In twelve months, humanoid runners in Beijing went from falling at the start line to beating the half-marathon world record. The feat is real. Its meaning is narrower than it looks.
On 19 April 2026, a humanoid robot named Lightning crossed the finish line of a half-marathon in Beijing in 50 minutes and 26 seconds. The fastest a human being has ever covered the distance is 57 minutes and 20 seconds, a record set by the Ugandan runner Jacob Kiplimo in Lisbon a few weeks earlier. The machine, built by the Chinese electronics company Honor, was almost seven minutes quicker, and it ran without remote control.
Footage of the race spread around the world. To many viewers it seemed to confirm what other videos had been suggesting: that robots have overtaken people. The truth is more specific, and more useful.
From stumbling to sprinting
The pace of improvement is the real story. Beijing held the first such race on the same date in 2025. Twenty-one robots started and only six finished. Several stumbled, wandered off course or simply lay down at the start. The winner, a machine called Tiangong Ultra, took 2 hours, 40 minutes and 42 seconds, a time most recreational runners would beat.
One year later, more than 300 robots from 112 teams took part, running in lanes alongside about 12,000 human participants. The winning time had fallen by almost two hours.
The rules were designed to push the technology. Robots could run autonomously or under remote control, but a remote-controlled robot had its time multiplied by 1.2. A remotely driven Lightning actually covered the course faster, in 48 minutes and 19 seconds, yet its adjusted time placed it behind its autonomous twin. About 40 per cent of the teams chose to run without a human driver.
The Robot Games
Four months later, Beijing hosted the second World Humanoid Robot Games. The first edition, in August 2025, had drawn 280 teams and about 500 robots, and its 100-metre race was won in 21.50 seconds. The 2026 edition, held from 22 to 26 August, was four times the size: 2,056 robots from 666 teams, competing in 51 events that ranged from football and table tennis to hotel service, firefighting and rescue.
The track times were startling. Tiangong Ultra won the opening 100-metre race in 9.39 seconds and later ran 8.86 seconds in a semifinal. Usain Bolt’s world record is 9.58. The robot’s makers, the Beijing Humanoid Robot Innovation Center, also reported 38.15 seconds for 400 metres, 2 minutes 21.64 seconds for 1,500 metres and a standing high jump of 2.88 metres.
The centre described these as four human world records beaten. Some caution is in order. The claim comes from the team that built the robot, and no athletics body ratifies robot performances. The high jump was a standing jump, set against a human record of 2.45 metres achieved with a run-up, so the two are not the same event.
What a race tests
Why would engineers who want to build factory workers spend their time on athletics? Because a race is an honest examination.
A promotional video can be edited. A robot can attempt a task fifty times and the company can publish the one success. A race, in public, on a fixed date, allows no second take. Motors overheat, batteries drain, joints loosen and the machine either finishes or falls.
In 2025 most of them fell. The improvement since then reflects real progress in the components that matter for any working robot: motors that deliver power without overheating, batteries that last, and control software that keeps a two-legged machine upright at speed. A robot that can run for fifty minutes can probably stand at a workbench for a shift.
The autonomy rule mattered too. By penalising remote control, the organisers forced teams to make their robots see the course and steer themselves, which is the capability industry most needs.
What a race does not test
Running is, in engineering terms, a narrow problem. The surface is flat and known. The motion repeats. Nothing has to be picked up, and nothing changes shape in the robot’s hands.
The tasks that would actually replace human workers are of another kind. Threading a needle, folding a shirt, wiring a plug, lifting a frail patient: these demand fine control of the fingers, a sense of touch and the ability to handle soft or unpredictable objects. Progress there is far slower. A machine that outruns a champion may still be defeated by a bag of crisps.
Even in the race itself, people were close at hand. Lightning struck a barrier near the finish and fell, and staff helped it up before it completed the course. The robots ran in their own lanes, apart from the human field, so the event was a time trial, not a contest between species.
A sport with a purpose
Robot sport is not new. The RoboCup football competition has run since the 1990s with the declared aim of fielding a robot team able to beat the human world champions by the middle of this century. What has changed is the scale, the public attention and the national investment behind it. China’s events are also a showcase for an industry the state has chosen to back, and the footage serves as advertising for the companies involved.
Older contests had other aims. The United States defence research agency ran a robotics challenge after the Fukushima nuclear accident, to test machines on disaster tasks, and Switzerland’s Cybathlon pairs disabled competitors with robotic aids.
None of that makes the achievements false. A humanoid that can run a half-marathon in fifty minutes did not exist two years ago. But speed over a measured course is evidence about legs and batteries. It says little about hands and judgment, and those are what most human work requires.
The sensible response to the videos is to be impressed by the engineering and unmoved by the conclusion drawn from it. The robot has learned to run. It has not yet learned to work.
