As self-driving cars race to market, their manufacturers are turning to video games, including Grand Theft Auto (GTA), to teach their autonomous vehicles.
The strategy is meant to save thousands of hours on the road worth of experience building.
Self-driving cars are powered by a range of radar, sensors, and artificial intelligence software. To inform its decision-making, that AI software needs data.
While that information can be recorded by a human-piloted car driving around, it’s far more efficient to simulate the experience because the same scenario can run ad infinitum to test the algorithm, Bloomberg reports.
GTA, a popular video game that focuses organized crime my seem at first blush a surprising choice for this purpose since it has garnered previous media attention for its excessive violence, drug use, and sexual content. However, GTA’s open-world layout, an expansive map with hundreds of cars and thousands of unpredictable scenarios, makes for “richest virtual environment that we could extract data from,” a Princeton simulation researcher told Bloomberg.
Technology Review reports computer vision researchers from Intel Labs and Darmstadt University showed in a 2016 paper they could extract useful information from GTA in 7 seconds compared to 90 seconds needed for a real-world image.