Driving intelligence tests are critical for autonomous vehicle (AV) development and deployment. Current methods use naturalistic driving environments (NDEs) in simulations, but require hundreds of millions of miles to demonstrate safety, which is inefficient. This paper introduces a naturalistic and adversarial driving environment (NADE) that uses sparse, intelligent adjustments to the NDE to significantly reduce required test miles without compromising evaluation unbiasedness. Background vehicles are trained to execute adversarial maneuvers at specific moments, creating an intelligent testing environment. The effectiveness of NADE is demonstrated in a highway-driving simulation, showing a multiple-order-of-magnitude acceleration in the evaluation process compared to NDE.

Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, Henry X. Liu