The idea of autonomous cars is no longer a dream of the far distant future. These vehicles are here right now and they are testing on public roads across the country. The reason we don’t all own one is largely due to the snail’s pace at which testing is done. New research found a way speed up testing and put fully autonomous cars in our driveways much sooner.
A study out of the University of Michigan determined the problem is the number of testing miles that don’t help developers learn anything. Think of it like your average road trip.
Most of the time you’re cruising down the highway listening to music and sucking down the caffeinated beverage of your choice to stay awake. It’s miles and miles that don’t require much of the driver. The same is true for autonomous vehicles.
The problem with that scenario is it doesn’t let researchers learn anything. They need a way to test more complicated driving scenarios. They need a way to see how the technology will perform in rush hour traffic in downtown Boston at an intersection with a broken traffic light. That kind of stressful situation for human drivers is exactly what needs to be duplicated for autonomous test cars.
Huei Peng, director of Mcity and the Roger L. McCarthy Professor of Mechanical Engineering at University of Michigan and Ding Zhao, assistant research scientist in the University of Michigan Department of Mechanical Engineering have a solution to the problem. In their white paper, the researchers propose a new way of testing that would require just 1,000 miles of driving to get the same results as up to 100 million miles of real-world testing.
The pair analyzed data collected in over 25.2 million miles of real-world driving during two programs. It involved nearly 3,000 vehicles driven by volunteers in several University of Michigan projects. They used that data to figure out which scenarios were most useful to improving autonomous technology.
Once they figured out which scenarios were the important ones to have in testing, they broke them down into their component parts. This will allow researchers to test these situations in simulations, repeating them over and over and learning from them rather than having to drive millions of miles hoping those same scenarios will randomly occur.
Their estimates put the number of real-world miles required to do this at 11 billion. Unfortunately, A decade of constant testing in urban conditions will yield only 2 million miles. A more efficient way to conduct autonomous testing with a focus on the areas that are the most challenging could dramatically reduce the time we’ll have to wait for fully autonomous cars.