China’s Humanoid Robots Outrun Bolt and Topple Into Flames at Beijing Games

Robots lost legs mid-sprint. They slammed into padded barriers and erupted in sparks. One caught fire after crossing the finish line. Another toppled a judges’ table during weightlifting. Yet the crowds cheered. Records shattered. And China left no doubt about its commanding position in the race to build capable humanoids.

The second World Humanoid Robot Games wrapped up last week in Beijing’s National Speed Skating Oval. Organizers fielded 2,056 robots from 666 teams across 16 countries. That marked a sharp jump from the inaugural 2025 edition, which drew roughly 500 machines. This time the schedule expanded to 51 events and 1,301 matches. Thirty focused on competition. Twenty-one tested practical scenarios such as hotel housekeeping, factory assembly, retail restocking and emergency response.

The Verge captured the chaos. One Honor-built sprinter shed a limb during its run. Finishers barreled past the line only to collide with safety mats, bounce backward and ignite. Cheerleader bots stumbled through routines. A high-jump competitor lost balance and crashed. The spectacle mixed engineering triumph with slapstick failure. But the failures themselves generated valuable training data.

Progress proved unmistakable. Last year the winning 100-meter time clocked a sluggish 21.50 seconds. This August one machine posted 9.39 seconds in preliminaries. Tiangong Ultra, developed by the Beijing Humanoid Robot Innovation Center, later lowered the mark to 8.86 seconds and then 8.64 seconds in the final. That beat Usain Bolt’s human world record of 9.58 seconds by nearly a full second. The same platform ran the 400 meters in 38.15 seconds, five seconds faster than any human mark.

Standing high jump records leaped even farther. The 2025 best stood at 0.95 meters. Tiangong Ultra cleared 2.8843 meters this year. Some reports placed the peak demonstration near 3.40 meters. Either figure dwarfed the human record of 2.45 meters. Long jump distances stretched from 1.25 meters to more than 4.8 meters in certain categories. The 1,500-meter times dropped dramatically. Autonomy rules tightened too. Most track events, including the 400-meter, 1,500-meter and relays, ran fully without remote control. Time limits shrank. The 100-meter window tightened from three minutes to one.

China dominated the medal table. AgiBot and teams tied to the Tiangong platform each collected more hardware than nearly every other competitor combined. AgiBot’s Genie G2 took gold in pitch-pot, a precision contest that demanded delicate hand-eye coordination to land arrows in a narrow vessel 1.5 meters away. Yobotics’ Xingzhe Taishan won the lightweight tug-of-war title. Its engineers taught the machine to read an opponent’s pull, adjust posture and apply steady counterforce. The X-Humanoid-Huake joint team secured heavyweight weightlifting with a 16-kilogram overhead lift and earned silver in tug-of-war.

Global Times quoted Zhou Tong, an algorithm engineer at Yobotics. “We taught the robot to pull on its own — to read the opponent’s force, adjust its posture, and then deliver force steadily.” Ding Gang from the Beijing center described the dual demands. “Weightlifting is static load capacity, while tug-of-war is dynamic interaction. Your upper body takes a heavy force, but your lower body cannot simply give way.”

These contests pushed machines toward opposite poles. Precision tasks such as pitch-pot and micro-manipulation — picking beans with tweezers, driving screws, sorting pills — honed fine force control. Power events built stability under load. Organizers framed both as direct preparation for factory floors, construction sites and rescue operations. Peg-in-hole assembly in manufacturing mirrors the finger dexterity shown in arrow-throwing. Carrying 10- to 20-kilogram loads during material handling echoes the tug-of-war strain.

Yet the games exposed persistent gaps. Many robots still required teleoperation in select events, though autonomy scored higher. Mundane chores produced fewer viral moments. Robots spilled water while trying to pour. They fumbled small objects. Shelf-stocking and bed-making proceeded in fits and starts. Wang Xingxing, founder of Unitree Robotics, told observers that limitations in underlying AI models remain the biggest bottleneck. Mass deployment in useful settings could still lie a decade away, he suggested.

The Guardian reported the mixed emotions. Viewers laughed at pratfalls while some expressed unease at the speed of advancement. One expert called the engineering gains “staggering.” Another noted that while stunt capabilities impress, the path to genuine utility stays steep. Liquid cooling systems have allowed Chinese teams to boost motor torque, producing faster, twitchier limbs. That hardware edge helped shatter records. But reliable perception, error recovery and generalized decision-making lag.

Scale tells its own story. China produced about 12,800 humanoids last year, roughly 90 percent of the global total according to the Mercator Institute for Chinese Studies. This year’s games featured 641 Chinese teams representing 157 companies and 200 universities. International participation existed — teams from the United States, Germany, Japan and a Brazilian RoboCup squad appeared — but remained limited. Major American developers such as Tesla, Boston Dynamics and Figure skipped the event. The absence underscored differing national priorities. Beijing has made embodied AI a strategic focus. Washington emphasizes frontier models and chips yet trails in hardware manufacturing and deployment volume.

Recent coverage reinforces the momentum. A TechTimes report on the closing ceremonies confirmed AgiBot’s overall medal lead with 46 total honors while Tiangong Ultra claimed the final 100-meter crown at 8.64 seconds. Beijing government updates highlighted the jump in participation and multiple broken records across sprint, jump and relay categories. State media emphasized transferability to real work. Robots that mastered hotel room tidying or industrial cable plugging could soon appear in commercial settings. Plans for a robot-run hotel next year and another half-marathon in 2027 already sit on the calendar.

Failures generated the most shared clips. A sprinter losing a leg. A weightlifter crashing backward. Robots flailing as cheerleaders or freezing mid-obstacle. Those moments supplied terabytes of training footage. Engineers study every tumble, every misjudged force, every balance failure. The data loop accelerates improvement. What looked comical on social media represented raw material for the next iteration.

So the games served dual purposes. They functioned as spectacle and propaganda. They also acted as a brutal, public stress test. Organizers compressed time limits and stripped away remote controls to raise the bar. Machines responded. Track times improved by factors of two or more within a single year. Dexterity contests grew more demanding. Scenario events simulated actual workplaces rather than laboratory tracks.

Analysts watching from outside China note the hardware lead. One observer remarked that while brains may originate in American labs, the bodies are increasingly made in China. Liquid-cooled actuators, high-torque motors and refined joint designs give domestic teams an edge in physical performance. Yet software that lets a single platform move fluidly from sprinting to folding laundry to fighting fires remains elusive. Current models excel at narrow tasks. Generalist behavior stays distant.

Beijing treats the entire effort as national policy. Government backing flows to both established players and university labs. The diversity of entrants — from startups to state-linked institutes — creates a thick talent pool and rapid iteration cycle. Circular investment worries and questions about viable business models persist. Some companies produce impressive demos but struggle to ship products that solve customer problems at scale. The games themselves do not settle those commercial questions. They do demonstrate raw technical velocity.

Tiangong Ultra’s progression from 9.39 seconds on opening night to 8.64 seconds in the final captured the pace. Each heat brought visible refinement. Similar gains appeared in manipulation events. A robot that could thread a needle or assemble power tools one day might handle warehouse picking or elder care the next. The gap between viral athletic feats and profitable deployment is narrowing, if unevenly.

Critics still point to the pratfalls. Robots that outrun Bolt can still trip over a curb or drop a glass. Those errors comfort some viewers. They remind audiences that these machines are not yet overlords. For engineers they represent solvable problems. Every recorded failure shortens the distance to reliable autonomy.

China has signaled its intention to keep hosting the games. A third edition seems likely. Additional challenges such as full-distance marathons and expanded scenario simulations will follow. Each cycle will tighten rules, add complexity and harvest more data. The robots will stumble less. They will finish tasks faster. They will handle messier environments.

The Beijing games delivered drama, data and dominance. They also issued a clear signal. The hardware foundation for widespread humanoid deployment is forming faster than many outsiders anticipated. Whether the software and economic models catch up in time to matter remains the open question. For now the machines keep running — sometimes into walls, often faster than before, always under the watchful eyes of engineers who turn every fiery crash into tomorrow’s improvement.