Driverless cars learn road safety from Grand Theft Auto 5
The popular driving computer game is being used as a simulation platform to train driverless cars to respond to obstacles, different weather conditions and chaotic driving.
With its disordered roads, criminal protagonists and controversies with each release, the Grand Theft Auto (GTA) games may seem like unlikely candidates when it comes to helping artificial intelligence (AI) operate a car safely.
However, the richness of the virtual environment and the many possible dangers it simulates may make it an ideal simulation platform for driverless cars to be tested.
GTA 5, produced by Rockstar Games, is a video game in which the player commits heists on the roads of the fictional state of San Andreas. It is one of the best-selling games of all time, having shipped 75 million copies.
The game features an open world design, containing more than a thousand roaming characters, 262 different vehicles, multiple different weather and lighting conditions and complex infrastructure, such as bridges and tunnels.
Professor Alain Kornhauser, professor of operations research and financial engineering at Princeton University described the game as “the richest virtual environment that we could extract data from”.
Driverless cars under development are undertaking extensive testing to ensure that they operate safely. Much of this testing is taking place on real roads: Apple has just been granted permission to begin testing its self-driving cars on roads in California.
However, it is not feasible for these cars to drive enough real-world miles to learn how to operate safely, so their AI systems also undertake simulated test drives.
Using video games for these virtual drives is cheaper, safer and offers engineers greater control over testing each variable in the more controlled environment. GTA 5 in particular has been singled out as a good candidate platform, as it mimics a range of hazardous and chaotic conditions.
“Just relying on data from the roads is not practical,” said Davide Bacchet, head of driverless car start-up Nio, told Bloomberg. Nio plans to introduce a driverless electric car in the US by 2020.
“With simulation, you can run the same scenario over and over again for infinite times, then test it again.”
Last year, researchers at Intel Labs and Darmstadt University developed a method to extract training data from GTA 5, using software to classify different objects in the road scenes and feed this information back to a machine learning algorithm. The algorithm uses this information to recognise cars, pedestrians and other objects.
AI systems have long been tested and trained using traditional games and video games, including Chess, Pacman and Angry Birds. More recently, Google DeepMind and Blizzard Entertainment announced that they would collaborate to use the military strategy game StarCraft 2 to test DeepMind’s ability to solve complex new problems. These games allow for researchers to use game scores as a simple indicator of an AI’s performance.
Test drives in GTA 5 provide just one platform for driverless cars for safety. Safety requirements for driverless cars remain under discussion, but are likely to compromise extensive real and simulated test drives.