Grand Theft Auto V: The Rise And Fall Of The DIY Self-Driving Car Lab
A few years ago, artificial intelligence researchers discovered thatย Grand Theft Auto V, the blockbuster 2013 video game, was good for more than stealing cars andย causing mayhem on the fictional mean streets of Los Santos. Its realistic graphics and physics engine provide an excellent virtualย environment for training self-driving cars. The game’s sprawling urban and rural environments are populated withย pedestrians, drivers and animals that dynamically interact with the AI. Training an autonomous car systemย is all the more efficient whenย running over a pedestrianย doesn’t result in real-lifeย injuries and legal troubles.
The fun didn’t last long, though.
Over the past year,ย GTA V’sย publisher, Take-Two Interactive, has quietly shut down a number ofย high-profile projects with cease-and-desist letters, according to multiple AI researchers involved in theย projects.ย Take-Two is slamming the breaks on more than just commercial endeavors. Academic researchers haveย also been dealt with aggressively.
In the most notable case, a joint project between OpenAI, theย Elon Musk-backed research group, and DeepDrive, a platform for training self-driving cars inย GTA V, was shut down. The collaboration allowed OpenAI’s bot, called Universe,ย to learn how to drive a car inย GTA V.ย The project was generating interest from researchers, but it didn’t take long forย the game publisherย to also take notice.ย Aย January 11, 2017 OpenAI blog post announcing the project was pulled downย and the code onย GitHub withdrawn.ย “We took down our project following the receipt of a letter from Take-Two,” said an OpenAI spokesperson.
DeepDrive was a personal project of engineerย Craig Quiter, who joined Uber-owned self-driving car group Otto last November.ย It’sย now defunct. “DeepDrive started with a mission to allow more people to work on self-driving AI, and the interest it received was absolutely incredible – something I am truly grateful to you all for,” reads a note from Quiterย on the site’s home page.ย “Unfortunately, however, we are restricted from providing resources usingย GTA Vย for legal reasons.”
Princeton’s Autonomous Vehicle Engineeringย program had been usingย GTA Vย for about two and a half years,ย before it shut down earlier in 2017.ย “We have used [GTA V]ย as part of our research in ‘proof of concept’ efforts andย [Take-Two is] not happy about it,” said Alain Kornhauser, a professor at Princeton University and faculty chair of the school’s autonomous vehicleย program, in an email.
Theย Princeton lab is now building its own simulatorย called “MaxiumusFurtum IV,”ย Kornhauser said.
Grand Theft Auto’sย developer, Rockstar Games, didn’t respond to multiple requests for comment, but has previously posted aboutย its policy on single-player modding:ย “Take-Two has agreed that it generally will not take legal action against third-party projects involving Rockstarโs PC games that are single-player, non-commercial, and respect the intellectual property (IP) rights of third parties.”
The companyย sent a statement in the past to Mashableย in response to AI researchers use of the game: “We welcome discussions about the use of our technology to help further academic research, but itโs obviously not appropriate for corporations to take our work and use it for their own financial interests or for researchers to distribute unlicensed copies of our code as part of their work without first seeking our permission.”
For the gaming world, Take-Two’sย tactic comes as no surprise. The game publisher’s aggressiveย pursuit ofย game modification software, whichย allows users to alterย gameplay logic and asset files,ย is well known.ย Most notoriously, the company went after popular modding tool OpenIV, butย backpedaled after backlash from fans.
Video games — not justย GTA Vย — have become an increasingly popular methodย for training AI systems how to solve complex problems. A technique known asย reinforcement learning enables a computer to learn how to accomplish tasksย through trial and error. The AI developerย sets a reward signal for aย specific objective, such as staying on the road. The computer (also called an agent) attempts to achieveย the reward by trying sometimes millions of times. Reinforcement learning works well in a game environment where errors — like crashing a car — won’tย cause real damage.
The AI can readย the game’s image frames as if it were the real world by applying deep learning techniques such as a convolutional neural network for image recognition. And by digging down deeper into the game’s code, AI developers canย simulate sensors important to real self-driving cars like LiDAR and radar.
Although usingย GTA Vย to train autonomous cars captured widespread attention in the press, the reality of using the game to train self-driving car systems for the real world may have never gone very far.ย Davide Bacchet, principal engineer of autonomous driving atย China-based autonomous car startup NIO, toldย Forbes:ย “GTAย has been extensively used for research by universities and many people around the world, but nobody in the Autonomous Driving team here in NIO ever considered using it, because we are building a safety critical system, and the level of certification that is involved is completely different.”
Bacchet saidย Nio makes its own virtual simulator to train its neural networks.
“I personally agree with the position taken by [Take-Two],”ย Bacchetย continued. “If people were using their game for anything outside of pure leisure, they should have been informed.”
Similarly, Alphabet’s autonomous car unit, Waymo,ย has developed its own AIย simulation environment it calls Carcraft and has recreated the cities ofย Phoenix, Austin and Mountain View.ย Waymo has 25,000 virtual self-driving cars navigating 8 million miles per day through the three virtual cities,ย according to the Atlantic.
Take-Two, however, hasn’t shut down all the AI projects associated with its game. Stephan Richter, a PhD candidate at the Technical University of Darmstadt in Germany,ย continues to publish research usingย GTA Vย related toย visual perception systems.ย Richter declined to answer specific questions around his research group’s relationshipย with theย GTA Vย developer, except that he has “taken care to comply with Rockstar Games’ policies,” he said. “All data we publish or have published is for research and education only.”
Harrison Kinsley, an independent Python programmer based in Texas, also continues to play around withย GTA Vย for his own custom self-driving AI system that he’s built for fun.ย Kinsleyย suspects Rockstar and Take-Two only goes after AI researchers who are mucking around inย the game’s underlying code. He said he doesn’t modify the code and his AI only reads the game’s visual information toย drive autonomously.
Kinsleyย maintains a 24/7 livestream of his AI — which he calls Charles — on Twitch. He trained the AI how to drive with only five hours of playing time (no reinforcement learning). Charles the AI is taught to drive as fast as possible and doesn’t bother following the law. The AI careens haphazardly down the road, frequently hitting pedestrians and cars along the way. If the AI loses the vehicleย it’s driving, it immediately steals another. โI wouldnโt get into a car that has this AI,โ he said.