Faster Than Bolt: How Humanoid Robots Are Redefining Speed
Updated: Aug 31
A robot just ran the 100 metres quicker than Usain Bolt ever did. What that does and doesn't prove about machines catching up to human limits.
The Race: Tiangong Ultra vs Usain Bolt
On the night of August 22, 2026, inside Beijing's National Speed Skating Oval, a white, headless robot called Tiangong Ultra lined up for a 100-metre heat at the second World Humanoid Robot Games. It trailed for the first half of the race, then closed the gap and crossed the line in 9.39 seconds. Usain Bolt's world record, set in Berlin in 2009, is 9.58 seconds. Tiangong Ultra beat it by nearly two-tenths of a second. Three days later, the robot went even faster, clocking 8.64 seconds in the final.
Honor's robot, Lightning, finished the same heat in 9.47 seconds, also inside Bolt's mark. In the days before the Games, Lightning had reportedly hit 9.32 seconds during a private trial, reaching a peak speed of 14.5 metres per second. That number never counted for anything official. It was practice, not competition.
Then came the part that actually went viral. Running fast was one problem; stopping was another. After its 8.86-second semifinal, Tiangong Ultra crashed into the stopping mat beyond the finish line and collapsed.
A year earlier, at the first Games, the winning 100-metre time was 21.50 seconds, barely faster than a fit adult jogging. In twelve months, the winning time fell by more than half.
The List Keeps Growing
The 8.64-second sprint was spectacular, but it wasn't an isolated display of progress. In 2026, humanoid robots have been pushing into a growing range of physical tasks, from long-distance running to jumping and even competitive table tennis.
In April, at the Beijing E-Town Half Marathon, an older, separate event from the Games, Lightning ran the 21-kilometre course in 50 minutes and 26 seconds, autonomously, beating the human world record of 57:20 held by Uganda's Jacob Kiplimo by close to seven minutes. A year before that, the fastest robot in the same race needed two hours and forty minutes, and only six of twenty-one entrants finished at all.
This time, more than a hundred teams brought 300 robots, and four of them broke the one-hour mark. Oddly, the fastest robot on the course that day wasn't even declared the winner: a remote-controlled Honor robot finished in 48:19, quicker than Lightning, but the title went to Lightning because the scoring rules favoured autonomous robots over remotely piloted ones.
Back at the Games, Tiangong Ultra also won a standing-jump event, reaching 2.88 metres, nearly three times the 95.6 centimetres that won the same event a year earlier.
In a sport played with a paddle instead of legs, Sony's table-tennis robot Ace, trained through reinforcement learning rather than hand-coded instructions, beat three of five elite human opponents under official international table tennis rules, in research published in the journal Nature.
And underneath all of it sits a quieter, more academic data point. A peer-reviewed comparison published in Frontiers in Robotics and AI found that modern electromagnetic and fluidic actuators already beat human muscle on raw speed, force density, and power density. The catch is in the fine print: that finding excludes the batteries needed to actually run them, which is where a lot of the advantage disappears.
The Fine Print: Did the Robot Really Beat Bolt?
None of this means Tiangong Ultra "broke" Bolt's world record in any sense that World Athletics would recognise. Its latest 8.64-second run was recorded inside a robotics competition, not a World Athletics-sanctioned athletics meet, with different rules, equipment and competition conditions.
Coverage sourced to wire agencies was careful to note that the result doesn't replace or threaten Bolt's actual world record, since the two events belong to entirely separate categories. Plenty of outlets ran "robot beats Bolt" headlines anyway, because it's the better headline.
The jump comparison is where the fine print matters even more. Reports comparing Tiangong Ultra's 2.88-metre standing jump to Javier Sotomayor's 2.45-metre world record are comparing two different events.
Sotomayor's mark is for the standard high jump, a running approach, then a Fosbury Flop over a bar. The robot's jump had no run-up; it measures the vertical distance from the ground to the lowest point of the robot's body at the top of a two-footed leap.
The real human record for that kind of jump, the standing high jump, belongs to the American Ray Ewry, who cleared 1.65 metres back in 1900, when it was still an Olympic event. Judged honestly, the gap between robot and human is still huge. It's just a different gap to the one that made headlines.
Even the half-marathon's scoring exposed the same problem from another angle. The fastest robot on the course wasn't the winner, because organisers decided that autonomy mattered more than raw pace. Someone has to decide what counts as a fair win, and right now, that someone is an engineer, not an athletics federation.
What Machines Still Can't Do
Set against all of this, humanoid robots remain oddly bad at things a seven-year-old manages without thinking. Rodney Brooks, a robotics researcher who co-founded iRobot and Rethink Robotics, has predicted that humanoid dexterity will still be "pathetic" compared to a human's well past 2036. Boston Dynamics' Atlas is one of the most advanced bipedal robots in the world, but impressive demonstrations of strength and movement don't mean it matches humans across every physical task.
Balance is a bigger problem than it looks from the outside. A wheeled robot is stable the moment it stops moving. A bipedal one has to spend power and computation constantly adjusting itself just to stand still, something the human body handles through the inner ear and a lifetime of unconscious practice. Ankles and wrists — the joints that give humans fine control on uneven ground and with delicate objects — remain some of the hardest parts of a robot to get right.
Energy tells a similar story. Battery life remains a major limitation for many untethered robots, with operating times varying considerably by model and task. UBTech, one of China's leading manufacturers, has described its own Walker S2 units as no more efficient than half of a human worker. Dexterous hands are so difficult to engineer that they reportedly account for close to a fifth of the total cost of building a robot like Tesla's Optimus.
Put another way: the actuators inside these machines may already out-punch human muscle in a lab test. The complete package — power source, joints, sensors, judgment, all of it working together in an unpredictable environment — still doesn't.
Why Humanoid Robots Are Getting Faster
Robots aren't producing these results because of one breakthrough. It's several arriving together: reinforcement learning that lets a machine train itself through trial and error instead of being hand-programmed for every motion, simulation software that lets that training happen safely before it's tried in the real world, lighter materials, and cheaper, more capable batteries and sensors than existed even three years ago.
China has also decided, deliberately, to make this a national project. A 2023 government directive named humanoid robots a strategic technology and called for a world-class domestic industry by 2025.
Since then, subsidies, tax breaks, and dedicated development zones have poured into the sector. Companies like Unitree, the Hangzhou firm behind some of the robots at the Games, now preparing for a public listing, and X-Humanoid, the state-backed lab that built Tiangong Ultra, have grown quickly on the back of that support.
Independent estimates suggest roughly 150 Chinese firms are now building humanoid robots, producing somewhere around 12,800 of them in 2025 alone.
None of this is happening in isolation. American and European companies are racing too, and the competition between them is its own story. But it's the engineering, more than the geopolitics, that explains the sprint times.
Where Humanoid Robots Actually Matter
Strip away the medals, and the more interesting question is where any of this becomes useful. Right now, the honest answer is: mostly factories. Tesla's Optimus robots have been sorting battery cells at the company's Shanghai plant since early 2025.
Figure AI's robots work alongside people at a BMW plant in South Carolina. Chinese automaker XPENG runs its own humanoids on its Guangzhou production line. These are narrow, repetitive jobs in controlled spaces, not general-purpose helpers.
China's government has set a public target of moving more than 10,000 humanoid robots out of demonstrations and into actual jobs in manufacturing, warehouses, retail, healthcare and emergency response by the end of 2026. Disaster response, specifically, is still mostly a demonstration rather than a deployment.
Boston Dynamics has shown Atlas moving through simulated rubble and operating tools, and South Korea's DRC-HUBO once completed a staged rescue course entirely on its own.
But the robots doing real search-and-rescue work today are still mostly wheeled or tracked machines, not humanoids, because a battery that dies within a couple of hours is a serious liability inside a collapsed building.
What Still Belongs to Bolt
Usain Bolt spent the better part of a decade training to run 9.58 seconds once, on one night in Berlin, in front of a stadium that knew exactly what it was watching.
Tiangong Ultra has now run 8.64 seconds, after first recording 9.39 seconds and then 8.86 seconds earlier in the Games. As far as anyone can tell, it felt nothing about either performance: no nerves beforehand, no relief afterward, no sense that it had done anything remarkable at all. It simply executed what its actuators and its training allowed, on a night when everything happened to line up, and then it ran into a wall because nobody had taught it how to stop.
Maybe that's the more honest way to read all of this. The robots aren't simply approaching the outer edge of what a human body can do. They're mapping a completely different set of edges — some further out than ours, some much closer in — and the two maps only occasionally line up.
Comments