I've worked on much less expensive, much smaller humanoid robots.

These robots are certainly running through a scripted set of poses which has been extensively tested for the conditions (Humans would also be choreographed and have to hit certain marks at certain times). If you covered the stage in loose gravel or a thick carpet they'd all start falling over. The things the robots hold are almost certainly taped into their hands.

Despite that, this is a very impressive demo. Those robots are $40k+, they've got 20+ of them. And not a single one fell over. They're fast too - and there are a load of corners they could have cut, but they didn't.

The floor has two textures, it would have been easier without that. The humans right alongside them? Much less safety paperwork without them. The robot wearing trousers and a cape? Much easier without that. The fewer robots you have, the lower the chances on falls over landing their backflip. Lose the audience and record it in multiple takes. Hell, you could have human acrobats in robot costumes and it'd cost far less and be much easier.

So this demo is very much a costly signal of confidence.

> They're fast too

That was of of the two things that impressed me most, along with the choreography involving close and direct contact

> The things the robots hold are almost certainly taped into their hands.

You can clearly see that the robots change their grip of their sword, so it cannot be taped to their hands.

When is that?

With the poles at the 1 minute mark, the robots enter holding them and their left hand never moves on the pole. Also note the stationary hand is matte grey while the moving hand is metallic silver.

Likewise with the wine gourds (?) at 2m30s and the nunchucks at 3m40s.

It’s a completely sensible design decision, much simpler to do cartwheels and vaults if you don’t have super delicate fingers fitted.

Probably a compliant magnetic coupling, as even simple force-sensing mechanical hands cost too much to practice back-flips. =3

Smaller platforms are actually harder to build: minimal power budget, weaker drive systems, less sensors, and fewer processing options.

Not a fan of bipedal platforms or 50kg of servos for a number of reasons.

Best regards. =3

Yes and no.

An Aldebaran Nao can fall over with no damage because it’s only 5kg and 58cm. And you can use relatively low power motors, so nobody can lose a finger to crushing in the joints.

But you miss out on the benefits of being able to operate in a human centric world - you’ll never get a Nao to climb stairs, open a door, or carry a cup of coffee.

>Nao can fall over with no damage

They were overpriced, hard to repair, and the university lab would lose one every few weeks. The consumer grade internal ribbon PCB interconnect harness would usually tear eventually.

Those 3kW to 5kW brush-less servos in the biped dance video are often only 12:1 planetary reduction, and are almost certainly backdrivable. As long as the platform has a safe-fall canned pose, than damage should be limited to shear pins in the worst case falls.

Fun to see people pushing the design limits. =3

Why do you think it would be the case about e.g. swapping to thick carpet would throw things off? Intuitively it seems like they must have a tremendous amount of dynamic adjustment going on. For instance think of how much variance, driven by dynamics, that there's going to be in the scene at 2:48 [1] where the robot [intentionally] falls over and then aerobically picks itself back up.

The motion is certainly scripted, but the exact mechanics in play there almost certainly vary radically from take to take. Imagine something simple like a pool/billiards break. Even if you set up a machine to rack the balls and break them in as close to identical as possible, you'd get wildly different results each time. And the dynamics in this motion is going to dwarf that.

[1] - https://youtu.be/mUmlv814aJo?t=168

To be more precise: I think they have a fixed repertoire of moves that they can blend together and slightly tweak on the fly, but only within certain limits.

This would also be normal for human performers - touring ballet companies travel with their own flooring the dancers are used to pirouetting on.

At the 40 second mark every robot does a backflip then when landing hops their supporting leg while pointing the toe of their working leg. Which works fine and looks great! But they arrive in that pose with a certain amount of momentum and needing a certain amount of grip on the floor.

So this is a rehearsed, tested performance - not proof we’ll have firefighter robots doing parkour through burning buildings any time soon!

> not proof we’ll have firefighter robots doing parkour through burning buildings any time soon!

I wonder how similar friction is across varieties of standard issue corporate carpet...

> To be more precise: I think they have a fixed repertoire of moves that they can blend together and slightly tweak on the fly, but only within certain limits.

Human acrobats do the same, they know a few fixed tricks and practice them often in order to stay in shape. The point here is to compare the dynamic abilities of bots vs those of humans, not vs some pink unicorns - the latter comparison isn't informative because there aren't any jobs for pink unicorns just yet.

> So this is a rehearsed, tested performance - not proof we’ll have firefighter robots doing parkour through burning buildings any time soon!

A bot assisted by humans can do better than them in a burning building, the lack of need for heavy SCBAs would allow the bot to perform acrobatics scripted by a human at the time of action.

It’s not a 1:1 human motion capture to servo translation. There is some work done to fix Center of gravity like you said and issues with friction and momentum.

The hard part with “autonomy” is interpretation of the environment and feeding that back into some control loop to accomplish a goal in real time. That is why most of these demos are basically recordings of movements, like choreography.

They're also interacting with the environment (vaulting boxes / walls), which implies they either know their 3D position very accurately, or they have some form of sensors and can adapt a bit.

That’s fine, static objects can be integrated in to sequences if each unit has enough positional accuracy.

My point is that if you interrupt the sequence by pushing it over or standing in front of it, the unit will not be able to recover or mitigate the issue. I’m not sure about the models in the video but they should enter a safe state, hopefully gracefully.

It looks like they have a point https://x.com/Osint613/status/2024068210660643009

As someone who owns a pair of Unitree G2s this blew my mind

what are you using those for ?

Performance and entertainment at events, along with four of their Go2 Pro dogs (the real crowd pleasers)

> balancing their own motion a bit individually.

check this 4 months old video below

https://www.youtube.com/watch?v=bPSLMX_V38E

I'd willing to bet that it is already close to impossible to get the robot lose its balance without some significant external forces.

Of course the robots have been pre-trained and the movements are scripted, and nobody is claiming otherwise. But there must be a lot of autonomous balancing taking place. At one point you can see the robots adjusting their feet slightly different although they are all in sync, and that catapult does not look like its movement is exactly the same every time. It is just super impressive.

Does anyone remember when Honda's Asimo robot clumsily fell down the stairs during a demonstration[1] and we thought we were safe from a robot invasion by just moving to the upper floor? That was about 20 years ago.

[1] https://www.youtube.com/watch?v=6mx6paHrnIE

The impressive part here isn't the movement itself. You can easily train a model to perform a "procedural animation" that includes a full body control policy. The hard part is making it reliable enough to perform long sequences of movements and adapting to differences in robot placement. In other words, performing a flawless stage play is the hardest part.

I'm afraid you might not understand what you're talking about. Animation is a geometry problem, while robotics is a dynamics problem. The latter is subject to constraints many times greater than the former. There is no such "easy" model as you imagined that can transform the former into the latter.

Those are the same people that say that China is 30 years behind in chip manufacturing.

I don't know if you're aware but robots and chips are different things that require different expertise.

It's still a completely farcical attitude. At this point it's just a matter of how many years it'll take for chinese manufacturing to outstrip western expertise with chips, too. Ten years? Five? Two?

Very exciting times to live through.

Yes, the autonomy level of these robots was what I was yesterday emailing with my former colleagues we were wondering. Two months ago CNET & PC-Mag posted following video which suggests more about robots movements being assisted by humans. And it also shows Chinese have being edge of the development at that point.

https://www.youtube.com/watch?v=jXTibM33SDg

However, then another short video bit alike popped up and is puzzling too.

Apparently Unitree robot is playing pingpong match like a pro. Sorry about german announcer, I couldn't find with english.

https://www.youtube.com/shorts/BgD1ukTyNnw

There is another match viewable by pressing that "Robot plays ping ppng #robot" arrow.

How about that robot? Is it human assisted or not? Our opinions diverted, I'm quite sure it is assisted but my former colleague thinks it's got to be autonomous as it would be too difficult and slow to do that fast movements with remote control assisted robot.

It would be nice to hear opinions about that playing robot too if anyone could provide some insight in that.

edit: I think the serve waiting robot hand movement and after losing wiping left eye gesture as a disappointing a bit in my opinion gives up it's human. Or if not, why would a robot do such a human like gestures.

edit2: OK, good points, I see now. It's definitely a fake. Thanks to all who replied :)

here is a real video of a unitree robot playing ping pong https://www.youtube.com/watch?v=tOfPKW6D3gE

that ping pong video is a CG robot, whether realtime superimposed or otherwise who knows. Look at the :27 when it gets out of tracking breaking all of physics, feet aren't planted to the ground, light, shadows.. etc.

I can think that future use of pingpot robot is to replicate specific pro player style (from various recording) and be used to spar by pro players before their specific matches.

I think the ping pong match video might be misleading you. Based on the visual artifacts around the robot, the original footage likely had a human player that was swapped in with a robot in this video. It also has an altered content warning.

The ping pong video you linked is clearly fake. Look at the paddle... anyways...

The pingpong video is very obviously computer generated. The robot feet give it away immediately

The second video you've linked is fake in every aspect in regards to the robot.

The robot is floating above the ground.

The paddle is phasing in and out of existence.

The robot has a realistic human hand and uses it to hit the ball.

The robot randomly turns around mid-air near the end of the video.

The robot looks nothing like a Unitree robot.

Oh, how could I forget, the entire robot looks so obviously fake even when disregarding all of the above that I can't believe you're even trying to analyze anything in that video.

I'm 99% sure that ping pong match is CGI. The whole robot has this green screen effect. Look at its feet. And at second 17 it just disappears entirely for a few frames.

Watch out, the two shorts you linked (both of robots playing ping-pong) are fake.

They do use keyframes most likely captured from a human controller. you can see this after they do the backflip at :29s they land a bit differently and recover in slightly different ways but all end up in a static pose for a moment before moving on to next movement. The advancement here is the dynamics to go between those frames. Looking at last years performance you can see they pretty much go from frame a to b then stabilize then to c then stabilize. This is what makes this years look much more lifelike there seems to be some active stabilization going on during the movements. It also seems to let them chain movements that can take advantage of momentum much better rather than needing to be at rest between frames.