For self-driving cars to ever be a genuine reality, they need to spend a lot of time driving and experiencing what’s called “reinforcement learning.” That’s basically a fancy way of saying “do the same task over and over again and learning how better to solve the problem it presents each time.”
In terms of driving down a highway, this is fine, but when things get more complex, “learning by doing” is really not how you want a car to figure out the best case scenario in a multi-car pileup or a small child sprinting into the road. So you need to simulate it, over and over, the problem being simulations are expensive. Unless you have sixty bucks and a basic gaming PC, it turns out.
Universe is a fascinating open-source AI project that uses video games as cheap simulators to hone the skills of various AI tools. The basic argument is that with the proper mods, video games are ready-built simulators for all sorts of AI tasks. One of those, DeepDrive, is teaching self-driving cars, and the environment chosen is, uh, Grand Theft Auto V. To be fair, they do turn the violence off, so this AI won’t learn to go off-road or ram a blockade when the police come after you. But, uh, even with those off, it is still GTA. When things go wrong, they go really wrong.
Joking aside, this is doing the grunt work of improving car safety. Even if we never truly get self-driving cars, automatic braking, blind-spot scanning, and other smart tools are making the cars we drive safer. Just hopefully nobody introduces DeepDrive to Trevor.
(Via VG 24/7)