Safety testing of autonomous vehicles has proven to be a time- and cost-intensive process. These vehicles are trained by artificial intelligence (AI) to navigate and respond to dangerous situations rarely confronted by drivers. To generate the data needed to test these vehicles in a simulated environment, real world test vehicles would need to drive millions to billions of miles to secure sufficient data pertaining to these infrequent safety critical conditions.

A new intelligent testing environment developed by researchers from Tsinghua University (China), the University of Michigan and Beihang University (China) can help automotive engineers steer clear of this testing impasse. The mixed reality testing environment combines real-world traffic data that contains rare safety-critical events with a dense deep-reinforcement-learning approach to train virtual autonomous vehicles.

Testing conducted on test tracks replicating urban roads and highways has demonstrated that the AI-trained virtual vehicles can accelerate the testing process by thousands of times. The study appears on the cover of Nature.

The approach applied to train the background vehicles strips away non-safety-critical information from the driving data used in the simulation. Time spans when other drivers and pedestrians behave in responsible, expected ways are eliminated, enabling tests to focus on dangerous conditions and moments that require action.

A new virtual testing environment simulates dangerous driving scenarios to quickly train autonomous vehicles. Source: University of Michigan

Reliance on only safety-critical data to train the neural networks that make maneuver decisions allows test vehicles to encounter more of those rare events in a shorter amount of time and drives down the cost of testing.

Testing was conducted at the highway test track at the American Center for Mobility in Ypsilanti, Michigan, and the University of Michigan Mcity Test Facility built for connected and autonomous vehicles. Outside researchers will soon be able to access this facility to conduct remote, mixed reality tests using both the simulation and physical test track, similar to those reported in Nature.