While automakers are still negotiating with local and state governments to let autonomous cars test drive on open streets, one programmer has found a more accessible proving ground to teach AI how to be a motorist: Grand Theft Auto V. It's not the first time folks have used the game to train their self-driving vehicles -- but you can watch this one learn in real-time on Twitch. One warning: If you're expecting a graceful, law-abiding AI driver...don't.
Programmer Harrison 'Sentdex' Kinsley created the AI (or "convolutional neural network"), named it Charles, and set it loose in the game to teach itself through deep learning. While that sounds advanced, so far Charles hasn't quite mastered avoiding collisions with cars, dividers, signs and people. If this AI hit the road today, it would have some real-life police after it quickly -- so long as it didn't hurl itself into the water first (a frequent fate on the livestream).
As Kinsley describes in the Twitch description, Charles "learns and takes all actions based on single frames at a time, and bases his decisions on just pixel data. Charles only sees exactly what you see." In other words, what a self-driving car would "see" if it just relied on cameras on real roads. What the AI can't do yet is remember: Kinsley didn't program in memory, forcing it to make split-second decisions one frame at a time, like so.
If you want to get into the nitty gritty, Kinsley documented his multi-part process building Charles in Python. Whether this AI becomes a better driver and validates educating neural networks through simulation, at least we can chuckle that even machines have trouble driving these games.