RAC Cars News


Grand Theft Auto training for driverless cars

By raccars Published


The popular Grand Theft Auto (GTA) game may not seem to be the ideal influence for driverless cars but it is still being used in their development.

Thankfully, it is not the frequent crashes, drive-bys and speeding that are commonplace in GTA that are being used in driverless cars research. Rather, it is GTA’s virtual world that is of interest.

A source of data for driverless cars

One of the major demands of driverless car technology is the need to be prepared for a vast range of different environments. The technology must be able to identify a multitude of different road types and configurations, plus be prepared for the addition of pedestrians and other vehicles.

One of the ways in which this technology can be made ready for as many scenarios as possible is by feeding the software with enormous volumes of visual data in order for it to ‘understand’ a wide variety of driving conditions.

A game such as Grand Theft Auto 5, therefore, can be used to provide high-fidelity road data which researchers can use for virtual cars to explore. Pedestrians and a range of other road users can also be added to prompt a reaction from the driverless car.

The game allows researchers to use the environmental data in algorithms without the safety issues and costs that would be involved in putting a real-life car onto the roads. It also makes the research process far quicker than it might otherwise be.

Other hyper-realistic games could also offer an efficient method of allowing artificial intelligence (AI) algorithms to ‘learn’ about real world conditions.

GTA’s widespread appeal

In Germany, researchers from Darmstadt University and Intel labs have developed a means of extracting this type of useful data from GTA.  Grand Theft Auto has long-been a favourite amongst gamers and now a number of different research groups are making use of the game to train their algorithms. Fortunately, the computers used will not be able to learn the bad behaviour and violence routinely displayed in GTA but realistic scenery should allow the machines to perceive real world elements correctly.

The use of GTA is reliant upon machine learning technology, which allows computers to perform a range of impressive tasks, ranging from identifying faces to recognising speech as accurately as a real person could. The constraints of this technology, however, relate to the time and practical challenges faced in compiling enough curated data to allow the computers to operate efficiently.

Some researchers involved in the development of driverless cars and other technology already use game engines to build their own 3D simulations to use as training data for algorithms. However, computer games, which can simply be bought off-the-shelf, can offer a simpler, quicker and cheaper alternative, whilst providing vast amounts of photorealistic imagery.

Extracting information for driverless cars

This is achieved by using a software layer which is placed between the computer’s hardware and the game. This can automatically classify different objects in various road scenes. Labels are then created which are fed to the researchers’ machine learning algorithm, making it possible for it to recognise cars, people and other objects.

The researchers claim that it would be virtually impossible for humans to label everything in the scenes with a similar amount of detail by hand. There is also the issue that some real life scenarios cannot be replicated easily in real life, such as a cars crashing into walls whilst travelling at high speeds.

It is also claimed that the researchers can improve training images by adding synthetic imagery, again aided by the use of games such as Grand Theft Auto.

Looking to Buy?
Search for cars