Autonomous vehicles learn to drive on the streets of GTA 5. However, the AI’s propensity for violence needs to be dialed back first.
The vast map of GTA 5 with its highways and off-road paths is used by many players daily as a racetrack. Whether in special races or in free roam mode, the different regions in Los Santos and Blaine County are well-suited for driving monster trucks, trucks, motorcycles, or supercars. Even driverless cars can learn a lot on the various tracks.
GTA 5 as a Simulator for the Real World
This project between Universe and DeepDrive demonstrates what driverless cars can learn from virtual situations and how they respond. Universe is a software platform that trains the intelligence of an AI. DeepDrive is an open-source AI agent for self-driving cars.
The map of GTA 5 is about one-fifth the size of Los Angeles. Various scenarios repeatedly occur to test the AI for autonomous driving. These scenarios arise from about 250 different vehicles, 14 different weather conditions, 7 types of bicycles, but also from objects at the roadside.
There are various regions such as mountains, deserts, snowy roads, freeways, or intersections in the city.
Violence Level Needs to Be Reduced
Not all scenarios from GTA 5 are well-suited for training the AI for autonomous vehicles. The mod that allows you to test the project on your PC reduces the propensity for violence of the people in the game. This way, no one calls the military, which ensures that you never again dent someone else’s car.
Although GTA 5 was originally developed for entertainment, this project shows how versatile the game can be. The vast map with its various terrains may be used for further tests in the future that can help us in the “real world”.
If you want to try out the project yourself, you just need a PC version of GTA 5. All further steps will be explained to you on the blog of Universe. Here is the link to the blog post on openai.com
The five coolest cars from the Import/Export DLC of GTA 5 can be found here.
