Very cool, but "Our approach lead to competitive play at human level and a robot agent that humans actually enjoy playing with": I would not call that remotely near "human level" yet. I think we'll get there at some point but I think people underestimate how difficult table tennis is and the level of top human play is insane. This was the World Cup match point recently for reference: https://www.youtube.com/shorts/-AJg2u7U5MU
As a little side project I'm working on a table tennis AI for VR which works by imitating real players, which is a much simpler problem since you're allowed to "fake" a lot of things in a game. I think VR holds more promise in the short to medium term for practicing TT than robotics.
It's exactly at the level I thought was communicated by the title and introduction.
"Human level competitive", "solidly amateur-level human performance", "beat 100% of beginners and 55% of intermediate players". That robot would definitely win some games in your local club league, except that it doesn't serve, and unless it's cheating in ways the announcement glosses over like extra cameras - DeepMind have some history here so I reserve the right to be skeptical.
The only thing I'd take issue with in the abstract is "Table tennis... requires human players to undergo years of training to achieve an advanced level of proficiency." While that sentence is true, it's irrelevant to this robot since this robot only plays at intermediate proficiency, a level reachable by a moderately athletic human with some practice.
By contrast, the AlphaGo [0] AlphaZero [1] and AlphaStar [2] papers claim "mastery", "superhuman", "world champion level", "Grandmaster-level", "human professional" ability - all defensible claims given their performance and match conditions in the respective games.
> That robot would definitely win some games in your local club league
Definitely not. If you go beyond the cherry-picked videos where some longer sequences happened, the longer match videos reveal how bad the robot is. It makes really bad mistakes and loses most points against players not even on intermediate player in any local club.
Yeah, looking it play makes me thinks it has a level comparable to mine, that is the level of someone that enjoys playing table tennis with friends and family members a every other year. Not at all the level you'd see in clubs, even among low-ranked people.
I'd describe the robot's level as "good for outdoor table tennis".
There are pretty much too distinct classes of players. Those that occasionally play for fun, typically at stone tables found in parks and open-air baths.
And then there are those that play and train at least once a week in indoor halls with wooden tables, and often try to learn proper stroke techniques, and often participate in leagues.
The robot is pretty good fit for the first category, and that's already a pretty impressive achievement.
In the second category, it'd lose to anybody playing for more than a year or two, so it would be on par with the lowest tier players there.
My local club is full of (mostly older) players who play once a week for a few hours. They don't train outside of this, but they play with proper equipment, sometimes play leagues, sometimes play a few strokes well, and beat anyone who just plays casually, apart from the occasional excellent tennis player who can transfer just enough skills to be competitive. But they also make lots of unforced errors and don't have good technique on all strokes.
The robot looks like it would be competitive with most of those players. Maybe my club is uniquely weak.
The AlphaZero paper claimed it was vastly superior to Stockfish, the top open source engine, based on a 100-game match against Stockfish. It turned out Stockfish was running with way less compute. It's not an apples to apples comparison between CPU and GPU, but IIRC there were orders of magnitude difference in the hardware cost and power budget. Additionally, they used a build of Stockfish that was tuned based on having access to opening books and endgame tablebases, but didn't give it those resources in the match.
The original AlphaStar announcement was also based on having serious advantages over its human opponents: it got a feed of the whole map, where humans could only view a section at a time, and the ability to perform an unrealistic number of actions per minute.
The equivalent in table tennis? Maybe having an additional high speed camera on the other side of the table, or a sensor in the opponent's bat. Actually, why is the opponent playing with a non-standard bat with two black rubbers? Presumably that's an optimization where the robot's computer vision has been tuned only for a black bat. But if that's so, it means none of the opponents got to use their own equipment, they used a bat which was unfamiliar and perhaps chosen to be easy for the robot to play against.
I could concede it as "lower intermediate", but the span is huge. I don't think it can beat an upper beginner consistently. It all depends on what ELO you consider intermediate.
Yeah, but I don't think anyone who's played TT regularly would call this "intermediate". It's like calling someone who managed to code fizzbuzz an intermediate programmer.
To be honest this feels a bit pedantic to me. "Intermediate" is a pretty vague word and unless there's some official rankings termed "intermediate" I'm unaware of, the authors of this paper are fine to apply it as they see fit.
There are a significant number of players worse than the robot, and a significant number of players better. That's fact, and that's all "intermediate" means.
Left-handed, one one leg, one eye closed and a mobile phone instead of a racket. Reminds me of a clip of Truls Moregardh playing a reporter with a phone: https://www.youtube.com/watch?v=3EJaT1b8QmA
As somewhat of a table tennis nerd, I really like this, but also cannot help to throw in some critical observations :-)
I've skimmed through the match highlights video, and it struck me that there wasn't really any engagement with spin.
You can hear from the relatively high pitch of the ball contact that these were rubbers with very thin sponges or no sponge at all, and likely not enough grip to cause any real rotation on the ball.
Even in the lowest leagues (here in Germany, at least), people use backspin to prevent attacks, and varying the spin is half the game.
Also, human players often have different rubbers on the forehand and backhand sides, which is why rules demand that both rubbers have very distinct colors. Using black rubbers on both sides kinda demonstrates their non-engagement with spin.
That said, kudos for making a robot that people enjoy playing with. That's not a given. Players anticipate the trajectory of the ball based on the opponent's body movement, which a robot could totally subvert.
Physical limitations also play a real role in table tennis, for example the forehand flick is much harder to play than the backhand flick due to the way our hand joint works. A robot wouldn't be subjected this particular limitation.
> that reaches amateur human-level performance in competitive table tennis
How are we defining amateur here? The presented video shows the human intentionally volleying with the robot, barely putting any force at all behind the returns. But it says the robot won 55% of matches against intermediate players? That requires being able to return much harder shots than shown.
They note that the robot achieved an "intermediate" skill level. This has been determined by letting it play matches against people of different skill levels as determined by a professional table tennis coach. The "Results" section explains this.
Google had a very competitive employee ping pong league and one of the coaches was on the USA Olympic team, I doubt they would lie about ping pong skills
Googlers are well known for speaking up when they disagree with something the company is doing, but yeah I mostly wanted to call out how big the ping pong culture is at Google at least in Mountain View
But they don't explain what metrics they use to differentiate skill level other than waving the hands of a puppet coach. The arm control is great but they don't show it returning non-softballs. Is that what "intermediate" means?
Probably, if you consider that a beginner can rarely volley at all. At an intermediate level, a fast return pretty much just wins the point, when it doesn't hit the net or miss the table.
The interesting result here is not the performance against human players, which DeepMind tries to upsell. The interesting bit is their approach to bridging the sim-2-real gap by iteratively training in a simulation and fine-tuning in real-world games. The approach is described here:
i-Sim2Real: Reinforcement Learning of Robotic Policies in Tight Human-Robot Interaction Loops
The sim-2-real gap is a real obstacle in adopting RL for rotobics and anything that pushes the envelope is worth the trouble. On the other hand, I can't tell how well this approach will work outside of Table Tennis.
Note that on top of the RL work there is, AFAICT a metric shit ton of good, old-fashioned engineering to support the decision-making ability of learned policies. E.g. the "High Level Controller" (HLC) that selects "Low Level Controllers" (LLC) using good, old-fashioned AI tools, tree search with heuristics, seems to me to be hand-crafted rather than learned, and involves a load of expert knowledge driving information gathering. So far, no bitter lessons to taste here.
Oh and of course the HLC is a symbolic component while the LLCs are learned by a neural net. Once more DeepMind sneakily introduces a neuro-symbolic approach but keeps quiet about it. They've done that since the days of AlphaGo. No idea why they're so ashamed of it, since it really looks like it's working very well for them.
It counts a lot for me. There's enough examples of humans playing table tennis; I'm very intrigued about how the same problem can be solved by alternate means.
I can only find papers from the late 1980s on this, but a professor around during the 60s told me hackers were working on robotic ping-pongers way back then.
I don't find this particularly impressive for DeepMind, mostly because it seems like a ballistics/physics problem rather than a machine learning problem.
If DeepMind wanted to emphasize the learning aspect of it (and they should) then it should be in the title. E.g. "Novel learning algorithm leads to competitive robotic table tennis"
The paper addresses this. Part of the goal was to let the world model be learned rather than modeled explicitly by the programmer.
> To date, the Omron Forpheus robot [73], [62] has the closest capabilities to the agent presented in this work, demonstrating sustained rallies on a variety of stroke styles with skilled human players. A key point of difference is that our agent learns the control policies and perception system, whereas the Forpheus agent uses a model-based approach. More specifically, Forpheus leverages rebound and aerodynamics models in order to identify the optimal configuration of the robot so as to return the ball to a target position. The Omron system represents a highly engineered system that cannot easily be customized to new players, environments, or paddles.
> While there have been many demonstrations of robots playing table tennis against human players in the past, we believe this research is one of the first human-robot interaction studies to be conducted with full competitive matches against such a wide range of player skill levels.
Feels like clickbait. People were doing similar things in the 60s at MIT. I thought by "human level robot" they meant actually human level, with a humanoid robot swinging the paddle, not a conveyor belt.
I'm pretty sure Omron has been bringing a table tennis robot to trade shows for at least ten years, and it's pretty good at the game. They mention it in their paper:
> A key point of difference is that our agent learns the control policies and perception system, whereas the Forpheus agent uses a model-based approach. More specifically, Forpheus leverages rebound and aerodynamics models in order to identify the optimal
configuration of the robot so as to return the ball to a target position. The Omron system represents a highly engineered system that cannot easily be customized to new players, environments, or paddles
But I think they're stretching a bit to claim that a model-based design can't be easily customized. Many of us would consider it much easier to plug in a new air viscosity or coefficient of restitution value into a model than to re-train a physical robot.
Based on what I'm seeing in the video, the robot isn't slicing the paddle in a way that would impart much spin other than what's already there. This is essentially a defensive play style.
That's not to say that the ball doesn't have spin. I play with a lot of spin, and a recent level up in my game was that I realized that when I put spin on the ball and my opponent returns it without slicing, I've got to account for the spin I put on it because it's still there.
You can get pretty dang far with a primarily defensive style, though. I was recently at an informal local tournament where the 2nd place guy had basically one serve and an insane ability to return shots. A lot of the times he scored in the final game were where his opponent hit a shot with such insane speed and spin that when when he returned it, his opponent couldn't defend the residual speed/spin of his own shot.
It is. It's just that we now live in an age where everything has to be solved with "AI" to get attention and justify the billions of investments. So we forget established tech and enter the era of AI enshittification where this kind of thing is considered an amazing achievement.
Had this been some high school kids I'd be impressed. But Google? C'mon.
it's a little disingenuous to claim human level performance at a sport when you use a robot arm on a track. Humans have to contend with a body not specifically suited to the problem.
I think on a population / humanity level, there is no such thing as "good enough". There is always something new to experiment with, something to improve, something to change. This bother seems to propel humans since time immemorial.
Very cool, but "Our approach lead to competitive play at human level and a robot agent that humans actually enjoy playing with": I would not call that remotely near "human level" yet. I think we'll get there at some point but I think people underestimate how difficult table tennis is and the level of top human play is insane. This was the World Cup match point recently for reference: https://www.youtube.com/shorts/-AJg2u7U5MU
As a little side project I'm working on a table tennis AI for VR which works by imitating real players, which is a much simpler problem since you're allowed to "fake" a lot of things in a game. I think VR holds more promise in the short to medium term for practicing TT than robotics.
It's exactly at the level I thought was communicated by the title and introduction.
"Human level competitive", "solidly amateur-level human performance", "beat 100% of beginners and 55% of intermediate players". That robot would definitely win some games in your local club league, except that it doesn't serve, and unless it's cheating in ways the announcement glosses over like extra cameras - DeepMind have some history here so I reserve the right to be skeptical.
The only thing I'd take issue with in the abstract is "Table tennis... requires human players to undergo years of training to achieve an advanced level of proficiency." While that sentence is true, it's irrelevant to this robot since this robot only plays at intermediate proficiency, a level reachable by a moderately athletic human with some practice.
By contrast, the AlphaGo [0] AlphaZero [1] and AlphaStar [2] papers claim "mastery", "superhuman", "world champion level", "Grandmaster-level", "human professional" ability - all defensible claims given their performance and match conditions in the respective games.
[0] https://www.researchgate.net/publication/292074166_Mastering...
[1] https://arxiv.org/pdf/1712.01815
[2] https://www.nature.com/articles/s41586-019-1724-z
> That robot would definitely win some games in your local club league
Definitely not. If you go beyond the cherry-picked videos where some longer sequences happened, the longer match videos reveal how bad the robot is. It makes really bad mistakes and loses most points against players not even on intermediate player in any local club.
Yeah, looking it play makes me thinks it has a level comparable to mine, that is the level of someone that enjoys playing table tennis with friends and family members a every other year. Not at all the level you'd see in clubs, even among low-ranked people.
I'd describe the robot's level as "good for outdoor table tennis".
There are pretty much too distinct classes of players. Those that occasionally play for fun, typically at stone tables found in parks and open-air baths.
And then there are those that play and train at least once a week in indoor halls with wooden tables, and often try to learn proper stroke techniques, and often participate in leagues.
The robot is pretty good fit for the first category, and that's already a pretty impressive achievement.
In the second category, it'd lose to anybody playing for more than a year or two, so it would be on par with the lowest tier players there.
My local club is full of (mostly older) players who play once a week for a few hours. They don't train outside of this, but they play with proper equipment, sometimes play leagues, sometimes play a few strokes well, and beat anyone who just plays casually, apart from the occasional excellent tennis player who can transfer just enough skills to be competitive. But they also make lots of unforced errors and don't have good technique on all strokes.
The robot looks like it would be competitive with most of those players. Maybe my club is uniquely weak.
> DeepMind have some history here so I reserve the right to be skeptical.
What history of cheating is there? I hadn't seen anything sneaky, but I don't follow everything. Do share.
The AlphaZero paper claimed it was vastly superior to Stockfish, the top open source engine, based on a 100-game match against Stockfish. It turned out Stockfish was running with way less compute. It's not an apples to apples comparison between CPU and GPU, but IIRC there were orders of magnitude difference in the hardware cost and power budget. Additionally, they used a build of Stockfish that was tuned based on having access to opening books and endgame tablebases, but didn't give it those resources in the match.
The original AlphaStar announcement was also based on having serious advantages over its human opponents: it got a feed of the whole map, where humans could only view a section at a time, and the ability to perform an unrealistic number of actions per minute.
The equivalent in table tennis? Maybe having an additional high speed camera on the other side of the table, or a sensor in the opponent's bat. Actually, why is the opponent playing with a non-standard bat with two black rubbers? Presumably that's an optimization where the robot's computer vision has been tuned only for a black bat. But if that's so, it means none of the opponents got to use their own equipment, they used a bat which was unfamiliar and perhaps chosen to be easy for the robot to play against.
I could concede it as "lower intermediate", but the span is huge. I don't think it can beat an upper beginner consistently. It all depends on what ELO you consider intermediate.
Most table tennis players don't play at the level of the World Cup.
Yeah, but I don't think anyone who's played TT regularly would call this "intermediate". It's like calling someone who managed to code fizzbuzz an intermediate programmer.
To be honest this feels a bit pedantic to me. "Intermediate" is a pretty vague word and unless there's some official rankings termed "intermediate" I'm unaware of, the authors of this paper are fine to apply it as they see fit.
There are a significant number of players worse than the robot, and a significant number of players better. That's fact, and that's all "intermediate" means.
Exactly this.
A world-class player could beat this thing left-handed with one eye closed.
Left-handed, one one leg, one eye closed and a mobile phone instead of a racket. Reminds me of a clip of Truls Moregardh playing a reporter with a phone: https://www.youtube.com/watch?v=3EJaT1b8QmA
As somewhat of a table tennis nerd, I really like this, but also cannot help to throw in some critical observations :-)
I've skimmed through the match highlights video, and it struck me that there wasn't really any engagement with spin.
You can hear from the relatively high pitch of the ball contact that these were rubbers with very thin sponges or no sponge at all, and likely not enough grip to cause any real rotation on the ball.
Even in the lowest leagues (here in Germany, at least), people use backspin to prevent attacks, and varying the spin is half the game.
Also, human players often have different rubbers on the forehand and backhand sides, which is why rules demand that both rubbers have very distinct colors. Using black rubbers on both sides kinda demonstrates their non-engagement with spin.
That said, kudos for making a robot that people enjoy playing with. That's not a given. Players anticipate the trajectory of the ball based on the opponent's body movement, which a robot could totally subvert.
Physical limitations also play a real role in table tennis, for example the forehand flick is much harder to play than the backhand flick due to the way our hand joint works. A robot wouldn't be subjected this particular limitation.
> that reaches amateur human-level performance in competitive table tennis
How are we defining amateur here? The presented video shows the human intentionally volleying with the robot, barely putting any force at all behind the returns. But it says the robot won 55% of matches against intermediate players? That requires being able to return much harder shots than shown.
They note that the robot achieved an "intermediate" skill level. This has been determined by letting it play matches against people of different skill levels as determined by a professional table tennis coach. The "Results" section explains this.
Google had a very competitive employee ping pong league and one of the coaches was on the USA Olympic team, I doubt they would lie about ping pong skills
What do you think the mechanism is that would make them honest?
Googlers are well known for speaking up when they disagree with something the company is doing, but yeah I mostly wanted to call out how big the ping pong culture is at Google at least in Mountain View
Just watch the videos. They don't support your statements.
It looks very slow in the videos though.
But they don't explain what metrics they use to differentiate skill level other than waving the hands of a puppet coach. The arm control is great but they don't show it returning non-softballs. Is that what "intermediate" means?
Probably, if you consider that a beginner can rarely volley at all. At an intermediate level, a fast return pretty much just wins the point, when it doesn't hit the net or miss the table.
Submitted on 7 Aug 2024
https://arxiv.org/abs/2408.03906
> Policy performance was assessed through 29 robot vs. human matches of which the robot won 45% (13/29)
The interesting result here is not the performance against human players, which DeepMind tries to upsell. The interesting bit is their approach to bridging the sim-2-real gap by iteratively training in a simulation and fine-tuning in real-world games. The approach is described here:
i-Sim2Real: Reinforcement Learning of Robotic Policies in Tight Human-Robot Interaction Loops
https://sites.google.com/view/is2r
The sim-2-real gap is a real obstacle in adopting RL for rotobics and anything that pushes the envelope is worth the trouble. On the other hand, I can't tell how well this approach will work outside of Table Tennis.
Note that on top of the RL work there is, AFAICT a metric shit ton of good, old-fashioned engineering to support the decision-making ability of learned policies. E.g. the "High Level Controller" (HLC) that selects "Low Level Controllers" (LLC) using good, old-fashioned AI tools, tree search with heuristics, seems to me to be hand-crafted rather than learned, and involves a load of expert knowledge driving information gathering. So far, no bitter lessons to taste here.
Oh and of course the HLC is a symbolic component while the LLCs are learned by a neural net. Once more DeepMind sneakily introduces a neuro-symbolic approach but keeps quiet about it. They've done that since the days of AlphaGo. No idea why they're so ashamed of it, since it really looks like it's working very well for them.
Until the robot is standing on its own two legs like a human I feel like it doesn't count. Sliding along a rail isn't comparible
It counts a lot for me. There's enough examples of humans playing table tennis; I'm very intrigued about how the same problem can be solved by alternate means.
People did the conveyor belt thing in the 60s.
Citation?
I can only find papers from the late 1980s on this, but a professor around during the 60s told me hackers were working on robotic ping-pongers way back then.
I don't find this particularly impressive for DeepMind, mostly because it seems like a ballistics/physics problem rather than a machine learning problem.
Edit: yep here's a better performing robot from 6 years ago based on ballistics: https://m.youtube.com/watch?v=u3L8vGMDYD8&pp=0gcJCYQJAYcqIYz...
If DeepMind wanted to emphasize the learning aspect of it (and they should) then it should be in the title. E.g. "Novel learning algorithm leads to competitive robotic table tennis"
The paper addresses this. Part of the goal was to let the world model be learned rather than modeled explicitly by the programmer.
> To date, the Omron Forpheus robot [73], [62] has the closest capabilities to the agent presented in this work, demonstrating sustained rallies on a variety of stroke styles with skilled human players. A key point of difference is that our agent learns the control policies and perception system, whereas the Forpheus agent uses a model-based approach. More specifically, Forpheus leverages rebound and aerodynamics models in order to identify the optimal configuration of the robot so as to return the ball to a target position. The Omron system represents a highly engineered system that cannot easily be customized to new players, environments, or paddles.
> While there have been many demonstrations of robots playing table tennis against human players in the past, we believe this research is one of the first human-robot interaction studies to be conducted with full competitive matches against such a wide range of player skill levels.
Feels like clickbait. People were doing similar things in the 60s at MIT. I thought by "human level robot" they meant actually human level, with a humanoid robot swinging the paddle, not a conveyor belt.
I'm pretty sure Omron has been bringing a table tennis robot to trade shows for at least ten years, and it's pretty good at the game. They mention it in their paper:
> A key point of difference is that our agent learns the control policies and perception system, whereas the Forpheus agent uses a model-based approach. More specifically, Forpheus leverages rebound and aerodynamics models in order to identify the optimal configuration of the robot so as to return the ball to a target position. The Omron system represents a highly engineered system that cannot easily be customized to new players, environments, or paddles
But I think they're stretching a bit to claim that a model-based design can't be easily customized. Many of us would consider it much easier to plug in a new air viscosity or coefficient of restitution value into a model than to re-train a physical robot.
> than to re-train a physical robot
Where we are going the robot retrains itself.
Does the ball hit by the robot have spin, or is it just a normal return? I'm kind of curious.
Based on what I'm seeing in the video, the robot isn't slicing the paddle in a way that would impart much spin other than what's already there. This is essentially a defensive play style.
That's not to say that the ball doesn't have spin. I play with a lot of spin, and a recent level up in my game was that I realized that when I put spin on the ball and my opponent returns it without slicing, I've got to account for the spin I put on it because it's still there.
You can get pretty dang far with a primarily defensive style, though. I was recently at an informal local tournament where the 2nd place guy had basically one serve and an insane ability to return shots. A lot of the times he scored in the final game were where his opponent hit a shot with such insane speed and spin that when when he returned it, his opponent couldn't defend the residual speed/spin of his own shot.
Smart :)
How far are we from pickleball? I’d pay big money for a pickleball robot.
I'd be more impressed if this was in a humanoid robot.
I don't wanna brag, but i remember building something similar when i was 6 or years old... by folding up one side of the table.
interesting, I would expected this problem to be much simpler than it turns out
It is. It's just that we now live in an age where everything has to be solved with "AI" to get attention and justify the billions of investments. So we forget established tech and enter the era of AI enshittification where this kind of thing is considered an amazing achievement.
Had this been some high school kids I'd be impressed. But Google? C'mon.
it's a little disingenuous to claim human level performance at a sport when you use a robot arm on a track. Humans have to contend with a body not specifically suited to the problem.
[dead]
What in the world is this huge push to have robots?
Are people not good enough?
I think on a population / humanity level, there is no such thing as "good enough". There is always something new to experiment with, something to improve, something to change. This bother seems to propel humans since time immemorial.