August 27, 2026

A Robot Ran 100 Metres in 8.64 Seconds in Beijing: So What Does That Number Mean?

The second World Humanoid Robot Games closed in Beijing yesterday, and one number made the headlines: 8.64 seconds, faster than Usain Bolt's 9.58. Let us go through where that number came from, what it actually measures, and what happened right after the finish line. The part after the finish line is at least as interesting as the record.

First, the names, because reports keep mixing them up. The event is the World Humanoid Robot Games. The robot that set the record is Tiangong Ultra. The organisation that built it is the Beijing Humanoid Robot Innovation Center, which also goes by X-Humanoid.

The Games by the Numbers

The Games ran from 22 to 26 August 2026 at the National Speed Skating Oval in Beijing, the building from the 2022 Winter Olympics that everyone calls the Ice Ribbon.

Do not skip past that last row, because those events answer the question people actually care about. The robots are not doing athletics there. According to the Associated Press they compete at real tasks: housework, hotel service, emergency response. Whether a humanoid is useful shows up in that column, not in the 100 metres.

The Robot That Set the Record: Tiangong Ultra

Tiangong Ultra stands about 1.8 metres and its reported mass is 55 kilograms, so it is the height of an adult and less than half the weight of one. Its published design points are low-inertia legs and high-power integrated joints, both of which are choices made for running long distances without shaking the structure apart.

The machine has a history worth knowing. On 19 April 2025 it won the world's first humanoid robot half-marathon in Beijing, covering 21.0975 kilometres in 2 hours 40 minutes 42 seconds. The winning human in that same race finished in 1 hour 2 minutes, and only four robots made it inside the four-hour cutoff.

Here are the three sprint times from this week, side by side.

Taking a time from 21.50 to 8.64 in twelve months is a 60 percent improvement and it is not a stunt. Running on two legs at that speed is a hard control problem, because on every stride the machine spends part of its time airborne and the rest balanced on one foot at the edge of falling over. Whatever the teams did to their controllers, their gearboxes and their thermal limits, they did it quickly.

What Happened After the Finish Line

The Associated Press reported that several of the runners, the champion among them, crashed into a padded barrier after crossing the line. Sparks came off some of them, and staff ran over with fire extinguishers to spray the machines lying on the ground.

This looks like a software failure and it is not one. As the arithmetic below shows, the distance needed to stop from that speed is far longer than the run-off behind a 100 metre finish line. The barrier is not an accident. It is part of the design.

Why Stopping Is Harder Than Running

Acceleration the robot produces for itself: the motors supply torque and the foot pushes on the ground. Braking runs into two separate limits, and neither of them is motor power.

The first limit is friction. The braking force comes from friction between foot and floor. Stopping distance is roughly the square of the speed divided by twice the deceleration. A robot covering 100 metres in 8.64 seconds averages 11.6 m/s. If it can decelerate at a full 1 g, which needs a friction coefficient near 1.0, roughly what a running shoe gets on a tartan track, it still needs about 6.8 metres of floor. Halve the friction and that becomes 13.7 metres.

The second limit is geometry, and on good ground it is the one that binds first. A walking machine can only decelerate as hard as it can plant a foot ahead of its own centre of mass, because a foot planted too far forward converts the braking force into a face-plant. Take a humanoid with its centre of mass 0.8 metres up, placing a braking step 0.15 metres in front of it. That ratio caps the average deceleration near 0.19 g, and killing 11.6 m/s then needs something like 36 metres. Lunge harder, land the foot 0.4 metres ahead, and you get 0.5 g and roughly 14 metres.

A warning about that table: the geometry figures are my assumption, not measurements of Tiangong Ultra. The real ones belong to the team that built it. But the order of magnitude does not depend on the details. The stopping distance is tens of metres and the space behind a finish line is not. That is why there is a barrier.

Why the Arithmetic Changes Off the Track

On a speed-skating oval the friction is high, known and the same everywhere, so geometry is what binds and the problem is well posed. The robot knows its own body and it knows the floor.

In a warehouse it inverts. There is a wet patch by the loading door, dust on the polished concrete, a steel plate somebody left out. Now friction is the term that moves, and the robot has to know its value before it commits to the braking step. So can it?

I Measured That Question on My Own Robot

This year I measured exactly this on a much slower wheeled robot in ROS 2, and the answer is not encouraging. The sensors available were wheel encoders and an IMU, that is an accelerometer and a gyroscope. The question: can the robot work out what the ground will give it from those two channels?

I set the test up in its sharpest form, putting two states with opposite consequences side by side. In one the robot is stuck on a slippery patch with its wheels spinning. In the other it is driving normally on solid ground.

I tried thirteen features. The best one had an effect size of 0.56, which counts as weak, and separated the two states at 0.552 balanced accuracy. Tossing a coin gets 0.500.

The reason came out of the measurement itself. The features are derivative-based, meaning they look at differences in acceleration. A robot stuck at constant wheel speed has zero acceleration. A robot cruising at constant speed also has zero acceleration. The two states look the same because the same thing is genuinely being measured. I then tried a terrain classifier, and it answered 8.3 seconds into a crossing that takes about ten, which is after the moment the decision would have been useful.

The size of the effect is worth writing down, because this is not a small error. On a two-degree ramp, the slippery arm at 0.02 friction produced a 245 percent odometry error. The same command on the grippy control arm produced 0.34 percent. The robot is not slightly confused about where it is. It is wrong by more than the distance it covered.

A running humanoid needs that same quantity, faster, and before the foot lands rather than after.

What to Watch For When Comparing These Numbers

The Bolt comparison. 8.64 is smaller than 9.58, true. But a time only means something together with the conditions that produced it, and almost none of the conditions match here: the surface, the start, the mass, the energy budget, and whether any of the run has to generalise to a floor the machine has never seen.

The high jump comparison. A robot cleared 2.88 metres, against 0.95 metres at last year's games. The year-on-year jump is the real content. The comparison made in headlines, against Javier Sotomayor's 2.45 metres, is a category error, because a standing jump and a jump with a run-up are two different events with different physics.

And one inconsistency I could not resolve. The same brand was running a half-marathon at an average of 7.9 km/h sixteen months ago. It now appears in the 100 metres at an average of 41.7 km/h. These are different races and very probably different configurations, since an endurance run and a sprint do not want the same machine. Even so, I could not find a technical account of how that five-fold jump was achieved in a source I trust. If I find one I will add it here.

None of this is me lecturing anyone. I published a slip result of my own at 0.869 accuracy with a 72 percent error reduction, and later had to mark the whole section invalid. The reason was that my robot had been driving nose down at 13.3 degrees, dragging the front edge of its chassis, and that dragging was producing the slip I thought I was measuring. Nobody asked me to quote that number out of context. I did that to myself.

The Measurement I Would Like the Games to Add

One event would tell the field more than the whole sprint programme, and it is cheap to set up.

Bring the robot to full speed, command a stop, and report two things: the stopping distance and the heading error once it is at rest. Then run the same test twice. Once on the competition floor, once on a surface the robot has never seen and has not been told about in advance. Publish both, the completed runs and the failed ones.

The gap between those two runs is the number that predicts whether a humanoid can work near people. Jonathan Hurst, a robotics professor at Oregon State University, put the general version of this to the AP: true progress will be measured in warehouses and factories, where humanoids work autonomously for hours and generate real economic value. A stopping test is one concrete way to start measuring it.

What I Could Not Verify

The autonomy rules decide how any of these times should be read, because a teleoperated run and a fully autonomous run do not measure the same thing. The reporting I can verify says the 400 metres, the 1,500 metres and the 4×100 metre relay ran as fully autonomous categories with no remote control. Several secondary sites say the same about the 100 metres, but I could not confirm that in a source I trust, so I am not claiming it either way.

Tiangong Ultra's 55 kilogram mass and 1.8 metre height also come from manufacturer material and secondary sources rather than an independent measurement.

Sources

Scale and results from the World Humanoid Robot Games record and Al Jazeera. The finish-line crashes, the final time, the scenario tasks and the Hurst quote from the AP report as carried by OPB. The half-marathon time from TechNode. My own numbers come from a ROS 2 study in a private repository, held private while a provisional patent application and a journal review are open; the measurement write-ups are here and here.

← Back to all posts