Driverless cars taught by Grand Theft Auto
UNITED STATES: Driverless cars are training for real-life situations by playing Grand Theft Auto, the computer game known for chases, car jackings and an unsavoury cast of drug dealers, pimps and prostitutes.
Developers at automotive giants such as Ford and Google’s Waymo division are using the virtual world of the computer game to train their driving software while opportunities for real road trials remain limited.
The latest game in the GTA franchise covers an area almost a fifth of the size of Los Angeles, with 262 vehicle types, a complex road system, 14 weather conditions and more than 1000 different unpredictable pedestrians and animals.
Last year scientists at Darmstadt University of Technology and Intel Labs developed a way to pull visual information from GTA V, stripping out the hoodlums and drive-bys.
In January researchers at Openai, a non-profit artificial intelligence research company, released a tool allowing developers to test-drive their software in the game’s Los Santos setting.
Now developers are using the game, as well as other simulators, to train their algorithms to cope with real-world situations such as avoiding obstacles or having multiple cars changing lanes.
Alain Kornhauser, a professor who advises Princeton University’s autonmous vehicle engineering team, told Bloomberg that the game’s universe would never be a substitute for asphalt.
‘‘But it’s the richest virtual environment we could extract data from.’’
The industry continues to harness real-world data too. In a UK project announced yesterday traffic cameras will be used to train driverless cars how to learn from motorists’ bad behaviour.
This should deliver important insights for the transition period when semi-autonomous and driverless cars share the roads with cars driven by humans.
The government-funded trial will employ images from CCTV units to teach autonomous vehicles how to negotiate busy city centre streets and interact with human-driven cars.
Under the plans, experts will use images from some of the estimated 14,000 cameras in London, feeding footage of millions of manoeuvres into the AI system.
It is hoped that the project will lead to the creation of a fully autonomous Uber-style taxi service in the capital, possibly by 2021. - The Times