이미지: METAL LAB 생성
Summary
- Tiangong Ultra clocked 9.32 seconds in a 100m heat, faster than the human world record of 9.58 seconds
- Foam pads stood behind the finish line, and all three robots that finished the 100m heat had to be carried off on stretchers
- 21 of the 51 events are real-world tasks like household chores and factory work, and teleoperated scores count for half as much as autonomous ones
- 대회
- 제2회 세계 휴머노이드 로봇 운동회 · 2026년 8월 22~26일 · 베이징 국가스피드스케이팅장
- 규모
- 666팀 · 로봇 2,056대 · 16개국 · 51종목 · 1,301경기
- 100m 최고 기록
- 톈궁 울트라 9초32 (8월 23일 예선). 결승은 8월 26일
- 사람 세계기록
- 우사인 볼트 9초58 (2009년)
- 1회 대회 100m
- 2025년 우승 기록 21초50 (톈궁)
- 채점 가중치
- 완전 자율 1.0 · 원격조작 0.5
- 시나리오 종목
- 51개 중 21개 · 가정 청소·호텔 서비스·공업 생산·응급 구조 등 9개 분야
- 참가 구성
- 666팀 중 641팀이 중국 팀 (157개사 · 30개 성시)
9.32 seconds — but there's a foam pad behind the finish line
On August 22, the second World Humanoid Robot Games opened at Beijing's National Speed Skating Oval, the "Ice Ribbon." On the first day, in a 100m heat for large-frame robots, Tiangong Ultra from the Beijing Humanoid Robot Innovation Center (X-Humanoid) posted 9.39 seconds. The next day, it broke its own mark with 9.32 seconds. For comparison, Usain Bolt's human world record, set in 2009, is 9.58 seconds.
One thing needs to be said upfront: both times came from heats, not the final. The 100m final doesn't happen until August 26. So these numbers aren't winning times yet — they're qualifying times.
There's also something at this race that human sprinting doesn't have. A thick foam pad stands just past the finish line. Robots can't decelerate on their own, so they slow down by running into the pad. After the 100m heats wrapped up, all three robots that competed had to be carried off the track on stretchers.
From 21.50 to 9.32 in a single year
The Games first launched in August 2025. Back then, the winning 100m time was 21.50 seconds, and the winner was also from the Tiangong lineup. In other words, the same event and the same robot family cut its time by more than half in just one year.
| Event | 2025 (1st Games) | 2026 (2nd Games, heats) | Human world record |
|---|---|---|---|
| 100m | 21.50s (Tiangong, winner) | 9.32s (Tiangong Ultra) | 9.58s |
Robot records move this fast because the upgrade cycle looks nothing like a human athlete's. Human records shift generation by generation, but robots can swap out joint motor output, batteries, and control algorithms and see the results reflected the very next season. According to specs published by IT之家, Tiangong Ultra stands 180cm tall and weighs 52kg, and this round of improvements focused on joint torque, rotational speed, battery capacity, and a gait refined through reinforcement learning.
9.32 seconds works out to an average speed of 38.6 km/h. Alternating legs at that speed while holding balance and staying in your lane draws on the same underlying capability that delivery robots or factory robots need to navigate around people. But here's the catch: a human 100m dash includes accelerating on your own and stopping on your own as part of the event. Here, deceleration was outsourced to a foam pad. That's why this number can't sit next to Bolt's 9.58 as an apples-to-apples comparison.
The real rule change this year is the scoring
Last year, the 400m title went to Team Gaoyi Robotics using a Unitree H1, with a time of 1 minute 28.03 seconds. But that run was done via teleoperation. Unitree's founder said at the time that the robot did have autonomous walking capability, but the team chose teleoperation for speed.
This year's Games put real weight behind that distinction. Gong Xiao, deputy director at the China Software Testing Center, explained that "fully autonomous performance carries a scoring weight of 1.0, while teleoperated performance carries a weight of 0.5." The reasoning behind doubling the gap, she said, is that organizers want robots to become genuinely autonomous collaborators — not have a human-controlled score sitting on the same leaderboard as a robot's own decision-making.
For reference, this year's 400m final for large-frame robots went to Tiangong Ultra with a time of 38.15 seconds, the Games' first gold medal. That's faster than the human world record of 43.03 seconds, but since both the manufacturer and the control method differ from last year's 400m result, you can't string them together as "50 seconds shaved off in one year."
21 of the 51 events aren't races — they're job sites
Let's start with the scale of the event.
| Category | 2025 (1st Games) | 2026 (2nd Games) |
|---|---|---|
| Teams | 280 | 666 (+138%) |
| Robots | 500+ | 2,056 (4x) |
| Countries | 16 (5 continents) | 16 (6 continents) |
| Events | 26 | 51 (30 competitions, 21 scenarios) |
| Total matches | 487 | 1,301 |
| Duration / match units | 3 days, 6 units | 5 days, 9 units |
| Referees | 123 | 223 |
What grew was the number of teams and robots — not the number of countries. Participating nations stayed at 16, same as last year, and 641 of the 666 teams are Chinese (157 companies across 30 provinces and cities). It carries the word "World" in its name, but the actual makeup looks a lot closer to a domestic Chinese competition — worth keeping in mind when reading these numbers.
What did clearly change is the event lineup. Of the 51 events, 21 are scenario-based. Robots are scored on handling real tasks across nine categories: household cleaning, hotel service, industrial production, emergency response, library management, retail, office service, landscape patrol, and vehicle charging. Over 40% of the competition now judges "did it finish the job," not "how fast did it run."
Zhang Feng, Party Committee Secretary of the China Institute of Electronics board, said at the opening ceremony that "humanoid robots are evolving from performing on stage and racing on tracks to working in homes and factories." According to the 2026 Humanoid Robot Industry Development Report he presented at the World Robot Conference on August 20, China's humanoid robot shipments topped 40,000 units in the first half of this year.
The track winner and the production leader are different companies
In the 100m heat for large-frame robots, Unitree's robot finished third in its heat with a time of 12.41 seconds. This is the company sitting on the largest volume in the market — it announced on August 12 that its cumulative production of bipedal humanoid robots had passed 18,000 units.
Unitree said it scaled back its event participation to focus resources on R&D and delivery for mass-production models. Tiangong, by contrast, is a platform built by an innovation center backed by the Beijing municipal government, where posting strong results at competitions like this is itself part of the development mandate. Track records and production volume are answers to two different questions, so reading this track ranking as a market ranking would be a mismatch.
The robot that ran with its face covered
One more moment drew attention. Tiangong Omni, from the same innovation center, ran the 400m for small-frame robots with both arms raised in front of its face, swaying its hips side to side as it moved. In China, it earned the nickname "wulianpao" — "face-covering run." Tiangong Omni finished the event final in 45.66 seconds.
Han Gang, a researcher who worked on motion control algorithms at the innovation center, explained what happened. The robot was originally designed to walk with a human-like arm swing. But during reinforcement learning, it shifted into this posture on its own. Swinging the arms quickly concentrates load on the shoulder joints and raises the risk of overheating; keeping the arms close to the face relieves that burden and lets the robot push more output into the legs. In simulation, the team only steered broad targets like speed and heat management — the specific posture wasn't something the developers had planned or anticipated. For what it's worth, a rumor that the robot "learned the move by watching a little girl" was officially denied by the team.
It's a good illustration of how a design that started out trying to mimic humans can, through training, drift toward whatever works best for the machine. It also connects to a broader trend: generating massive amounts of motion data in simulation for training is becoming standard practice in robotics.
Editor's take
The number that will get quoted the most is 9.32 seconds, but the numbers that will actually move the industry are 1.0 and 0.5. Autonomous gets 1.0; teleoperated gets 0.5. Put that next to the fact that last year's 400m winner chose teleoperation for the sake of speed, and it becomes clear what the organizers were trying to fix. Every humanoid demo video draws the same skeptical question — "isn't a person controlling that?" — and this year's scoring system is the Games' answer to it.
Track records will keep falling. Joint motors and batteries get better every year. The bigger news, honestly, will be the day organizers no longer need to lay down a foam pad behind the finish line. What homes and factories actually need isn't top speed — it's the ability to stop and get out of the way.
One more thing worth watching. This event is hosted by the Beijing Municipal People's Government and China Media Group (CMG), among others, and it also functions as a showcase for an industry China has designated as strategic. That means records and scale get the big headlines, while failures tend to leave a smaller trace. This competition will become genuinely useful as an industry benchmark once we see how much of the scoring data from the 21 scenario events gets published — including how many robots simply failed to finish the task.
Comments