Humanoid robots broke human world records in the 100-meter, 400-meter, and 1,500-meter races at the second World Humanoid Robot Games in Beijing, cutting the 100-meter time from 21.5 seconds a year ago to 9.32 seconds — then slammed into padded walls because they could not brake1,3,10.
The five-day event ran August 22–26 at Beijing's National Speed Skating Oval, drawing 666 teams and 2,056 robots from 16 countries across 51 events2,12. The competition split into 30 sports events and 21 scenario-based contests, with more than 40% requiring full autonomy, according to Huawei, a tech partner of the games11,13.
Honor's Lightning robot posted the fastest 100-meter time at 9.32 seconds, reaching a peak speed of 14.5 metres per second, according to China's state broadcaster. Tiangong Ultra, built by the Beijing Humanoid Robot Innovation Center, clocked 9.39 seconds in a preliminary heat on August 22. Both beat Usain Bolt's 9.58-second record set in Berlin in 2009. A robot of the same model as the 100-meter winner ran 400 meters in 38.15 seconds, beating Wayde van Niekerk's human world record of 43.03 seconds. In the 1,500-meter final, the top three finishers all came in ahead of the human world record of 3 minutes 26 seconds, with the winner finishing in 2:21.63. The high jump reached 2.88 meters, nearly triple last year's 0.95-meter mark and above the human record of 2.45 meters.
The speed came with visible limits. Tiangong Ultra stumbled off the track near the finish line. Lightning collapsed and was carried away on a stretcher. Organizers placed thick mats several meters past the finish line to absorb impacts.
ANALYSIS The gap between raw locomotion speed and reliable deceleration captures the state of the field: the hardware can generate force, but the control systems governing transitions — stopping, turning, recovering — lag behind.
The more consequential shift was structural. Organizers scrapped remote controls for most events, requiring robots to perceive surroundings, make decisions, and execute actions using onboard AI. Scenario-based contests tested robots in simulated factories, restaurants, hotels, hospitals, logistics centers, and emergency settings4. Dexterous-hand tasks required weighing 20 grams of powder within 0.5 grams, stacking blocks, picking soybean seeds with tweezers, and connecting cables — all within five minutes. Beijing-based Galbot staged what it called the world's first human-robot tennis match, with robots tracking balls, serving, and returning shots autonomously. During one rally, a robot lost its balance and toppled backward, then rolled over, pushed itself upright, and resumed play.
Guo Yijie, who leads the Tiangong Ultra project at the Beijing Humanoid Robot Innovation Center, cited improvements to the robot's body, joints, electrical reliability, and "cerebellum" motion control. Unitree, a leading humanoid maker whose founder Wang Xingxing watched from the stands, separately unveiled a robot called Superman that it says clears a 2-meter standing jump and hits 12.66 meters per second. Unitree's founder has said the humanoid robot industry is still a decade away from its "ChatGPT moment".
The investment backdrop is substantial. Bloomberg reported that Chinese investors have put at least 100 billion yuan (about $14.8 billion) into the humanoid robot sector this year. Unitree closed its Shanghai STAR Market debut up more than 460% after raising about $900 million, lifting its market value to roughly $66 billion. The Beijing municipal government has established a 10 billion yuan investment fund for robotics, expected to attract over 53 billion yuan in additional private capital. China's Ministry of Industry and Information Technology projected full-year humanoid robot production will exceed 100,000 units, after the country shipped more than 40,000 in the first half of the year — 97% of global output, according to the Chinese Institute of Electronics.
Zeroth chief marketing officer Hua Rong framed the gap between demonstration and deployment: "The real test will come after the demonstrations end... whether the product can really solve users' problems after it is sold".