You've heard of Michael 'Air' Jordan – well, get ready for 'AI-R' Jordan
We apologize in advance for this machine-learning basketball player pun
Skynet is getting closer. Ish. Artificially intelligent software has now picked up a devastating new skill after observing humans. It can now, er, dribble a basketball. Boomshakalaka!
A pair of brainiacs – Jessica Hodgins at Carnegie Mellon University in the US and Libin Liu at DeepMotion, a Silicon Valley startup creating realistic graphics using AI – have trained a virtual character that can show off its believable b-ball skills in real time.
It’s harder than it sounds, honest. The joints in the arms, hands, and fingers have to be realistically simulated in order to demonstrate its basketball powers. And even though it exists in a virtual world, the character has to take gravity into account, and maintain its balance and control while moving. Here’s a video of the player, and how it was trained...
The deep-learning-based system, dubbed DeepDribble, is trained to mimic humans' poses from a dataset containing motion-capture video of people playing basketball, by computing two components. One measures the position of the joints in the arms, hands, and fingers to manipulate the basketball, and another is concerned with the joints in the rest of the body. Both components are trained separately using a reinforcement-learning algorithm.
A separate neural network is required for performing various basketball skills, such as dribbling between the legs of opponents, and crossover moves. “Once the skills are learned, new motions can be simulated much faster than real-time," explained Hodgins, a Carnegie Mellon professor of computer science and robotics, this week.
So, what will this be used for? Liu, chief scientist at DeepMotion, said it “opens the door to simulating sports with skilled virtual avatars.” It’ll help developers craft more realistic characters in games, animation, and possibly even robotics. A more difficult sport to tackle is football because the balance of a player is trickier to model, he added.
The research is expected to be presented at this year's SIGGRAPH conference in Vancouver, Canada, from August 12 to 16. ®