Cars that drive themselves? That’s crazy talk! Well, it was crazy talk when the idea first became popular amongst sci-fi writers. Now, it’s a thing that actually exists. They’re far from a common sight though, and almost nobody has seen one outside of TV or the internet. That’s because the race to perfect driver AI is still far from over, one of several major challenges being teaching these things to recognise other cars on the road in the first place. One researcher at the University of Michigan has found that an excellent tool for teaching autonomous cars to identify other vehicles is… Grand Theft Auto V. Yes, really.
As reported over at New Scientist, Matthew-Johnson Roberson – Assistant Professor of Engineering at UoM – says: “One evening, after a long day of hand-labelling images, I was playing Grand Theft Auto V […] I thought, ‘this is just so realistic that it would be a perfect simulation of the real world’.
This isn’t just idle chatter. An algorithm based on GTA V was developed – and used – to recognise road vehicles. Amazingly, the GTA-trained program was just as good at recognising vehicles as one based on the traditional (and laborious) process of compiling individual images. Although the GTA algorithm needed about a hundred times more images to achieve this standard, New Scientist points out that half a million images can be generated from the game overnight.
Try not to panic; GTA V isn’t being used to train car behaviour (so far as we know). That, er, probably wouldn’t end well.