Learning athletic humanoid tennis skills from imperfect human motion data
Comments
ilaksh
mplappert
That seems like quite an extrapolation and an extraordinary statement. This is a single task, in a lab setting. What your describing are extremely open-ended tasks in people’s homes.
What is informing these timelines?
ilaksh
Look at recent developments/announcements involving novel increasingly generalizable learning capabilities from projects like 1X/Neo, Figure 03, Skild AI. Also see open published work like MimicDroid, HDMI, GenMimic, Humanoid-Union Dataset, RoboMirror, Being-H0
Figure 03:
https://www.youtube.com/watch?v=e-31-KBBuXM
https://www.youtube.com/watch?v=ZUTzuhkDG3w
1X Neo:
https://www.youtube.com/watch?v=lS_z60kjVEk
Skild AI
https://www.youtube.com/watch?v=YRmjBdKKLsc (Learning by Watching Human Videos)
mplappert
Yeah those are demos. I think we‘re pretty far away from this becoming a real thing. I wrote up why here: https://matthiasplappert.com/blog/2026/humanoid-robot-in-the...
ilaksh
You're not really trying to see the advances in things like the data flywheel. If you were you would see that those demos represent real movement towards generality.
mplappert
Where’s the data flywheel exactly?
jonkthomas
I know labs are throwing an insane amount of money wherever they can to try and build datasets for this - at claru.ai we're working with a few and they are even grabbing all the players like labelbox, toloka etc to throw armies at getting bespoke data for these needs. Most of the time its get data then figure out what to do with it after...the unfortunate strategy from what I'm seeing but data is a moat in their eyes
nmaley
" Do we have any standard benchmarks for humanoids to do domestic tasks?" The answer is yes. Steve Wozniak proposed the Coffee Test. See https://www.youtube.com/watch?v=MowergwQR5Y
It's actually very clever. Despite the apparent simplicity, no current model could pass it.
Re your forecasts, I think they are optimistic in terms of timing but not ridiculously so.
p1esk
I think coffee test for robots will be similar to Turing Test for LLMs, which was quietly achieved and forgotten somewhere between gpt-3.5 and gpt-4. Real tests are tasks like cooking or plumbing - I expect that to come in 2-3 years.
andai
The "AGI" (-ish) moment for AI was shoving Common Crawl into a transformer.
What's the animal intelligence (physical int.) equivalent of that? I don't think such a dataset exists? (e.g. NVidia is trying to compensate for that with simulated worlds, i.e. synthetic data)
impossiblefork
I think that's a bit too optimistic, but I still think the direction is right-ish. It feels hard to give a timeline though. Robotics is hard.
vasco
Hey guys I rooted my humanoid and it killed my mom when I disabled the "slow limb motion" mode. It just wacked her in the head as she walked in the kitchen and she's not moving what do I do??
SiempreViernes
You should @linustechtips and hope he picks it up, then you have a good chance of getting a voucher for the funeral and getting a shout out in the manufacturers next demo when they talk about their new safety features.
auggierose
>I'm guessing by next year or 2028 there will be services where you can order a robot to come cook and or clean for you.
Bullshit.
ohyoutravel
omg Elon Musk posts here! Are we also going to get full self driving, no interventions from NYC to LA within this timeframe, sir???
p1esk
I’ve been thinking about this too, and I share your optimism: https://news.ycombinator.com/item?id=47213310
ordu
It is interesting to watch. The movements of the robot are robot-like. I mean, wtf, there were no robot playing tennis before, but I have an idea how a robot playing tennis would be like, and this video confirms my expectations. Sharp, unsure movements, a lot of hesitation, ...
Movies pictured robots like this long before this become possible, but how did producers guessed it?
Or maybe movies rendered different kinds of robots, but this video bring into my memory only those, that look like this. A kind of confirmation bias?
rob74
Also, I can't help noticing that the guy(s) playing with the robot are doing their utmost to make it look good: playing the ball as gently as possible (so it has time to react and doesn't have to exert too much force to return it), aiming for places the robot can comfortably get to etc.
thethirdone
I agree that the movements look quite robotic (though not as much as you might expect), but I don't think any movies have depicted robots moving like that. A much more common depiction is moving only a single joint at a time.
> Sharp, unsure movements, a lot of hesitation, ...
I like these particular descriptors. Another I would add is holding poses unnaturally still. While waiting for the ball, the robot holds its racket extremely consistently relative to its body even while sharply turning.
KolmogorovComp
Nothing constructive to say, besides that the video really shows we're entering into a Sci-fi era.
blueblisters
Very impressive. But it doesn’t solve the whole problem yet.
The robot and ball pose is estimated by high speed mocap cameras, and is fed to the policy.
I imagine estimating that with onboard cameras - how humans do it - is much harder.
Almost all of closed loop robotics is a state estimation problem. Control is “solved” if you can estimate state well enough.
ohyoutravel
We know. Just appreciate it for what it is. Which is…awesome.
SiempreViernes
Look at the guys above posting that within 18 months these sorts of robots will be able to cook in anyone’s home; the above reminder is very necessary.
blueblisters
I agree it is pretty awesome!
V__
This is so interesting. Especially since it's kinda weird to train a robot to mimicking human play. I wonder what a perfect robot what actually behave like.
It wouldn't need to split-step to activate muscles, the footwork would probably be minimal. I imagine a lot of different unusual looking swings to confuse human players, while still making perfect contact. It could make really late drop shots or even rotate the racket at the last moment for crazy angles.
Would love to watch this.
pmontra
The humans in the video shot easy balls to the robot, which returns more difficult balls. It's the human that is doing all the running. The robot is quite static. However with better software and better hardware is possible that the robot will be so fast that it will miss no ball, and so strong to return balls faster than any human can reach. So there is no need to play fine shots. That could be a goal if we want to provide automated training partners to humans. If we want to win games against humans, stronger and faster is more than enough.
kadoban
> I wonder what a perfect robot what actually behave like.
Really depends what its hardware is. One with hardware a lot like a human would behave like a human.
Since you didn't specify, I'm going to go with a robot that looks like a giant pong paddle.
SiempreViernes
I don't know much about tennis, but the perfect opponent is probably some form of slightly concave wall that will always bounce the ball into the court no matter the angle you send the ball at it
hbcondo714
Impressive! Looks like a nice alternative or evolutionary step for a ball machine. Either way, teach it to serve :)
Aboutplants
Really impressive. In a few years there will be robotic AI instructors for the wealthy and their kids
squibonpig
Maybe for novelty, but the rich usually just pay humans to act like robots.
nradov
Ironically something like this could eventually make elite level tennis training cheaper and more accessible. Families of some top US juniors already spend $100k per year, much of that on 1:1 coaching. Some fraction could eventually be automated, at least for repetitive basic skills practice. Like the next level of a tennis ball machine.
dandaka
human instructors for wealthy, robotic AI for the rest of us
Void_
This just makes me want to play tennis right now. Such an addictive sports.
sam1r
So this is all pretty much theoretical, but very tightly woven strictly bounded protocols to be brought to production-- perhaps an accelerated alternative to perceive a much sooner ETA of 18months...
Maybe its moreso about reaching out to the right people about this "white paper" worthy research.
AFAIK, billions of dollars are poured into tennis mechanics at the highest level.
Introduce this to the right group of people, I truly can see this funded to play Janik Sinner where he would pay as a service to play against his worst nightmare.
ohyoutravel
Why can some Temu humanoid robot do this sort of impressive, coordinated, high-speed thing, but Tesla Optimus completely sucks at everything unless they’re moving at 0.02m/s (and even then they’re not great)? Like, train this thing on the latent space of folding my clothes out of the dryer and I will send you my money.
10xDev
Relax, it is one demo. It probably can't handle the millions of edge cases that exist in real life.
ohyoutravel
I’d be OK (and from a product perspective think it would be a win) if Optimus just mastered one high-value skill like clothes folding. Yet, here we are.
kace91
Now we intellectual workers can race physical workers to see who becomes obsolete first!
We have just started ramping up practical use of imitation learning from human demonstrations in humanoids. A bigger development is that one or two projects are working on training foundational vision action language models based on large video datasets.
I think before the end of summer general purpose physical knowledge and capabilities will start to be demonstrated by one or more humanoid AI or robotics groups.
Maybe 18 months at the absolute latest.
I'm guessing by next year or 2028 there will be services where you can order a robot to come cook and or clean for you. By 2029 it should be quite affordable to get a humanoid on a short term rental.
Do we have any standard benchmarks for humanoids to do domestic tasks?