Self-Driving Cars Are Learning From ‘Grand Theft Auto,’ No Danger From Self-Aware Death Machines

In what might seem like the most ill-advised move by scientists regarding self-driving technology, researchers are now using the hyper-violent video game franchise “Grand Theft Auto” to teach cars how to navigate city streets. As anyone who is even slightly familiar with the games may know, “GTA” driving often involves manic destruction of property and high pedestrian casualties. However, the researchers are assuring people that there’s no way the program would result in destructive, self-aware vehicles.

According to MIT Technology Review, there are now several groups of researchers using “GTA” to teach self-driving software how to navigate roads in the real world. Thanks to the realistic structures of streets, traffic, and buildings in the games, these researchers believe that spending thousands of hours driving around the streets of Los Santos is going to make autonomous cars safer drivers.

This method is preferred over others mostly because of the existing world and realistic graphics that already exist within the games. Self-driving is possible through machine learning, but machine learning only works by accumulating mountains of data, which inevitably means thousands of hours and millions of miles driven.

Doing this in the real world is not ideal and building another 3D terrain similar to “GTA’s” environment would take too much time. By using the game, the algorithm can drive around town for as long as necessary to get the data the researchers need.

Among the institutions using this method to teach software to drive, include Intel Labs and Germany’s Darmstadt University, The Next Web reports. The scientists then created software that allows the algorithm to identify many of the objects within the virtual world including trees, buildings, pavements, and of course, people.

More than simply allowing the algorithm to gather the thousands of hours of driving needed, however, there are also simulations that are necessary for machine learning but are impractical to apply in the real world. Collisions are a good example of this since it would be too expensive to crash two vehicles together or one vehicle against a wall thousands of times in the real world. In “GTA,” car crashes are a dime a dozen.