Now Anybody Can Use 'GTA V' to Develop a Self-Driving Car

Great news: Anyone with a computer and a copy of Grand Theft Auto V can develop autonomous car technology.

On Wednesday, OpenAI, the non-profit artificial intelligence company chaired by Elon Musk and Y-Combinator founder Sam Altman, announced that it was releasing an open-source integration of Universe, its artificial intelligence training software, with GTA V. Thanks to the DeepDive Project, a platform that turns GTA V into a researcher-friendly environment (making the NPCs non-violent is a main factor), development of self-driving technology can progress much faster than they can today.

“By integrating DeepDrive into Universe, [DeepDive creator Craig Quiter] has made it easier for people to start running DeepDrive,” Jack Clark, Strategy and Communications Director at OpenAI, tells Inverse. With Universe, it takes about 20 minutes to set up DeepDrive, and it can run on Linux or Mac systems. The release on Universe also includes a pre-trained A.I. that has seen about 21 hours of GTA V driving. People can use this as a basis for their system, or train their own, Clark says.

A growing body of research shows that using simulated environments can help deep-learning systems gather data that can be used in the real world. And as simulations go, GTA V is an extensive, detailed world about a fifth of the size of Los Angeles — easily modified to any size or city. Not only does it have winding city streets, but also mountains, deserts, and highways that you can explore in 257 different cars and 7 types of bike. Oh, and it has 14 different weather simulations. Minus the actual gameplay, it’s basically the dream test-driving range for autonomous cars.

Various companies, including Intel have modified GTA V to test artificial intelligence for self-driving cars, but they haven’t released these modifications to the larger research community.

In 2015, Quiter started DeepDrive as a way to configure GTA V so it could be used to develop and train artificial intelligence for self-driving cars. The system required a local PC computer and almost an entire day to install. In July of 2016, he started working with OpenAI to integrate DeepDrive with Universe, which allows researchers to test artificial intelligence systems with open source programs. And now, it’s open-source.

With the range of simulated environments and vehicles, using DeepDrive on Universe allows researchers a huge amount of flexibility in their training. But the real benefit of using the system is its speed. “Right now, you need a real car to test the end-to-end performance of a self-driving car,” says Quiter. “Researchers can use this to iterate on A.I. much faster than with a real car.”

So DeepDrive can speed up the testing process immensely, and reduce the costs of testing both virtually and on the road. It takes thousands of hours of road experience to train A.I. system on fluke safety events, like sliding on ice in traffic. So running your A.I. through a simulation, where you can throw it into those events at no cost to lives or an actual vehicle is a huge savings. This could also make the autonomous systems being road tested that much safer, because they already experienced any number of fluke events in the simulation.

Quiter says he is working on creating better benchmark measures of the different A.I.s, so that researchers can compete to see who has designed the best driving A.I. Now you just have to wait for the mod that lets researchers race their A.I.s against each other while all playing as the satirical version of the Google self-driving car.