Sign In

Communications of the ACM

ACM TechNews

Self-Driving Cars Can Learn a Lot By Playing Grand Theft Auto

View as: Print Mobile App Share:
An image from Grand Theft Auto in which different elements have been automatically annotated.

Researchers are using the "Grand Theft Auto" videogame to teach algorithms about driving.

Credit: Rockstar Games

Several research groups are using the "Grand Theft Auto" videogame to educate algorithms with potential application to self-driving vehicle navigation.

The researchers say the game's scenic elements are extremely realistic, and can be used to produce data on a par with that generated using real-world imagery.

A team from Intel Labs and Germany's Darmstadt University created a software layer positioned between the game and a computer's hardware, automatically classifying different objects in the game's road scenes. The team says this provides the labels that can be entered into a machine-learning algorithm, enabling it to recognize cars, pedestrians, and other objects displayed either in the game or on an actual street.

A key challenge in artificial intelligence that applies particularly to automated driving is the work involved in collecting and labeling real-world imagery, and the annotation required is not easily scalable at present, says University of British Columbia (UBC) postdoctoral researcher Alizera Shafaei. He and UBS professor Mark Schmidt demonstrated videogames can not only be used to train a computer-vision system, but also easily vary the environmental conditions found in training data. "We showed that this synthetic data is almost as good, or sometimes even better, than using real data for training," Shafaei says.

From Technology Review
View Full Article


Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account