A Robot Ran 100m in 9.39s. It Has No Idea How to Stop.

TL;DR

On August 22 in Beijing, a humanoid robot called Tiangong Ultra ran 100 meters in 9.39 seconds at the second World Humanoid Robot Games. Usain Bolt's world record is 9.58. Honor's robot Lightning went 9.47. Twelve months ago, at the first edition of the same games, the winning 100 meter time was 21.50 seconds, set by the same robot lineage. That is the result worth reading: not that a machine beat Bolt, but that the machine got 2.3x faster in a year. The detail that did not make the headlines is that none of these robots can stop. They cross the line at full speed, hit a wall of crash pads, and get carried off.

The times

Tiangong Ultra is built by the state backed Beijing Humanoid Robot Innovation Center, which trades as X-Humanoid. The 9.39 came in a preliminary heat on opening day. Honor, better known for phones, took the final in 9.47 and had already clocked 9.32 in a pre-games trial at a peak of 14.5 meters per second. For scale, Bolt's peak in Berlin was around 12.3 m/s. Nobody is quietly nudging a record here.

The same games produced a 2.88 meter high jump, against Javier Sotomayor's 2.45 from 1993. Treat that one carefully: robot event rules are not World Athletics rules, and none of these marks touch the official human record books.

2.3x in twelve months is the real number

Going from 21.50 to 9.39 on the same track, in the same event, with the same robot family, is a genuine capability jump. It is also a fair benchmark in one important respect: the games tightened their rules this year. Flat track events now have to run with no remote control and no human intervention, with the 100 meter and 400 meter hurdles the named exceptions. Last year plenty of entries were driven by a person with a tablet, and the organizers papered over it with a handicap, giving autonomous robots a 0.8 time coefficient. One autonomous machine physically finished third in a race and was declared the winner.

So the 9.39 is a robot that saw the track and ran it. The engine underneath is reinforcement learning: X-Humanoid runs Tiangong on an embodied intelligence platform it calls Huisi Kaiwu, with a locomotion controller the center describes as state memory based predictive reinforcement imitation learning. Strip the branding and it is the standard modern recipe, the same one you can read in open work like Humanoid-Gym: train a gait policy for millions of episodes in simulation, randomize the physics until the simulator is wrong in every direction at once, then transfer the policy to a real chassis and hope reality lands inside the distribution you randomized over.

That last step is the hard part, and the honest way to picture it is learning to drive entirely in a racing game. Everything transfers except the parts that matter, so most of your effort goes into making the game wrong in exactly the right ways, wrong friction, wrong motor lag, wrong sensor noise, until the real road is just another level you already played.

The finish line is a wall

Here is what a 9.39 second run actually consists of.

Reporters at the venue watched robots slam into padded barriers past the finish and get lifted off the track. Others tripped, fell, came apart, and in at least one case caught fire. Bolt decelerates by jogging a victory lap. Tiangong Ultra decelerates by hitting a wall.

This is not a manufacturing defect, it is the reward function. Braking on two legs from 14 m/s is a harder control problem than accelerating, because you have to dump kinetic energy through the same two contact patches that are keeping you upright, and you have to do it while your center of mass is already ahead of your feet. Nothing in the race scores that. So nothing in training optimized for it.

What the games certify, and what they do not

Autonomy on the flat is real, and that matters more than the stopwatch. But the comparison to Bolt is loose in ways worth naming:

• The hardware is a free variable. Al Jazeera reported that Honor lengthened Lightning's legs to 1.05 meters before the games, on a 169 cm frame. Sprinters cannot bolt on longer shins the week before a final.
• Only the scored window counts. Start, run, stop is the actual task. The event scores the middle third.
• Autonomy is event by event. Obstacle work and weightlifting still allow teleoperation. Flat running and soccer do not. "Robot did X autonomously" is a claim you have to check per event.
• Nothing here is deployed. Every wire report carried the same expert caveat: humanoids remain demo and research hardware.

Why a builder should care

Because this is a benchmark story, and you already know how benchmark stories end.

When a metric covers a fixed interval, when you control the hardware the metric runs on, and when a national industrial policy is watching the leaderboard, the metric moves. It moved 2.3x in a year. What did not move is the thing you would actually need a humanoid to do, which is finish a task and then be standing afterward, ready for the next one.

Swap the nouns and this is the argument you have every week about model evals. A score on a bounded, scored, publicly ranked interval is a real signal about that interval and a weak signal about anything past it. The robot that runs 9.39 and then needs a stretcher is the most literal version of a benchmark you will ever see.

The useful counter-signal in Beijing is not on the track at all. It is the scenario events, industrial assembly, restaurant service, plugging in an EV charger, a chunk of which now require the robot to work things out on its own. Those numbers are boring and nobody will put them on a chyron. They are the ones to watch. Meanwhile Unitree, X-Humanoid and the rest keep shipping open assets around the edges, like X-Humanoid's RoboMIND manipulation dataset, which is where the transferable work actually lives.

Key Takeaways

• 9.39 seconds is real and it was autonomous. Flat track events at the 2026 games ban remote control, a genuine tightening from last year's teleoperation plus a 0.8 handicap coefficient.
• The year over year jump matters more than the Bolt comparison. Same event, same robot family, 21.50 to 9.39 in twelve months.
• The robots cannot decelerate. They finish into crash pads and get carried off, because braking is not in the scored interval and therefore not in the reward.
• The hardware is a free variable. Honor reportedly lengthened its robot's legs before the games. That is not a race against a human, it is a different sport with a shared unit of measure.
• Read it as an eval lesson. A bounded, ranked, publicly scored interval will be optimized. Judge deployment readiness on the unscored parts.
• Watch the scenario events instead. Assembly, service and EV charging tasks, a large share of them now autonomous, tell you far more about real capability than the sprint does.

Sources: Associated Press, CBS News, Al Jazeera, Engadget, CGTN, Xinhua via Yahoo, The Robot Report, World Humanoid Robot Games