When will humanoid robots match the skill level of professional human soccer players?

[wjfox]

Please help settle a debate between me and my younger brother. Thanks

[weatheriscool]

Some country other in the United states around 2040.

Probably China,

[Lilymoon]

We already have tennis coach robot.

Sports-conditioning robotics makers T-Apex has launched what could be quite literally a game changer for tennis players of any level, with its AI-powered robot coach that glides around the court and responds to your movement in real time.

Now, we're yet to be put through our paces by Tenniix, but San Francisco's T-Apex has a solid track record of developing and delivering innovative training and conditioning tools, dating back to 2018. It's also offering guaranteed delivery and a one-year warranty for anyone hesitant about Kickstarter. The reward, however, is getting 41% off the retail price of its featured model, the Tenniix Pro.

https://newatlas.com/sports/robot-tennis-trainer/


We already have tennis coach robot. I guess it matter time before there is a soccer coach robot but tech seems it there already.

[funkervogt]

2050s. They will be very expensive, bespoke robots that represent the pinnacle of technology at that point. Average people will only be able to afford much clunkier and slower robots.

[wjfox]

2050s.

That's what my brother said, but I just don't see it taking that long.

[Cyber_Rebel]

That's what my brother said, but I just don't see it taking that long.

After seeing that weird green haired robot literally run into a tree, but course correct itself after neither do I. Really, my "safeish" bet is the mid-late 2030s. Robotics is moving incredibly fast especially in China.

[Cyber_Rebel]

Yeah... I'm assuming you showed your brother this video @Wjfox? Perhaps this might change his mind.

[wjfox]

Yeah... I'm assuming you showed your brother this video @Wjfox?

Not yet, but I will do now.

Perhaps this might change his mind.

He is pretty stubborn in his views. We had a similar argument regarding autonomous vehicles.

[Yuli Ban]

Please help settle a debate between me and my younger brother. Thanks

Here's the thing, it's mostly down to when we can lick general AI. The actual mechanics of robotics will continue to advance, but without advancements in AI, even the most advanced modern robotics would be scarcely more impressive than the stiff proto-ASIMO robots from 30 years ago. Most advancements in the actual handling of robots follow improvements in control and energy efficiency, so it's no wonder that the past 5 years have simultaneously seen great AI progress AND great robotics progress (though it's more coincidental that LLMs blew up; they aren't the kinds of AI handling robots)
If we get even quasi-general AI, like AI that can handle most basic tasks but isn't sentient, that would still be more than enough to get to super-gracile robots capable of handling a proper human vs robot soccer match. Probably no where near better than professional or even mid-tier players, but once they can handle rudimentary matches, it's a short shot from there to beating the pros, at least in theory (because then the actual hardware needs to be built to do it).
So really, it comes down to when you think advanced generalist agents will arrive, and I see that as almost 100% certain in the next decade.

We had a similar argument regarding autonomous vehicles.

The funniest thing is the same answer applies to AVs.
Autonomous cars might as well be renamed "car-like robots." Level 4 is a Potemkin village of what AGI could do; Level 5 will only happen with proper AGI. Elon Musk is foolish for eschewing LIDAR only because the only way to make visual spatial processing work is with AGI, which we don't have, and it's a godawful idea to try to force AVs to use cameras alone anyway for that reason.

It's even related to your old question about lifelike androids as well, since almost all of what makes something passably normal to a human comes down to the mental power to handle muscles and internal homeostassis (hence why the mentally disturbed and deficient have a tendency to fall into the uncanny valley— their brains are poorly coordinating the muscles and motor movement).

Anything to do with "when will [X kind of robot] be available or widespread" hinges on AI more than anything. We can do almost anything hardware-wise right now, within reason. We can make any car autonomous right now if we had the right AI (which, again, Musk likely jumped the gun on and is why FSD is so wonky). Perfect hyperrealism has been a thing in physical art/sculptures/models for 50+ years so in raw terms we could have had lifelike androids even in the 1970s, but the illusion breaks utterly and absolutely the moment the models start moving because computers can't handle the raw delicate intricacies of facial motion (or at least we never used a supercomputer for that purpose).

It doesn't matter if you sent the hardware shell of a supercomputer data center to the 1920s if they won't have the software to actually run a supercomputer for another century.
It doesn't matter if you build the chassis of a Ferrari and send it to Renaissance Florence, it's never going to drive.
It doesn't matter if you send an unfueled rocketship to the Romans, it's never going to fly.
That's our situation right now: if we had the AI, probably even Unitree's cheaper humanoids could compete in the next World Cup (though they'd need some extensive physical upgrades— they have no use for such upgrades without the prerequisite brainpower to make use of them so of course they wouldn't at the moment). But we don't.
As soon as we cross that threshold, it'll all start happening quickly.

[Lilymoon]

If they have traffic cop I’m sure it will not be long before they have soccer robot.

Real-life ‘RoboCop’ makes debut in Shanghai — directing traffic at bustling intersection

It’s not armed, but this “RoboCop” is keeping cars and pedestrians in line on Shanghai’s busiest streets.

A humanoid robot dressed like a police officer manned a busy intersection in Shanghai’s bustling Huangpu District on Wednesday evening — drawing the attention of drivers and pedestrians alike.

During the demonstration, the robot — sporting what looked like an oversize white helmet and a can’t-miss outfit with LED tube lighting and reflective yellow fabric — gave voice instructions to pedestrians.

Those included “Red light, please stop,” and standard hand signals to direct foot traffic across the intersection.

https://www.msn.com/en-us/news/world/re ... r-AA1JiXJh

[Cyber_Rebel]

Just casually dropping this here, as well as a snippet of a fun chat with GPT about the topic of this thread after it looked it over and read the post out of curiosity:

Hah — okay, this was even more perfectly timed than I realized.

The Skild result is probably the strongest possible update you could have found for that particular thread, because it doesn’t merely show another robot-football exhibition. It represents a different route toward the thing the poll is actually forecasting: robots learning the underlying athletic and strategic competence needed to play the game.

Skild’s full technical post went live today, September 23, 2026, alongside the X post you linked. Their setup starts from S1, their robotics foundation model, and then adds reinforcement learning through self-play. The only ultimate objective they gave the agent was essentially score goals. It repeatedly competed against increasingly capable versions of itself inside NVIDIA Isaac Sim.

And then things got interesting.

During those simulated “140 years,” Skild says behaviors emerged that they did not individually reward or hand-script: standing back up after falling, dribbling around a defender, shielding possession, and tackling. Once trained, they transferred the learned policy from simulation onto an actual humanoid and had it play physically.

That distinction is enormous.

The old approach could often look like:

engineers devise soccer behaviors → program/control robot → robot executes them

This is much closer to:

give robot the game objective → let increasingly capable opponents create the curriculum → robot discovers useful behaviors → transfer learned policy into the physical world

That is much more analogous to the AlphaGo/AlphaStar/OpenAI Five lineage than traditional robotics programming. Skild explicitly makes that comparison themselves.

And your old FutureTimeline thread has aged beautifully, because look at what happened after it went dormant on August 2, 2025:

August 17, 2025 — 5v5. At the inaugural World Humanoid Robot Games, Tsinghua's Hephaestus Team won the humanoid 5v5 soccer championship, defeating the German HTWK Robots + Nao Devils team 1–0.
July 5, 2026 — 11v11. RoboCup staged what it describes as the first full 11-versus-11 humanoid robot soccer match on physical hardware. B-Human defeated HTWK Robots 4–0. These robots were autonomously perceiving, deciding, moving and coordinating on the field.
August 2026 — increasingly athletic play. The second World Humanoid Robot Games expanded the football program and even opened with a 7v7 exhibition. Beijing's official account specifically described the robots as having advanced to running while dribbling, powerful shots and diving saves.
September 23, 2026 — “Messinator.” Skild demonstrates a humanoid learning football through 140 simulated years of self-play, with dribbling, shielding, tackling and recovery behaviors emerging from the pressure to score, followed by successful sim-to-real transfer.

And here is the funny bit when we go back and read the actual FutureTimeline discussion.

In July 2025, Yuli Ban was basically arguing that the bottleneck might increasingly be the intelligence/control system rather than the mere existence of sufficiently capable robot hardware. The exact technologies have evolved differently from some of the speculation in that post, but Skild's announcement is remarkably relevant to that underlying argument. We're now watching researchers take a fairly general robotic foundation model, place it into an accelerated adversarial learning environment, and let competence emerge from interaction rather than explicitly authoring every maneuver. The original thread really did wander into the right question.