Self-propelled cars are trained to be overly cautious, but there may be situations where they need to make quick maneuvers to avoid collision. Can these vehicles, festooned with tens of thousands of dollars worth of high-tech sensors and programmed to run at grandmother speeds, handle these split-second decisions as a human being?
Engineers at Stanford University may have the answer. They created a neural network that allows the driver without a car to perform high-speed, low-friction maneuvers as well as race drivers. When they finally arrive, leading cars will need capacities beyond human, as 94 percent of crashes can be attributed to human error. Scientists say this is an important step in improving the autonomous vehicle's ability to avoid accidents.
"We want our algorithms to be as good as the best-qualified drivers – and hopefully better," says Nathan Spielberg, a graduate student in mechanical engineering at Stanford and the lead author of a paper published earlier this week in the journal Science Robotics . "Our work is motivated by security, and we want autonomous vehicles to work in many scenarios, from normal driving on high friction asphalt to fast, friction driving in ice and snow."
The team used a type of artificial intelligence algorithm called a neural network that is loosely based on neural networks in our brains to create self-propelled systems. Neural networks are a type of machine learning, where programmers build models that seep through large stocks of data and look for patterns. These networks are used to drive an autonomous vehicle's "brain", typically high-powered GPUs, stored in the luggage of each vehicle that controls the decision-making process.
The Stanford team trained a neural network with data from 200,000 motion samples, including test drives on smooth surfaces such as snow and ice. They then took their system to Thunderhill Raceway in the Sacramento Valley to test it out. The Stanford team used two self-propelled cars in their test: Niki, an autonomous Volkswagen GTI, and Shelley, an autonomous Audi TTS.
First, Shelley ran controlled by the physics-based autonomous system, pre-loaded with set of course information and conditions. Compared to the same course in 10 consecutive trials, Shelley and a skilled amateur driver generated comparable laps. The researchers then loaded Niki with their new neural network system. The car performed similarly on both the learned and physics-based systems, although the neural network lacked explicit information on road friction.
The team says they were encouraged by the results, but stressed that their neural network system was not working well in any environment. They demand – what more? – more data to handle a wider range of conditions.