A year ago this [0] table tennis robot backed by Google DeepMind was discussed on HN.
It plays much worse and the HN discussion is anchored around whether it's OK to call it "human-level" or if the authors should have clarified that they meant a human who doesn't actually play table tennis. But it was accepted as being SOTA at that time.
What happened since then? This looks like the kind of level of advance we see in, say, coding AIs, but I thought physical robotics was advancing much more slowly.
A partial answer is that the new robot cheats in ways that DeepMind didn't seem to. It has high speed cameras all over the room and can detect spin by observing the logo on the ball. But I'm not sure this explains such a big advance.
As a human player (of a not-high standard) I cannot see the spin of the ball directly. I can only infer it from the movement of my opponents bat. So I would wonder that a camera could pick it up in real time.
As a player myself, and having seen much higher level player than me, reading the spin from the spin (and in fact trajectory) of the ball is a common (if advanced) skill. Sometimes the movement of the bat can be deceptive (since with the same movement, where it contact on the bat, the finger pressure can affect the spin).
For example, backspin/underspin balls will move slower after the first bounce and feel 'damper' while topspin will jump. So it's def. possible (and in fact reliable) to read the spin from the spin and trajectory of the ball.
I had a look at Google trends for France. Table tennis is slightly more common than ping pong but the latter is much more stable. Table tennis has huge peaks, the biggest one being during the OG in Paris. These parks are not reflected in there ping pong trend
Interestingly, for Youtube searches this is the other way, with a much bigger difference in favour to ping pong
My biggest fear at the moment is robot armies and police forces.
Case in point : we're all expecting China needs to invade Taiwan soon, or they will run out of soldiers because of the one child policies of the 70s/80s.
Meanwhile, Ukraine is holding up against a "modern" army with quickly assembled drones.
So it all seems a bit like "they'll never put tanks through the Ardennes", sort of ?
Where and when will the first invasion of a country by a purely remote controlled, AI assisted army take place ?
Will robot battalions embed civilians to act as human shields ? Will the AI learn to mistreat the locals to maintain fear, or will they see it as a needless distraction and rush to the center of powers ?
If war is mostly played out from a disrance, will years of playing RTS give South Korea an edge ?
Autonomous suicide drone swarms are easily countered by autonomous interceptor swarms.
>Marching humanoid terminator robots
ground bots, not necessarily marching, do have their value. They can have bulletproof armor, while still be relatively lightweight and small and fast. They can easily carry even 20-25mm autocannon - very destructive weapon, sometimes can even succeed against a real tank.
And imagine when a swarm of drones lifts a ground bot, brings and drops it right into the needed point and protects it from the enemy drones while the ground bot just destructs the things around.
I don't think Russian army is very modern -- but maybe that's the reason of your quotation marks.
I kinda think that the competitions among the big dogs (US/Russia/China/etc.) would eventually green light ANY AI/Robots projects if they can justify tipping the scale somehow, and in the process completely destroys the last element of any political counter-weight. Because "fear gives men wings".
I would really hate to live in a dystopian world worse than what is described in the books/movies.
Peter Ilyich Tchaikovsky was born in Votkinsk, May 7 1840.
When he was a little boy he never played out in the streets of Votkinsk like the other little children of Votkinsk, because when Tchaikovsky was one month old, his parents moved to St. Petersburg.
Much like the robots beating half marathon records in China recently… who cares? Cake making robots can make cakes way faster than human bakers. Cars and motorcycles go faster than bicyclists. It is a boring given that purpose made machines perform the tasks they are built to perform better than humans.
Yeah, thinking through it a bit further, the real story here, aside from the mechanical engineering, is the application of AI/machine learning/computer vision processing. The advancements that have made it possible to reason about, simulate, and react to the complexities of a spinning ball in a fraction of a second are pretty cool. My gripe is mostly that this article isn't focusing on and detailing this.
The article's main focus is on the "vs. human" aspect and is light on technical details. I would love to hear specifics from the engineers behind this.
I would love to see a video of this thing that shows the whole table. From the paper I guess they have to light the area very brightly. But it seems like a pretty serious set up.
Don't table tennis players learn to predict how the ball will act based on their opponents movements? Seems like if they aren't able to do that with a robot opponent (who doesn't look or behave like a human) then they wouldn't be able to play at their best.
I do expect this to have a "novelty edge" over human opponents - which can be closed with practice, on the human end.
And, like many AIs, it can have "jagged capability" gaps, with inhuman failure modes living in them - which humans can learn to exploit, but the robot wouldn't adapt to their exploitation because it doesn't learn continuously. Happened with various types of ML AIs designed to fight humans.
Chess players learned to exploit chess computers’ weaknesses in the beginning too, but they can’t any longer. This version of the robot might not learn continuously, but the next will be better.
You can predict the movement of the ball (speed, direction, spin) based on the movement of the bat relative to the ball. What the rest of the player's body is doing is irrelevant to predicting what the ball will do - but relevant to predicting where they will be when you make the return shot.
Rui Takenaka, an elite-level player who has won and lost matches against Ace, said in comments provided by Sony AI: "When it came to my serve, if I used a serve with complex spin, Ace also returned the ball with complex spin, which made it difficult for me. But when I used a simple serve - what we call a knuckle serve - Ace returned a simpler ball. That made it easier for me to attack on the third shot, and I think that was the key reason why I was able to win."
It seems like the human players might be playing in a way that tacitly overestimates their AI opponents' intelligence and underestimates their skill. AFAIK the SOTA Go AIs are still vulnerable to certain very stupid adversarial strategies that wouldn't fool an amateur (albeit they're not something you'd come up with in normal play, more like a weird cheat code). I wonder if this could get ironed out with a bit more training against humans vs. simulation.
Am I correct in my understanding that- they had specialized software that not only tracked the ball, calculated spins using the logo, and fed calculated trajectories?
Even club level players have access to tennis table 'robots'. They fire the ball at you and collect the return in a net. You can set the speed, position and spin. They are very basic compared to this robot, but useful for training.
I've always wished for something similar: autonomous car racing. No human drivers. No remote controls. Just program the cars before the race, and let them go. Maybe even load the cars with mild explosives so they go BOOM when they crash.
A year ago this [0] table tennis robot backed by Google DeepMind was discussed on HN.
It plays much worse and the HN discussion is anchored around whether it's OK to call it "human-level" or if the authors should have clarified that they meant a human who doesn't actually play table tennis. But it was accepted as being SOTA at that time.
What happened since then? This looks like the kind of level of advance we see in, say, coding AIs, but I thought physical robotics was advancing much more slowly.
A partial answer is that the new robot cheats in ways that DeepMind didn't seem to. It has high speed cameras all over the room and can detect spin by observing the logo on the ball. But I'm not sure this explains such a big advance.
[0] https://news.ycombinator.com/item?id=43861207
As a human player (of a not-high standard) I cannot see the spin of the ball directly. I can only infer it from the movement of my opponents bat. So I would wonder that a camera could pick it up in real time.
Also IT'S TABLE TENNIS, NOT PING PONG!
As a player myself, and having seen much higher level player than me, reading the spin from the spin (and in fact trajectory) of the ball is a common (if advanced) skill. Sometimes the movement of the bat can be deceptive (since with the same movement, where it contact on the bat, the finger pressure can affect the spin).
For example, backspin/underspin balls will move slower after the first bounce and feel 'damper' while topspin will jump. So it's def. possible (and in fact reliable) to read the spin from the spin and trajectory of the ball.
According to this video it can read the spin:
https://www.youtube.com/watch?v=EH8kZDc7OLk
I had a look at Google trends for France. Table tennis is slightly more common than ping pong but the latter is much more stable. Table tennis has huge peaks, the biggest one being during the OG in Paris. These parks are not reflected in there ping pong trend
Interestingly, for Youtube searches this is the other way, with a much bigger difference in favour to ping pong
My biggest fear at the moment is robot armies and police forces.
Case in point : we're all expecting China needs to invade Taiwan soon, or they will run out of soldiers because of the one child policies of the 70s/80s.
Meanwhile, Ukraine is holding up against a "modern" army with quickly assembled drones.
So it all seems a bit like "they'll never put tanks through the Ardennes", sort of ?
Where and when will the first invasion of a country by a purely remote controlled, AI assisted army take place ?
Will robot battalions embed civilians to act as human shields ? Will the AI learn to mistreat the locals to maintain fear, or will they see it as a needless distraction and rush to the center of powers ?
If war is mostly played out from a disrance, will years of playing RTS give South Korea an edge ?
Marching humanoid terminator robots will never be as cheap as a drone. Autonomous suicide drone swarms are what should terrify you.
Autonomous suicide drone swarms are easily countered by autonomous interceptor swarms.
>Marching humanoid terminator robots
ground bots, not necessarily marching, do have their value. They can have bulletproof armor, while still be relatively lightweight and small and fast. They can easily carry even 20-25mm autocannon - very destructive weapon, sometimes can even succeed against a real tank.
And imagine when a swarm of drones lifts a ground bot, brings and drops it right into the needed point and protects it from the enemy drones while the ground bot just destructs the things around.
China had more births in 2025 than all of europe and russia combined so I don't think they're going to run out of soldiers.
I don't think Russian army is very modern -- but maybe that's the reason of your quotation marks.
I kinda think that the competitions among the big dogs (US/Russia/China/etc.) would eventually green light ANY AI/Robots projects if they can justify tipping the scale somehow, and in the process completely destroys the last element of any political counter-weight. Because "fear gives men wings".
I would really hate to live in a dystopian world worse than what is described in the books/movies.
Reminds me of the Mitch Hedberg joke: "The depressing thing about tennis is that no matter how good I get, I'll never be as good as a wall."
I used to love Mitch Hedberg. I still do, but I used to, too.
Peter Ilyich Tchaikovsky was born in Votkinsk, May 7 1840.
When he was a little boy he never played out in the streets of Votkinsk like the other little children of Votkinsk, because when Tchaikovsky was one month old, his parents moved to St. Petersburg.
— Victor Borge
Put up in a place
where it is easy to see
the cryptic admonishment
T.T.T
When you feel how depressinglyslowly you climb
it's well to remember that
Things Take Time
-- Piet Hein
If you don't like a parade, run in the opposite direction to fast-forward it.
Much like the robots beating half marathon records in China recently… who cares? Cake making robots can make cakes way faster than human bakers. Cars and motorcycles go faster than bicyclists. It is a boring given that purpose made machines perform the tasks they are built to perform better than humans.
It's an amazing feat of engineering because it requires constant micro-adjustments, something that robots couldn't do a few years ago.
Yeah, thinking through it a bit further, the real story here, aside from the mechanical engineering, is the application of AI/machine learning/computer vision processing. The advancements that have made it possible to reason about, simulate, and react to the complexities of a spinning ball in a fraction of a second are pretty cool. My gripe is mostly that this article isn't focusing on and detailing this.
isn't this a technology forum?
The article's main focus is on the "vs. human" aspect and is light on technical details. I would love to hear specifics from the engineers behind this.
Here is the paper:
https://www.nature.com/articles/s41586-026-10338-5
I would love to see a video of this thing that shows the whole table. From the paper I guess they have to light the area very brightly. But it seems like a pretty serious set up.
quite surprised to see SAC, considering the deepmind ping pong paper resorted to evolutionary strategies, iirc
The motion system constrains the problem quite a bit. This video of high speed vision/actuators is 16 years old - https://www.youtube.com/watch?v=bfdHY26E2jc
I was expecting/hoping for a humanoid robot.
Nature video https://www.youtube.com/watch?v=EH8kZDc7OLk
Makes sense that it would.
Reminds me of this old The Onion story: https://theonion.com/ping-pong-somehow-elicits-macho-posturi...
What happens when two of them play each other?
How easy is it to introduce artifacts that reduce accuracy and performance?
Don't table tennis players learn to predict how the ball will act based on their opponents movements? Seems like if they aren't able to do that with a robot opponent (who doesn't look or behave like a human) then they wouldn't be able to play at their best.
I do expect this to have a "novelty edge" over human opponents - which can be closed with practice, on the human end.
And, like many AIs, it can have "jagged capability" gaps, with inhuman failure modes living in them - which humans can learn to exploit, but the robot wouldn't adapt to their exploitation because it doesn't learn continuously. Happened with various types of ML AIs designed to fight humans.
Chess players learned to exploit chess computers’ weaknesses in the beginning too, but they can’t any longer. This version of the robot might not learn continuously, but the next will be better.
You can predict the movement of the ball (speed, direction, spin) based on the movement of the bat relative to the ball. What the rest of the player's body is doing is irrelevant to predicting what the ball will do - but relevant to predicting where they will be when you make the return shot.
Yes, you're dead on:
It seems like the human players might be playing in a way that tacitly overestimates their AI opponents' intelligence and underestimates their skill. AFAIK the SOTA Go AIs are still vulnerable to certain very stupid adversarial strategies that wouldn't fool an amateur (albeit they're not something you'd come up with in normal play, more like a weird cheat code). I wonder if this could get ironed out with a bit more training against humans vs. simulation.Is there a video of this in action? Pictures are not satisfying at all!
https://www.youtube.com/watch?v=EH8kZDc7OLk
https://sonyresearch.github.io/ace_public/
Am I correct in my understanding that- they had specialized software that not only tracked the ball, calculated spins using the logo, and fed calculated trajectories?
I'll be impressed when it's a humanoid robot that has to contend with similar kinematic limitations as a human player.
I wonder if a top player with access to a robot like this can get an extra edge in training?
Even club level players have access to tennis table 'robots'. They fire the ball at you and collect the return in a net. You can set the speed, position and spin. They are very basic compared to this robot, but useful for training.
Cool. Now let's see two robots play and if it's fun let it become it's own thing. Other than that, this could be used for training actual players.
I've always wished for something similar: autonomous car racing. No human drivers. No remote controls. Just program the cars before the race, and let them go. Maybe even load the cars with mild explosives so they go BOOM when they crash.
robot ping pong league
The greatest blernsball player was a machine for playing blernsball.
> Now, Wireless Joe Jackson! There was a blern hitting machine.
> Exactly! He was a machine designed to hit blerns. I mean come on, Wireless Joe was nothing but programmable bat on wheels.
> Oh? And I suppose Pitch-o-mat 5000 was just a modifier howitzer?
> Yep!
AI gets all the fun jobs. Yet again.
Now build a robot that can catch a bullet.
careful what you wish for.