These self-driving cars are learning to drive like whom?

Most of the time, we want autonomous vehicles to take driving slowly and carefully. But we also want them to be able to eventually make snap decisions and quickly change their actions when the occasion calls for it, like human drivers do.

A team of engineers at Stanford University has set out to make sure self-driving cars are physically capable of making low-friction maneuvers at high speeds, akin to what a professional race car driver can do. They believe that learning these techniques will make self-driving cars better able to avoid collisions, especially at the last second.

The team used 200,000 motion samples from test drives in a range of weather and road conditions on a machine learning neural network. They installed their neural network in a self-driving Volkswagen GTI and compared its performance on the Sacramento Valley’s Thunderhill Raceway to that of a self-driving Audi TTS and a skilled amateur driver. The results were encouraging, though the team wants more data to further develop the system so it can perform in more varied conditions.