The Finish Line Was a Wall

On the evening of 26 August 2026, inside Beijing’s National Speed Skating Oval, a bipedal machine named Tiangong Ultra crossed a painted line in 8.64 seconds. The distance was one hundred metres. The time was nearly a full second faster than the human world record set by Usain Bolt in 2009. Then the robot kept going. It struck a thick padded barrier placed several metres beyond the finish, folded at the waist in a shower of sparks, and was carried off on a stretcher. Nearly every other finalist followed the same trajectory: high-speed impact, collapse, removal.

That sequence—record, crash, stretcher—became the signature image of the second World Humanoid Robot Games, the five-day event that organisers and media alike called the Robot Olympics. From 22 to 26 August, 2,056 humanoid robots from 666 teams competed across 51 disciplines at the Ice Ribbon arena and in staged workplaces around the city. Of those teams, 641 were Chinese. The rest came from fifteen other countries. The programme mixed Olympic-style sports with practical tests designed to measure whether machines that can sprint faster than any human can also make a bed, screw in fasteners, or extinguish a simulated fire.

The athletic numbers were startling even by the standards of a field that advances by leaps. Tiangong Ultra, developed by the Beijing Humanoid Robot Innovation Centre (also called X-Humanoid), lowered the 100-metre mark three times in five days: 9.39 seconds in a preliminary heat, 8.86 in the semi-final, and 8.64 in the final. The same family of robots recorded 38.15 seconds for 400 metres and 2 minutes 21.64 seconds for 1,500 metres. A standing high jump reached 2.88 metres in one report and 3.40 metres in another—well above the human record of 2.45 metres that has stood since 1993. Last year’s inaugural Games had produced a winning 100-metre time of 21.50 seconds and a high jump of less than one metre. The improvement was measured in orders of magnitude rather than incremental gains.

Yet the same footage that showed the new times also showed the cost of speed. Robots that could accelerate with remarkable coordination could not decelerate with equal skill. After the finish line they slammed into the crash mats. Some broke at the waist. Others emitted sparks or small flames that staff extinguished with handheld units. An Honor robot lost a leg mid-sprint. In the weightlifting debut a machine holding a modest 15-kilogram barbell toppled toward the judges’ table, arms raised in an unintentional cartoon of failure. Organisers had anticipated the problem; the mats and stretchers were part of the design. The spectacle remained irresistible. Clips of the collisions circulated widely, often paired with the record times that had immediately preceded them.

The sports programme was only half the competition. Twenty-one scenario-based events tested robots in environments modelled on factories, hotels, restaurants, libraries, offices, logistics hubs, and emergency sites. In a hotel challenge, machines had thirty minutes to wheel luggage to a designated room, restock towels and water, clear linen, and make the bed. In a library, they collected returns and shelved books correctly. A restaurant task required taking an order, heating food in a microwave, dispensing a measured volume of drink, and delivering the tray without excessive spill. An outdoor firefighting scenario asked robots to identify hazards, close three different types of valves, locate a fire, and use an extinguisher; only three of twelve teams completed the full sequence. Dexterous-hand contests demanded tweezers to pick up beans, precise screw-driving, powder weighing, and nail hammering without angle error. AGIBOT’s OmniHand, with sixteen degrees of freedom, took seven of the eight gold medals in that category. LinkerBot completed eighteen autonomous screw installations in five minutes to win the power-tool assembly event.

These tasks revealed a different hierarchy of difficulty. Locomotion over a flat, predictable track had advanced dramatically. Fine manipulation, task switching, and recovery from unexpected interruptions remained harder. Many robots still relied on some degree of teleoperation, though organisers emphasised that certain categories—particularly the longer races and selected scenario events—required full autonomy. The medal table reflected the split emphasis. AGIBOT, a Shanghai company strong in manipulation and scenario work, finished first overall with 46 medals (18 gold). Tiangong’s developer placed second with 45, having dominated the track events. A German team, B-Human, became the first foreign squad to win gold across the two editions of the Games, taking the medium-size 5-v-5 football final 13–3.

The Games functioned simultaneously as publicity, benchmark, and data-collection exercise. Organisers released a dataset containing more than 2,500 hours of operational recordings gathered during training and competition. Developers described the event as a public test bed for embodied AI—the combination of perception, decision-making, and physical action in unstructured environments. Spectators interviewed on site noted visible progress from the previous year: robots that had once looked stiff now moved with something closer to fluid coordination. Industry voices outside the arena were more measured. Progress on a controlled track, they observed, does not automatically transfer to warehouses or homes where floors are uneven, objects shift, and no crash mat waits at the end of a task.

The second World Humanoid Robot Games therefore closed on a double image. One image is the stopwatch: 8.64 seconds, a time no human has approached. The other is the stretcher and the extinguisher: machines that can generate that speed still struggle to stop, to recover balance under load, or to complete a multi-step domestic chore without human help. Both images are accurate. The gap between them is the current state of the technology, measured in public under lights, with the whole world watching the finish line turn into a wall.