Watch Google’s ping pong robot pull off a 340-hit rally • TechCrunch
As if it wasn’t enough for AI to hide humanity (figuratively at the moment) in every board game out there, Google AI has us all pinged pong as well. For now, they stress it’s “collaborative,” but at the rate at which these things are getting better, he’ll be dealing with the pros in no time.
The project, It’s called i-Sim2RealIt’s not just about table tennis but about building an automated system that can work with fast-paced and relatively unpredictable human behavior. Ping pong, AKA table Tennis, is heavily restricted (as opposed to playing basketball or cricket) and balances complexity and simplicity.
“Sim2Real” is a way of describing the process of creating artificial intelligence in which a machine learning model is taught what to do in a virtual or simulated environment, and then applies that knowledge in the real world. Essential when it takes years of trial and error to come up with a working model – doing it in a sim gives you years of real-time training in a matter of a few minutes or hours.
But it is not always possible to do something in simulation; For example, what if a robot needs to interact with a human? It’s not easy to simulate, so you need real-world data to start with. You end up with a chicken-and-egg problem: you don’t have the human data, because you need it to get the robot to interact with it and generate that data in the first place.
Google searchers He escaped this predicament By starting simple and creating a feedback loop:
[i-Sim2Real] It uses a simple model of human behavior as an approximate starting point and alternates between training in simulation and deployment in the real world. With each iteration, both the human behavior model and policy are revised.
It’s okay to start with a poor appreciation of human behavior, because the robot is also just beginning to learn. More real human data is collected with each game, improving accuracy and allowing AI to learn more.
The approach was successful enough that the team’s table tennis robot was able to run a 340-person race. check it out:
She is also able to return the ball to different areas, while not giving her exact mathematical precision, but is good enough that she can start to implement a strategy.
The team also tried a different approach to more goal-oriented behaviour, such as returning the ball to a very specific place from a variety of positions. Again, it’s not about creating the ultimate table tennis machine (though that’s a possible outcome nonetheless) but finding ways to effectively train with and for human interactions without making people repeat the same action thousands of times.
You can learn more about the technologies the Google team used in the video summary below: