Watch this AI learn to drive in GTA V on Twitch

Since the start of May, an AI called Charles has been trying its best to navigate the mean streets of San Andreas in Grand Theft Auto V. If that wasn’t hard enough, it’s had to do so in front of a live Twitch audience.

If you have a deep-rooted fear about our driverless car future, then this probably isn’t the live stream to tune into. Charles spends its time ploughing into oncoming traffic, running down pedestrians and driving straight into walls. That may sound quite similar to every other Grand Theft Auto stream on Twitch, but if this seems even more directionless, there’s good reason for it.

Charles is the work of programmer Harrison Kinsley and is learning to drive through deep learning, which involves much more trial and error than would be permitted by a human driving instructor on real roads. “At the moment, Charles learns and takes all actions based on single frames at a time, and bases his decisions on just pixel data,” Kinsley writes in the description. “Charles only sees exactly what you see.” In other words, what Charles sees at any given time is distinctly limited.

If it doesn’t seem to be getting any better, there’s good reason for that: “In time, I intend to give Charles some short-term memory to hopefully improve his driving.”

It’s a fascinating watch – I have it open in a second window. Just when you think Charles is beginning to get the hang of things, drifting deftly between cars like a slalom skier, it goes and does something stupid, like driving straight into the sea.

Still, considering its limitations, Charles seems to be doing a pretty good job where I’m sitting. Which is mercifully far away from the passenger seat. Still, I’d struggle to concentrate on the road too if I were constantly being hounded by police and sworn at by passing motorists. You go Charles: you’ll get there eventually. (Maybe.)

Leave a Reply

Your email address will not be published. Required fields are marked *

Disclaimer: Some pages on this site may include an affiliate link. This does not effect our editorial in any way.