Autonomous car technology is the Holy Grail for auto manufacturers today. Elon Musk claims self-driving cars will be as common as self-service elevators in the very near future. But first, the systems used to allow cars to drive themselves need to be validated. Currently, that means collecting data from billions of miles of real-world driving.
Tesla has already passed the 2 billion mile mark. Once its Model 3 hits the market, its pace of data collection will accelerate dramatically. Other manufacturers are scrambling to catch up but are years behind after leaving the starting gate late. Waymo, the autonomous car division of Google, and Uber are pushing forward with their data-collection programs, but don’t have anything close to the fleet that Tesla has. Self-driving cars are now being tested on public roads in California, Arizona, Michigan, Texas, and Pennsylvania.
Researchers at MCity at the University of Michigan announced this week that they have devised a new way to test autonomous car technology that will allow validation to occur in just a few thousand miles of driving instead of billions of miles. If so, its procedures would slash the cost of developing self-driving systems, and may help other automakers catch up with Tesla.
“Even the most advanced and largest scale efforts to test automated vehicles today fall woefully short of what is needed to thoroughly test these robotic cars,” said Huei Peng. He is the director of MCity and a professor of mechanical engineering at the University of Michigan.
Based on data collected from more than 25 million miles of real-world driving, the process created by the researchers eliminates the majority of miles during real-world driving when nothing special happens and concentrates on those situations that occur only rarely but often lead to serious collisions.
In almost all such cases, the wild card in the mix is a car driven by a human being. The accelerated evaluation process concentrates on just those moments so they can be simulated and tested repeatedly while eliminating the majority of uneventful miles driven in between.
As MIT professor Bryan Reimer told the NEMPA Technology Conference in Cambridge last week, computers are rational 100% of the time. People? Not so much. As he says, the biggest challenge for those designing self driving systems is taking into account the “nut behind the wheel.”
The researchers programmed their computer simulations to consider human drivers as the primary threat to automated vehicles. Their simulation inserted unexpected interactions with human drivers randomly throughout the analysis.
During real-world driving, the two most common areas of conflict occurred when an autonomous car was following behind a human driver and when a car driven by a person merged in front of an autonomous car. Other areas of conflict can be programmed into the accelerated analytics and studied without the necessity of conducting time-consuming and expensive real-world driving tests, the researchers claim.
Decreasing the time and expense devoted to autonomous car research is obviously a good thing. The question then becomes, do people really want to spend their hard-earned money on self-driving cars. Research by MIT indicates that interest in autonomous cars may be eroding. One person told the MIT team, “My iPhone doesn’t work 100% of the time. Why would I expect a self-driving car to?” Good question.
Source and graphic credit: University of Michigan
Hat tip to Leif Hansen