A humanoid robot has just run 100 meters in 9.39 seconds.
Another can play tennis against a human, fall over, get back onto its feet, and continue the rally.
Meanwhile, in a model apartment elsewhere in Beijing, robots are struggling with a rather less spectacular challenge: putting clothes into a washing machine and hanging them up again.
That contrast may tell us more about the state of humanoid robotics in 2026 than another polished manufacturer demo.
The second World Humanoid Robot Games are still underway in Beijing through August 26, so this isn’t a final verdict on the event. Instead, we looked at what the robots have already demonstrated.
The picture is becoming clear.
Humanoid robots have made enormous progress in locomotion, balance and autonomous movement. Manipulation, multi-step tasks and dealing with normal, imperfect environments remain much harder.
A Robot Just Ran 100 Meters in 9.39 Seconds
Tiangong Ultra, developed by the Beijing Humanoid Robot Innovation Center, completed 100 meters in 9.39 seconds during a heat on August 22. Honor’s Lightning followed in 9.47 seconds. Both times were below Usain Bolt’s 9.58-second human world record.
But the more useful comparison is with last year’s robots.
At the inaugural World Humanoid Robot Games in 2025, Tiangong Ultra won the same distance in 21.50 seconds. Going from more than 21 seconds to below 10 in roughly a year is extraordinary progress.
The 2026 rules are also stricter on autonomy. Except for the 100m and 400m hurdle races, all track events and other competitive events must now be completed fully autonomously, according to the organizers.
That matters more than the Usain Bolt comparison.
The machines aren’t simply being steered down a track. They are increasingly doing the perception and movement control themselves.
Fast Doesn’t Yet Mean Fully Controlled
Some of the fastest robots have another problem: stopping.
Several sprinting robots crossed the finish and relied on padded barriers beyond the line rather than decelerating neatly like a human runner.
So yes, a robot can now cover 100 meters extremely quickly.
Stopping cleanly is another engineering problem.
The speed continued over longer distances. On August 23, Tiangong won the large-robot 400m final in 38.15 seconds.
The interesting achievement isn’t really that robots can “beat humans.” Cars can do that too.
What matters is how rapidly dynamic balance, joint control and autonomous locomotion are improving in machines that still have to remain upright on two legs.
Tennis May Tell Us More Than Running
One of the more revealing moments came during a tennis demonstration.
A Galbot humanoid was playing against human opponents, moving sideways, tracking the ball and returning forehands and backhands.
Then it fell.
While chasing a shot, the robot lost its balance and toppled backwards. Instead of waiting for technicians to reset it, it rolled over, pushed itself back onto its feet and resumed playing.
For a useful humanoid, that may matter more than a perfect sprint.
A robot working in a warehouse, hotel or eventually a home will make mistakes. Floors vary. Objects move. People get in the way.
A machine that can recognize a failure, recover and continue without human intervention is much closer to being useful.
The tennis itself is also demanding. The robot has to detect a fast-moving ball, predict where it is going, reposition its body and coordinate its arm quickly enough to return it.
Galbot says its robots managed more than 100 consecutive autonomous rallies. We would treat that as a company claim rather than an independently verified competition record, but the public demonstration still shows how quickly perception and whole-body control are developing.
Then We Get to the Laundry
Running is getting fast. Tennis is starting to look surprisingly competent.
Now ask a robot to tidy a room.
This year’s Games include 21 scenario-based events, up from six last year. They cover household work, hospitality, emergency response, industrial tasks and other attempts to move robots away from controlled demonstrations.
There are also precision tests involving jobs such as picking up beans and tightening screws.
German broadcaster ZDF observed part of the household competition in a real smart-home showroom with furniture, a kitchen and a washing machine.
The robots had 30 minutes to complete a sequence including:
- tidying the living room;
- folding clothes and putting them away;
- loading dirty laundry into a washing machine;
- hanging clean clothes;
- receiving a package at the door.
During the roughly five hours ZDF observed the multi-day competition, none of the robots completed the entire sequence.
It is difficult to find a better snapshot of humanoid robotics in 2026.
A robot can cover 100 meters in 9.39 seconds.
Give it a pile of clothes and a washing machine, and things become much more complicated.
Why Is Laundry So Difficult?
Because the apparently simple task is actually many problems stacked together.
The robot has to identify objects, work out where they belong, navigate through the room and manipulate soft materials that constantly change shape.
A T-shirt doesn’t behave like a rigid factory component.
Then there are doors, handles, positioning and sequencing. If something sits a few centimeters away from where the robot expected it, the machine may have to perceive the difference and change its approach.
The Games’ dexterity events expose the same problem on a smaller scale.
Robots are being asked to use tweezers to pick up beans, open containers, weigh powder and manipulate small objects precisely. Beijing’s own reporting notes that even small errors in the angle of a spoon or the speed at which a robot pours can ruin a task.
That is the gap between a robot that can perform a skill and one that can reliably do a job.
Autonomy Matters More This Year
The 2026 Games also make it easier to distinguish what the robot itself is doing from what a remote human operator is doing.
As noted above, most competitive events now require full autonomy.
Scenario events still permit teleoperation in some cases, but the scoring explicitly favors autonomy: a fully autonomous performance receives a 1.0 weighting, while a remotely operated performance receives 0.5.
That distinction is important.
A remotely controlled humanoid can produce an impressive demonstration without showing much autonomous intelligence.
A robot that perceives its surroundings, decides what to do, acts and then recovers when something goes wrong is solving a much harder problem.
What the Games Show So Far
Based on what has happened during the opening days, the state of humanoid robotics looks roughly like this:
| Capability | State in 2026 | What we have seen |
|---|---|---|
| Running | Very advanced | 100m in 9.39 seconds |
| Dynamic balance | Advanced but imperfect | Very high speeds, but stopping and falls remain issues |
| Fall recovery | Improving quickly | Tennis robot gets itself back up |
| Autonomous locomotion | Advanced in structured environments | Most competitive events now require it |
| Fast perception and reaction | Improving quickly | Autonomous tennis and ball tracking |
| Fine manipulation | Still difficult | Bean, screw, powder and container tasks remain meaningful tests |
| Handling soft objects | Difficult | Folding and moving clothes remain challenging |
| Multi-step household work | Early | Full household sequences remain difficult |
| Error recovery in complex tasks | Major bottleneck | Small failures can still derail an entire sequence |
| General-purpose humanoid worker | Not there yet | Strong individual skills don’t yet combine reliably |
The striking part is how uneven the progress is.
An individual ability can suddenly look extremely advanced. Combine perception, manipulation, planning and recovery in a messy environment, and reliability falls quickly.
The Real Test Is Whether the Skills Work Together
A robot can be optimized heavily for a particular task.
Running quickly is one problem. Lifting something heavy is another. Returning a tennis ball can be trained as a specialized perception-and-motion task.
A useful general-purpose humanoid has to connect everything.
It may need to enter an unfamiliar room, understand an instruction, identify the correct object, grip it properly, carry it elsewhere, notice that something has changed and adapt without an engineer stepping in.
That is why the household competitions are so revealing.
The organizers have expanded scenario events from six in 2025 to 21 in 2026, and several now take place in real settings such as factories, hotels and model homes rather than purely simulated competition spaces.
The question is shifting from can this robot perform an impressive skill? to can this robot actually work?
Why the Games Are Useful
There is still plenty of spectacle.
A total of 2,056 robots from 666 teams are registered for 51 events this year, ranging from athletics and football to martial arts and weightlifting.
The event is also clearly a showcase for China’s rapidly expanding robotics industry.
But public competition has one major advantage over a promotional video: the robots are allowed to fail where everyone can see it.
They fall over. They miss objects. Sometimes they cannot finish the task.
Those failures are useful.
The Games still have several days to run, so the picture may change. But what we have seen already is enough to make one point clear.
Humanoid robots are becoming remarkably good at some individual physical tasks.
Combining those skills into reliable everyday work is still much harder.
Running fast has become an engineering achievement.
Doing the laundry is still robotics research.
