What makes a video game good? Not being just a video game
I wonder if there's someone who hasn't heard of Grand Theft Auto Five. This game is so awesome millions are still playing it online on a daily basis, despite it being three years old.
Take-Two Interactive had invested almost $140 million developing this amazing interactive open world platform (and some other $130 million in marketing), numbers that translated into more than 60 million copies sold (and a gross profit of almost $2.5 billion).
OK, let's get to our news now. If you've played GTAV, did you ever hail a cab without skipping the drive to your destination? The game gives you this opportunity – in exchange for a higher fare, naturally – so you can just "skip" the drive. Try to avoid "hurry up" command, though, as that can lead you to death-by-taxi.
Most of the time, I did skip it also, but then it strouck me: is this how it would be to be driven around autonomously? I mean, while you annoy/amuse yourself over the never-to-be-forgotten "El sonidito" song most taxis feature, the car you're in actually drives itself, with no human corrections needed.
You afterwards realize this is not just a scripted, syncopation-prone ride. No, the "car" is adapting to whatever situation encounters, that even involving going around an obstacle that's blocking the street. However, it recognizes humans, so it doesn't go full Carmageddon.
Consequently, a huge open-world already featuring a lot of real life like elements might be as well an amazing playground for self-driving software to be taught reading the world around it.
That's what a group of researchers from Inter Labs and Darmstadt University in Germany has actually done – create a "software layer" between the game and a computer's hardware. The layer classifies the objects appearing on the screen, thus helping the software "learn" faster – the simple reason being that, in real life, it wouldn't encounter as many different situations as in a game.
“With artificial environments, we can effortlessly gather precisely annotated data at a larger scale with a considerable amount of variation in lighting and climate settings,” Alireza Shafaei, a Ph.D. student at the University of British Columbia says. “We showed that this synthetic data is almost as good, or sometimes even better than using real data for training.” Shafaei is the coauthor of a paper showing how video games can be used to train a computer vision system.
via TechnologyReview