A computer learns how to make dribles as the best basket player

The players of basketball need a lot of practice to become good at the art of dribbling, and it seems to be that this is true when you consider players with computer animation also. Using deep learning reinforced, the players in a video game of basketball can be discovered thanks to the motion capture, how to improve their skills dribles.

VIDEO: This Artificial Intelligence is Wally in fraction of seconds

This process of trial and error takes much time and requires many millions of attempts, but the result is interesting, because the movement of the arms is coordinated in a plausible way with the physical movements of the ball. Players learn how to dribble between the legs, behind them and even make movements cross, to make the transition from one skill to another. “Once you learn the skills, new movements can be simulated much more quickly in real time”, says Jessica Hodgins, professor in the department of computer science and robotics at Carnegie Mellon.

Hodgins and Libin Liu, head of science at DeepMotion, will present its progress at the SIGGRAPH 2018, the Conference on Graphics Computers and Interactive Techniques, which will be held from 12 to 18 August in Vancouver, Canada.

“This research opens the door to the sports simulated skills required of the online avatars,” said Liu. “The technology can be applied beyond the sports simulation to create characters more interactive games, animations, motion analysis, and in the future, robotics”, stated the researchers.

Since the motion capture adds realism to the state of the art of video games, but these games sometimes include devices puzzling, notes Liu, as the balls follow paths impossible, or stick to the player’s hand. The method of the researchers has to do with the physical, which ultimately speaks to us of a simulation, which allows, potentially at least, to make games more realistic. The problems start when it comes to the movements of, for example, a basketbolista, as by dribbling, in contact with the ball is very short and the position of the fingers is critical. Some details, for example, as the ball continues to spin briefly when you touch the fingers of the player are difficult to reproduce. And once you release the ball, the player must anticipate when and where the ball will be.

Liu and Hodgins have opted to use learning reinforced deep to allow the model to get the important details. The AI programs that use this form of deep learning have been used in a variety of video games and in the program AlphaGo, who beat the best player in the oriental board game a couple of years ago.

Motion capture is used as input data to see how people were doing things like rotating the ball, rotate the waist, dribble while running and dribble with the right hand, or when exchanging hands. This motion capture does not include the movements of the ball, which Lui explains, are difficult to reproduce with precision. Instead, we used trajectory optimization to calculate the most likely path of the ball to a certain movement of the hand.

The program learned the skill in two stages. First, how to move and control the arms and hands, and then, the movement of the ball. This approach decoupled is sufficient for actions such as dribble and even make pirouettes with the ball, where the interaction between the character and the object have no effect on the balance sheet of the character.

Coolest-hacks.com and Partners.