Autonomous car uses AI to learn from its mistakes
A robot car that learns from experience, like a human motorist with years behind the wheel, has been tested for the first time.
The Volkswagen GTI was put through its paces at Stanford University in the US, and performed as well as a skilled racing car driver.
The idea was to create a control system that allowed the driverless car to handle unexpected conditions, such as ice and snow.
Current autonomous cars are good at making on-the-spot assessments of their environment. However, the new system incorporates data from past driving experiences.
As part of the project, one of the Stanford team’s autonomous vehicles, a Volkswagen GTI known as “Niki”, was let loose on an icy test track near the Arctic Circle.
Skidding around corners, the car’s computer “brain” learned from its mistakes and stored the information away.
Lead researcher Nathan Spielberg said:
“Our work is motivated by safety, and we want autonomous vehicles to work in many scenarios, from normal driving on high-friction asphalt to fast, low-friction driving in ice and snow.
“We want our algorithms to be as good as the best skilled drivers – and, hopefully, better.”
A major hurdle facing autonomous cars is dealing with unexpected emergencies. Safely recovering from a skid on ice, for instance, requires planning based on advanced information about the way the car is likely to behave.
To overcome this problem the Stanford researchers built an artificially intelligent neural network that “remembered” past driving experiences.
Tests of the system were conducted at Thunderhills Raceway, a motor sports complex in Willows, California.
Niki performed as well as another autonomous car, an Audi TTS nicknamed “Shelley”, which had been pre-loaded with information about the course and conditions.
Unlike Shelley, Niki had no pre-programming and had to rely on memories of its earlier driving experience.
Both cars achieved similar lap times to a skilled amateur racing driver.
Mr Spielberg said:
“With so many self-driving cars on the roads and in development, there is an abundance of data being generated from all kinds of driving scenarios.
“We wanted to build a neural network because there should be some way to make use of that data. If we can develop vehicles that have seen thousands of times more interactions than we have, we can hopefully make them safer.”
