You might have heard that some folks making self-driving cars used Grand Theft Auto V to help train their AIs. You might have wondered what that looks like. You… might not learn that watching this Twitch livestream of an AI learning to drive in GTA V. The compubrain, named Charles, is less advanced, though Charles sounds fancy in creator Harrison ‘Sentdex’ Kinsley’s description: “a convolutional neural network that learns to drive through deep learning.” Charles often ends up atop highway barriers, ramming walls, or battered in ludicrous police chases. This makes Charles no less fun to watch.
Kinsley, a man robots will burn for trapping their forebears inside a crimeworld for education, explains:
“At the moment, 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 time, I intend to give Charles some short-term memory to hopefully improve his driving.”
Like with Brent Watanabe’s wonderful San Andreas Deer Cam (which is still my favourite botstream), I’m coming to recognise behaviours and quirks in how Charles sees the world, and enjoying simply watching Charles’s weird adventures. Charles is living on my second screen right now, pootling around and ploughing through pedestrians.
Occasionally Charles will seem almost like a real player, swerving to side-swipe a scooter then do a sweet flip or repeatedly driving over the same man.
Oh, Charles has just driven into the sea.
If you want to know more about how Charles works, check out this brief explanation of how Charles sees and Kinsley’s tutorial on building self-driving cars with the programming language Python. Yes, a magic snake hisses the words of life inside Charles’s brain.