The Grand Theft Auto video game franchise isn’t perhaps the most obvious means of developing real-world, autonomous driving capabilities, considering that general wrongdoing is, literally and figuratively, the name of the game; gunfire and flying semi trucks come to mind.

However, scientists from Darmstadt University in Germany have demonstrated that it is, in fact, not only possible, but feasible, to import data from GTA to use for teaching better autonomous on-road behaviour in real life.

Not only has it been found to be cheaper to extract data with this method compared to gathering real-world information, the researchers also said that data from modern computer games can be almost as good, and sometimes even better than real-world data.


Special software was created for this purpose, which discovers and classifies different kinds of objects within the open-world game’s virtual environment – vehicles, buildings, pedestrians, and the like. The scientists have extracted 25,000 frames from the game, simulating varying weather conditions and times of day.

The data was annotated in a total of 49 hours, compared to a process that would have taken days instead if they had used real-world data. According to TheNextWeb, this process is conducted in order to teach autonomous vehicles the difference between pedestrians and other objects when navigating in real life.

The German team aren’t alone in using the computer games approach to gathering data for self-driving vehicles; According to an article by MIT Technology Review, a group at Johns Hopkins University in Baltimore, USA is developing a tool to adapt a machine-learning process to any virtual environment built using the Unreal game engine.