We know. That's a lot of quotation marks. Also the robot's baseball bat looks a bit like an oversized fly swatter. Whatever – this thing is still impressive.
The real-time cerebellum was created by Tadashi Yamazaki and Jun Igarashi, of Japan's RIKEN Brain Science Institute. An accelerometer located at the rear of the batting cage detects the flight path of the ball, and relays that information to a graphics processor (GPU), similar to what you might find in a gaming computer.
This GPU is the robot's artificial brain. According to the researchers, it comprises a large-scale spiking network model of a cerebellum, including "more than 100,000 spiking neuron units within realistic parameters." (The neurons in a spiking neural network, like the neurons in your brain, fire in accordance with the propagation of simulated neuron potentials. Basically, the GPU's artificial neurons obey the same go/no-go rules as the neurons in your head, making spiking neural networks some of the most realistic brain simulators out there.) The GPU does some number-crunching and tells the robot how best to move in order to make contact. But the GPU also learns from its mistake. Even if the robot swings and misses, it's only a matter of time until it's nailing every pitch (provided the ball is tossed at the same speed and along the same flight path every time). If the pitch-speed and path are changed, the robot learns again.
As Wired's Daniela Hernandez points out, this is not the first time researchers have used a neural network to control a robot:
A team of scientists in Europe, for instance, have used an artificial cerebellum to control a robotic limb. But according to [Yamazaki]... the baseball-playing robot is the second largest model of its kind and it runs in realtime, meaning its much faster than other systems. That means the GPU brain is better suited to controlling external hardware, he says.
The paper describing the research is published, free of charge, in Neural Networks.