Researchers from the University of Michigan have developed a new way of testing autopilot cars, with which you can not pass the billions of kilometers of road tests required for admission to use.
The process, which was developed using data from more than 40 million kilometers in the real world, can reduce the time required for evaluating robotic vehicles in potentially dangerous situations by 300 thousand to 100 million times. The researchers point out that this can save 99.9% of the time and cost of testing.
Professor of Mechanical Engineering at the University of Michigan Roger L. McCarthy argues that even the most advanced and large-scale efforts to test autopilot cars today do not justify the need for thorough testing of machines. Instead of an integrated approach that measures everything at once, the new method breaks up the process into separate components that can be retested during modeling, providing automatically controlled vehicles with compressed sets of the most difficult traffic situations. If, for example, you need to obtain data on the reaction of the mobile to an unexpected obstacle on the way, then efforts should be focused on this situation. As an unmanned vehicle will behave in "boring" moments, you can get a statistical analysis.
The researchers give a simple comparison: only 1,000 miles (1.6 thousand km) of tests are equivalent to 300 thousand - 100 million miles (482.8 thousand kilometers - 161 million km) in the real world. And let 100 million miles sound like an unnecessarily long distance, this is not enough. For researchers to get enough data to certify safety without a driver, we need complex and extremely rare road scenarios. For example, an accident that is fatal occurs only once every 100 million miles of drive.
For consumers to agree to get into a car without a driver, it is necessary to prove with 80% certainty that it is 90% safer than man-driven. For this, cars would have to go through 17.7 million km (11 million miles), either in the real world or in simulation. Even over a decade of round-the-clock testing in typical urban conditions, a little more than 3 million km (2 million miles) will be accumulated.
In addition, fully automated vehicles require a completely different type of check that is different from that used for today's cars. Even the questions asked by researchers are more complicated. Instead of finding out what happens in an accident, they should determine how successfully cars can prevent its occurrence.
Scientists drew the analogy of testing a car with a doctor's appointment. Methods of testing cars with a man’s control is a measurement of blood pressure or pulse, testing of unmanned vehicles is a patient IQ check.
To develop a four-stage accelerated approach, researchers analyzed data from 40.6 million real kilometers (25.2 million miles) collected by two projects by the University of Michigan Institute for Transportation Studies. Within two years, they employed about 3,000 cars and volunteers.
They then identified events that may contain “significant interactions” between the automatic vehicle and the driven person, and created a simulation that replaced the entire driving without incidents with these significant interactions. After that, they programmed the simulation in such a way that unmanned vehicles would treat people on the road as the main threat, and put vehicles driven by people randomly everywhere.
The programming phase was followed by mathematical tests to assess the risk and likelihood of certain results, including injuries, accidents and moments close to the accident. In conclusion, the researchers interpreted the results of accelerated testing, using the method of sampling by significance
, to find out how automatically controlled cars will statistically react to everyday situations on the road.
The process of rapid assessment can be performed for various potentially dangerous maneuvers. The research team has so far evaluated only two of the most common situations that, as expected, will lead to serious accidents: an unmanned vehicle following a human driver, and, conversely, a person in a car that follows an autonomous vehicle. The accuracy of the estimate was determined by conducting and comparing accelerated and real simulations. The team notices that it is necessary to continue research and consider other situations on the road.