The new drone training methodology will help reduce the time required to study the various driving situations that a driverless vehicle must cope with minimal health risks. The results of this methodology of the Center for Connected and Automated Transport (CCAT) are published in Nature Communications

One of the obstacles preventing drones’ mass adoption is the ineffectiveness of existing methods for testing and evaluating such vehicles. There is no single standard for assessing the readiness of drones to drive on public roads to date. State of the art tests combines software simulation, closed-circuit testing, and real-world road testing.

The main problem is that road events of interest, including accidents, are too rare. Thus, systems can require hundreds of millions (sometimes billions) of kilometers to demonstrate the required safety performance. To date, Waymo has modeled only 15 billion km. Therefore, the work carried out by Dr. Liu and his team at the University of Michigan is aimed at creating a natural and competitive driving environment (NADE),

Liu built a simulated driving environment using large-scale driving data collected by the University of Michigan Transportation Research Institute (UMTRI). In this environment, “background” cars (those that simulate road traffic) are trained to perform certain hostile maneuvers towards the drone. This removes bias and improves efficiency.

NADE is a continuous learning method that enables continuous interaction between the drone and many background vehicles. For example, if a researcher wants to test his car in an urban environment, this approach would allow the drone to drive continuously and experience adversarial scenarios, including switching on and hard braking at a higher frequency. The results show that this environment eliminates the inefficiencies of the currently available options by orders of magnitude. This approach is expected to accelerate the adoption of autonomous vehicles.

“Driving a kilometer using augmented reality simulations overlaid on a test track equals hundreds or thousands of kilometers on public roads. This will result in a significant reduction in the overall cost and time of testing drones in a safer, more controlled, and repeatable test environment”, said ACM President and CEO Ruben Sarkar.