The cost and time associated with safety testing has hampered the push toward self-driving vehicles, but a new system developed at the University of Michigan shows that AI can reduce the test miles required by 99.99%.
It could trigger a paradigm shift that enables manufacturers to more quickly verify whether their autonomous vehicle technology can save lives and reduce crashes. In a simulated environment, AI-trained vehicles perform risky maneuvers, forcing AVs to make decisions that drivers rarely encounter on the road but are required to better train vehicles.
To encounter these types of situations frequently to collect data, test vehicles in the real world need to drive hundreds of millions to hundreds of billions of miles.
said Henry Liu, professor of civil engineering at UM and director of both Mcity and the Center for Connected and Automated Transportation, a regional transportation think tank funded by the U.S. Department of Transportation.
UM researchers refer to the problem as the “curse of scarcity,” and they address it by learning from real-world traffic data that contains safety-critical rare events. Testing on test tracks that simulate urban as well as highway driving has shown that AI-trained virtual vehicles can speed up the testing process thousands of times. The study appears on the cover of Nature.
“The AV test vehicles we use are real, but we created a mixed reality test environment. The rear vehicles are virtual, allowing us to train them to create difficult scenarios that rarely occur on the road,” Liu said. .
The UM team used a background vehicle training approach that removes safety junk from the driving data used in the simulation. Essentially, it gets rid of long distances when other drivers and pedestrians act in responsible, predictable ways — but it spares dangerous moments that require action, like another driver running a red light.
By using only safety-critical data to train the neural networks that make maneuvering decisions, test vehicles can experience more of those rare events in a shorter amount of time, making testing much cheaper.
said Xu Feng, an assistant professor in the Department of Automation at Tsinghua University and a former assistant research scientist at the UM Transportation Research Institute.
“It also opens the door to accelerated training of safety-critical autonomous systems by leveraging AI-driven test agents, which may create a symbiotic relationship between testing and training, accelerating both areas.”
Training, along with time and expense, is clearly a barrier. An October Bloomberg article said that although robot leader Waymo’s cars have logged 20 million miles over the past decade, more data is needed.
The author wrote: “This means that its cars must drive 25 times more than they do before we can say, even with a vague sense of certainty, that they cause fewer deaths than bus drivers.”
Testing took place in the urban environment of Mcity in Ann Arbor, as well as the highway test track at the American Center for Mobility in Ypsilanti.
Launched in 2015, Mcity was the world’s first purpose-built testing environment for connected and autonomous vehicles. With new support from the National Science Foundation, outside researchers will soon be able to conduct mixed reality tests remotely using both a simulated track and a physical test, similar to those described in this study.
The real-world datasets that support Mcity’s simulations are being collected from smart intersections in Ann Arbor and Detroit, with more intersections being prepared. Each intersection is equipped with privacy-preserving sensors to capture and classify each road user, determining their speed and direction. The research was funded by the Center for Connected and Automated Transportation and the National Science Foundation.