As if having AI tan humanity’s hide (figuratively for now) at every board game in existence wasn’t bad enough, Google AI has one trying to wipe us all out at ping pong. The goal of the i-Sim2Real project, which is not only about playing ping pong, is to create a robotic system that can adapt to and function alongside quick-moving, unpredictable human behaviour.
A machine learning model is taught what to perform in a virtual environment or simulation, then uses that information in the real world. This process is known as “Sim2Real.” It’s essential because developing a functioning model may take years of trial and error; by simulating the process, years of real-time training can be completed in a matter of hours or minutes.
But not everything can be done in a sim; for instance, what if a robot has to communicate with a person? Since that is difficult to mimic, you must start with real-world information. [i-Sim2Real] alternates between training in simulation and deploying in the real world using a basic model of human behaviour as a rough starting point. Starting with a crude simulation of human behaviour is acceptable since the robot is still learning as well.
With each game, more actual human data is gathered, increasing accuracy and allowing the AI to gain more knowledge. Again, the goal here isn’t to build the perfect ping pong machine (though that is a possible side effect), but rather to identify effective ways to train for and with human interactions without having individuals perform the same activity repeatedly.
Again, the goal here isn’t to build the perfect ping pong machine (though that is a possible side effect), but rather to identify effective ways to train for and with human interactions without having individuals perform the same activity repeatedly.