Faster than Bolt, failed at 15 kilos: What Beijing's robot games really show

On August 26, the humanoid Tiangong Ultra ran one hundred meters in 8.64 seconds, coming in almost a second under Usain Bolt's record. It was the third sprint record within five days at the second World Humanoid Robot Games in Beijing. The real story, however, was in the medal table. There, not the sprinter led, but AGIBOT from Shanghai with 46 medals from grasping and simulated work scenarios.

This means the industry is split into two rankings that barely intersect. How fast a machine can run on a flat track says nothing about whether it can deliver an order in a hotel corridor. The situation becomes explosive through a figure from the margins: 96 percent of the 2056 robots came from China.

Racing took place in the speed skating oval from the 2022 Winter Games. Already in the preliminaries on August 22, times of 9.39 seconds were recorded, and in the same race, Honor's Lightning robot also beat the human record. A year earlier, the winning time had been 21.50 seconds. Such jumps sound like a solved problem, and for this one discipline, that's true. However, it only applies to a flat track, a standing start, and a rulebook written specifically for machines. World Athletics won't recognize any of these times, and the organizers say so themselves.

How narrow this capability is became clear in a scene at the finish line. A robot crashed into the barrier, stopped, and caught fire, requiring track officials to extinguish it. They can run fast now. But they can't stop on their own yet.

Two days earlier, weightlifting had its debut, and of the four robots that started, one lost its balance at fifteen kilos on the barbell and tipped against the referee's table. What's telling is that a human is allowed to sit at the remote control here, but not in the sprint. The rules thus unwittingly draw an honest map of what's possible.

It was precisely in the work scenarios that AGIBOT collected its points. The Shanghai firm entered for the first time and won gold in hotel service, library operations, and emergency response, plus in the eight dexterity disciplines that were held for the first time this year. There, the machines had to assemble a screwdriver, weigh powder, or handle a bottle with its cap. However, a competition venue cannot determine whether the measured precision also holds up with a different box size or a misaligned component.

The fact that precisely this company is leading has a concrete background, as it has built over 15,000 robots by mid-2026 and is considered the industry's largest manufacturer by volume. Jiang Guangzhi, director of the Beijing Office of Economics and Information Technology and deputy head of the games, never concealed the purpose of the event. "Once robots master these final grips, they can convert medals into orders," he told the Xinhua news agency.

666 teams with 2056 robots from sixteen countries participated, compared to just 280 teams and around 500 machines the previous year. As foreign participants, the organizers name the United States, Germany, Japan, and a Brazilian selection of five RoboCup teams. Switzerland does not appear in these announcements.

Something did fall to local developers nonetheless. At the conclusion, the organizers released a dataset comprising over 2500 hours of operational data from twelve application fields, which companies and universities can obtain free of charge. The bottleneck in robotics is no longer motors, but data from real movements under messy conditions. One caveat remains, however, and it is not technical in nature: Every machine that competed in Beijing comes from a manufacturer subject to Chinese law, which requires the same care in data transfer and remote maintenance as any other networked system.

The third edition comes in August 2027, again in Beijing, and the organizers have already announced they will increase the proportion of work scenarios. Whoever shifts the focus away from the sprint thereby says that there's nothing left to gain on the track.