by nearly a metre. That gap is not a fluke of one competition, it follows from how these legs are built. Running rewards a machine that can put power down continuously, cycle a limb quickly and stay upright — all things a torque-dense actuator with a fast control loop does well, and does without getting tired over the distance. A long jump is one single explosive extension, where the whole result is decided in the fraction of a second the foot spends on the board, and where a human athlete converts a full-speed approach into lift with tendons that store and release energy. Robot joints have no equivalent of that elastic return; a motor can only push as hard as its peak torque allows. The other number worth reading is the podium itself. All three long-jump medals went to the same model from the same team, and the centre's own bulletin claims the first four places. Its running total after 23 August was five golds, two silvers and six bronzes, including every athletics gold awarded that day. A clean sweep looks like dominance, and it partly is — but it also means the machines were mostly racing copies of themselves. At this stage of the sport, the standings measure how many units a well-funded lab can bring to the start line at least as much as they measure the state of the art.
Tiangong Ultra →The first gold of the robot games was won in the joints, not in the brain
The first gold medal of the second World Humanoid Robot Games was decided on the afternoon of 23 August 2026, in the 400 m large-class final at Beijing’s National Speed Skating Oval. Tiangong Ultra, from the Beijing Innovation Center of Humanoid Robotics, won it in 38.15 s. The human world record over the distance is 43.03 s. The night before, the same machine had taken the qualifying round in 39.70 s — also inside the human mark. In the small-class final an hour earlier, the centre’s smaller Tiangong Omni won in 45.66 s, so both 400 m titles went to the same laboratory. Asked what changed in a year, the engineer who leads Tiangong Ultra development, Guo Yijie, did not talk about intelligence. He listed joints: torque and rotational speed both raised substantially, higher motor power for acceleration and take-off, better reliability in the electrical components. On the software side the answer was equally mechanical — the control framework was retuned event by event, because the 100 m rewards acceleration while the 400 m and 1500 m reward taking corners at speed. That is the honest reading of these headline times. A sprint record measures actuators, thermal headroom and a controller that has been tuned for one specific track, and those are exactly the things that a well-funded hardware team can improve in twelve months: on this distance the winning time has come down from 21.50 s at the first games to 9.39 s in a qualifying heat this year. It says much less about whether a robot can decide anything for itself. The same centre entered four machines — Tiangong 3.0, Tiangong Ultra, Tiangong Omni and Tianyi 2.5-TE — across more than forty events, and the interesting half of that list is not on the track. Tianyi 2.5-TE is running the dexterous-hand contest and the retail, catering, industrial, office and home scenarios; Tiangong 3.0 is in the emergency-response and all-round events. Those events have no human record to beat and will produce no headline number, which is why they are the ones worth watching: they are the only part of the programme that tests the part of the robot nobody can upgrade with a stronger motor.
Tiangong Ultra →Two robots went under Bolt’s record in one qualifying heat — and the third one fell at the line
Beijing’s National Speed Skating Oval hosted the ninth qualifying heat of the 100 m at the second World Humanoid Robot Games on 22 August 2026. The Tiangong entry, built by the Beijing Innovation Center of Humanoid Robotics, crossed the line in 9.39 s. Lightning, entered by the smartphone maker Honor, led almost all the way and lost the race near the finish, clocking 9.47 s. Usain Bolt’s human record, set in Berlin in 2009, is 9.58 s — in this single heat two machines went under it. The number worth keeping is not the comparison with Bolt but the one with last year. At the inaugural games in 2025 the best time over this distance was 21.50 s. Twelve months later the leading machine covers the same 100 m in less than half of that, more than 12 s quicker. What the results sheet does not show is how the heat ended. The Unitree machine, third across the line, fell just past it and stayed down; organisers had lined the run-off with padded barriers for exactly this reason. Sprinting is a control problem with a beginning and an end, and only the beginning has been solved: these robots can now accelerate harder than a human, but stopping on their own terms is still beyond some of them. That is a fair description of the state of humanoid locomotion in general — peak performance is arriving faster than the ability to fail safely. The games run from 22 to 26 August and gather 666 teams with 2,056 robots from 16 countries, across 51 events. Of those teams, 641 are Chinese.
Tiangong Ultra →Beijing's robot games quadrupled the machines on the floor — and 25 of the 666 teams come from outside China
The second World Humanoid Robot Games open in Beijing on 22 August and run to the 26th, at the National Speed Skating Oval. The organisers put the field at 2,056 robots entered by 666 teams from 16 countries on six continents. Against the first edition — 280 teams and just over 500 machines in August 2025 — that is 138% more teams and about four times as many robots. The same briefing contains a second set of numbers that is quoted far less often. Of the 666 teams, 641 are domestic: 157 Chinese companies plus 200 universities and research institutes. They bring 1,975 of the 2,056 robots. Subtract those and the entire foreign presence at the world championship comes to 25 teams and 81 machines — 3.8% of the field and 3.9% of the robots. Spread across the fifteen countries other than the host, that is under two teams each. The count of countries is the figure that did not move. The 2025 edition also drew 16. Everything that grew this year grew inside China: the number of entrants more than doubled, the number of nations stayed exactly where it was. The growth is also lopsided in a way worth naming. In 2025 the average team brought 1.8 robots; this year it brings 3.1. Teams are not merely more numerous, they arrive with deeper benches — which is what a manufacturing base looks like when spare units stop being precious. One caution on the headline growth rate. The deputy executive director, Jiang Guangzhi, said the robot count had doubled. Set against the officially reported 500-plus machines of 2025, 2,056 is a fourfold rise, and China Daily itself put the increase at 311% in a follow-up piece five days later. We use the underlying counts rather than either summary. The programme has widened too: alongside track and field, football, gymnastics, street dance and tai chi, this edition adds long jump, weightlifting, tug-of-war and table tennis, plus scenario events in industrial assembly, logistics, retail, hotel and family service, and firefighting and rescue. The 2025 edition ran 26 events. Note that these Games are a separate event from the World Robot Conference trade show, which ran in the same city earlier this month.
China's robot makers needed six months to book 55% of a whole year's revenue — and 49 state firms now have their own hall in Beijing
The World Robot Conference opened in Beijing on 19 August 2026 and runs to 23 August at the exhibition centre in Beijing E-Town. Organisers report more than 300 exhibitors, 36% more than a year ago, filling 50,000 square metres across four halls with over 2,000 exhibits — among them more than 150 products shown publicly for the first time anywhere. The numbers quoted at the opening say more about the industry than the floor plan does. China's Ministry of Industry and Information Technology put robotics revenue at 300 billion yuan (about 44.45 billion US dollars) for the whole of 2025, after five years of average growth above 20% a year. For the first half of 2026 alone the figure is 165.5 billion yuan, up 24.5% year on year — in six months the sector booked roughly 55% of what it took all of last year to earn. One change at this edition is easy to miss and hard to overstate: for the first time the show has a dedicated zone for central state-owned enterprises, 49 of them, presenting 12 application scenarios. Robotics in China is no longer only a start-up story told in funding rounds; it is being staged as state industrial infrastructure, next to a consumer-facing "robot consumption street" where machines cook and serve. The programme itself has been split into themed days — a release day, a procurement day, a developer day and two public open days — which puts buyers and developers in the building on separate dates rather than in one undifferentiated crowd. Organisers also announced a free trial programme lending humanoids, quadrupeds and dexterous hands to innovation teams, and an initiative on intelligent robots "for the positive and for good" aimed at international cooperation. For a catalogue like ours the practical consequence is the 150 global debuts: the profiles worth writing in the coming weeks are being unveiled this week.
Xiaomi shows its humanoid in public for the first time — and says in the same breath it will not be a product
Xiaomi's robot will make its first public appearance at the World Robot Conference in Beijing, which runs from 19 to 23 August 2026. The announcement came not from a product launch but from the company's quarterly earnings call on 18 August, where president Lu Weibing said Xiaomi does not plan to develop the robot as a standalone product. The work is to be folded into the company's "human, car and home" ecosystem instead — smart factories, mobility and household applications. That sentence is the news. A public showing normally starts a sales countdown; here the maker closed that door before anyone could ask about a price. It also sets the frame for whatever is wheeled onto the stand this week: a demonstration of an internal automation programme, not a machine you will be able to buy. The last humanoid Xiaomi put in front of cameras was CyberOne in August 2022 — 177 cm, 52 kg, walking at 3.6 km/h, with roughly 1.5 kg of grip per hand. It never became a product either, and from 2024 it was repurposed into automation work inside Xiaomi's own plants. Four years on, the company has still published no spec sheet, no mass, no runtime and no price for its successor. What we are not writing: how tall this robot is, what it weighs or how well it performs on an assembly line. Those numbers are circulating in reports of the same call, but Xiaomi has not put them in a document of its own. Until it does, our profile of the company's humanoid work stays with the 2022 machine and its verified figures. We will open a separate profile for the new robot the day Xiaomi publishes its data — the catalogue keeps old versions and gives every new one its own entry.
Xiaomi CyberOne →