The science of ping-pong

What can a robot table-tennis player tell us about skill?

Man vs. Machine

It’s been 27 years since IBM’s DeepBlue supercomputer shocked the world by beating the grandmaster Garry Kasparov in chess. AI has since proven capable of outperforming humans in a lot of things, and counting. Where we still have the edge clearly - very clearly - is sports.

But is that changing? Google last week introduced DeepMind’s first “AI-powered robotic table tennis player,” and let’s get one thing out of the way to start: It’s a big step up from the ridiculous 3-point shooting robot introduced by Toyota a few years back:

Google’s PongBot doesn’t loom menacingly — it doesn’t have any anthropomorphic traits at all. It’s an industrial robot arm that has been endowed with AI software and given a ping-pong racket. And in a study involving 29 human participants, the bot was capable of winning plenty of matches.

In fact, it didn’t lose any matches against beginning players, and it won 55 percent of its matches against intermediate opponents. It zips from side to side along the table and even adjusts itself back and forth to square each stroke before returning to the midline to await a return. According to Google, the bot struggled with extremely “fast or high balls, has difficulty reading intense spin, and shows weaker performance in backhand plays,” which is why the robot struggled to win at all against experienced or expert players. These are problems that, over time, with training and a few engineering tweaks, I would suspect that the robot could eventually overcome.

Google trained the AI by having it watch videos of a lot of ping-pong matches, and then by letting the system play real games against opponents. By the final round, the robot had learned from over 14,000 rally balls and 3,000 serves. Google says the computer performs by analyzing the game state at each moment, adapting to the opponent’s style, and choosing one of a few shots out of a repertoire it has developed, like a forehand or backhand, for each incoming ball.

In a nutshell, we do the same thing.

Fewer unforced errors

The bot has a few advantages over us. For one thing, it shouldn’t get nervous. There’s no threat of an automated table-tennis system choking under pressure or developing the yips in a tense bout. It won’t tire during a furious rally or overthink its next shot after flubbing the last. Take psychology, emotion, and fatigue out of the picture, and it’s good to be a robot athlete.

Robots also shouldn’t have problems moving. By this, I mean a few things. They’ll need no warmup, no stretching; just turn them on, they’ll be in peak form. Absent a mechanical breakdown, they’re not susceptible to injury. And their actions also aren’t subject to the natural perturbations that we encounter in our day-to-day movements. That’s called motor noise, and it’s mostly imperceptible — until we’re trying to do repetitive, fine tasks like signing our name or reaching for a glass and you may notice slight deviations, the result of the motor signal encountering minor turbulence on its route from the brain to the muscles. If motor noise didn’t exist, we’d never have a sport called darts — it’s part of what makes physical performance unpredictable and competitive. It’s what makes us human.

So the robot would be motorically pure, which is annoying, too, because assuming its calculations are correct, then each shot should be consistent, smooth, and flawless. Not even Roger Federer could say that.

Declarative vs. Procedural

What I’ve described above sounds like an unbeatable opponent, and yet Google’s bot couldn’t win against experts. What did the experts have that AI didn’t?

In the literature on skill learning, there is a very famous story about a man named Henry Molaison. Patient H.M., as he was known, had severe anterograde amnesia that left him with no short-term memory, forgetting events sometimes as little as 30 seconds after experiencing them. But his ability to draw and trace improved over time, despite having no memory of ever performing the task each time.

H.M. received declarative instructions on how to trace. Though he might not have remembered them, his brain was encoding the necessary signals to get his muscles to move in the right fashion. In its simplest form, this is how motor learning is achieved — going from declarative to procedural so that, over time, you don’t need the declarative anymore.

Why is that important? It frees the brain up for higher-order cognition. Serena Williams doesn’t need to hear the declarative instructions on how to serve each time she takes the court. She has encoded that information into procedural memory, with access to quick retrieval of it whenever needed. And without those cognitive resources spent thinking about the declarative, she can focus on strategizing, improvising, and improving upon those skills to become an even better … Serena Williams.

Analyzing thousands of hours of video gives a robot enough declarative information to beat an amateur, but taking its game to the next level will require a lot more practice — and maybe a certain creative spark that’s uniquely human. We’ll see.

Links

Why elite athletes are harnessing their own brain waves for sporting success … The advice mental health experts give experienced Olympians … Why do women suffer more ACL injuries? … Why contrast therapy could be the next big thing in Olympic recovery … How sprinters run the perfect race … The sports where women outperform men … The world’s fastest climber trains by playing chess … Why baseball’s switch hitter is trending toward extinction … What can Olympians teach CEOs? … Public health experts want IOC to dump Coca-Cola as sponsor for the good of the planet