It came. It ran. It crashed.

The 2026 World Humanoid Robot Games rewrote their own rulebook this year, and the rewrite had one purpose: make autonomy the headline.

The changes are unusually explicit. The 100-metre time limit fell from three minutes to one; the 400m limit from 15 minutes to five; the 1,500m from 40 minutes to 15. Global Times reported that the track events were now to be completed fully autonomously, with no remote control, while scenario events gave an autonomous completion a score of 1.0 and a remote-controlled one 0.5. It is a decision about what the audience should see as the benchmark.

You do not need to add “no remote control” to a supposedly autonomous showcase unless remote control has been part of the story until now. Forbes’ John Koetsier made a similar point from the event schedule: autonomy is mandated for some of the simpler sporting tasks, while harder activities still allow a human at the controls. Huawei, a sponsor of the Games, put the share of fully autonomous events at more than 40 per cent. That means the games were not suddenly autonomous. They were curated so the autonomous parts became the headline.

I read the rule changes against the coverage, and the tell is what the rules don't touch. Associated Press photographers at the 2025 games caught operators crouched by their robots between events, swapping batteries and carrying the fallen ones off the track. The 2026 footage is not much different. Newsweek reported that almost every competitor left its event on a stretcher. None of that breaks a rule, because the rules only cover the one minute the race is now allowed to last. And the people who wrote them are not neutral: the Games are run by the Beijing city government and China Media Group, the state broadcaster. And Tiangong Ultra, the robot that took the sprint records, came from a lab that same government funds.

None of this means the runs were fake. Some were really autonomous. A bipedal robot running the 100 in nine seconds is a real feat. Just last year it took twice as long. The progress would have carried the week on its own. The organisers dressed it up anyway: they chose the events, rewrote the rules, and set the clock so the autonomous runs were the footage that went viral. Take the staging away and the achievement still stands. So why couldn't they leave it at that?

The most revealing part of the humanoid sprint was not the sprint. It was the wall.

Tiangong Ultra ran 100 metres faster than Usain Bolt's human world record. TechRadar reported the 8.86-second run and the collision with the padded barrier; Tom's Hardware clocked its top speed at 23.8 mph. The Associated Press later reported an 8.64-second final, when the same robot hit the barrier again and sparks flew.

That is an odd way to finish a record. The race measures how quickly a machine can accelerate and stay upright for 100 metres, then effectively hands the last part of the job to a wall. Several robots crashed at the Games. Newsweek reported robots breaking apart, catching fire, and nearly every competitor carried off on a stretcher. The viral footage made for a very efficient blooper reel, but the joke obscures the engineering problem underneath it.

Stopping a two-legged robot is not as simple as running the start sequence backwards. David Nield, writing for TechRadar, laid out the problem: a robot at full speed carries a lot of momentum, and bleeding it off on two small points of contact takes thousands of balance corrections computed at once. The motors that make these machines quick off the line are far worse at absorbing the forces braking sends back through them. Safe stopping, he writes, is still the unsolved problem for sprinting humanoids — which is why the organisers set the crash mats out in the first place.

I checked the record trail because the number itself was starting to look suspiciously slippery. In the space of one week, reports put the mark at 9.32, 9.39, 8.86 and finally 8.64 seconds, depending on which run and announcement they were referring to. AP’s final report settled on 8.64 seconds. The improvement is real, but so is the measurement problem: every number describes a run designed to end at the same place, with the stopping problem left outside the race.

That matters because humanoids are already being put into warehouses and factories, where there are shelves, workers, cables and other things that do not politely become foam barriers at the finish line. The spectacular capability on display was acceleration. The unglamorous capability still missing from the highlight reel is knowing when to stop.

Beijing held its second World Humanoid Robot Games last month — 2,056 robots, five days, a stack of broken world records and a fair few broken robots.
Underneath the spectacle, both of this week's stories are about the same move: deciding what the audience looks at. The rulebook made autonomy the headline; the race format made acceleration the record. The hard, unsolved parts like staying in control, staying upright once you stop were conveniently omitted.