Researchers from DeepMotion and CMU have learned an Artificial Intelligence system to dribble with a basketball. How did a computer learn to imitate such a specific human movement?
For this the researchers used a method with which animation characters can learn to dribble through experience. First, the system learned basic movements. If it can do these movements, it also learns new and more complicated movements easier and faster. By practicing and making mistakes, the algorithm learns to manipulate the ball as well as possible.
The movements learned the system by imitating human movements. The ball itself moves too fast to transfer its details to the computer, so the researchers added it later. With each movement, the most likely direction of the ball was calculated.
The avatar always learns to adapt to other situations. Through positive feedback he learns to manipulate the ball, based on knowledge about how the body should move. The avatar can also dribble if the floor is not completely straight, for example. Eventually the avatar can hold the ball between his legs, dribble behind the back and make crossing movements. The avatar even manages to switch fluently between movements.