Title: Intel­li­gent Dri­ving Intel­li­gence Test for Autonomous Vehi­cles with Nat­u­ral­is­tic and Adver­sar­i­al Dri­ving Envi­ron­ment
Speak­er: Hen­ry Liu
Abstract: Dri­ving intel­li­gence tests are crit­i­cal to the devel­op­ment and deploy­ment of autonomous vehi­cles. The pre­vail­ing approach tests autonomous vehi­cles in life-like sim­u­la­tions of the nat­u­ral­is­tic dri­ving envi­ron­ment. How­ev­er, due to the high dimen­sion­al­i­ty of the envi­ron­ment and the rareness of safe­ty-crit­i­cal events, hun­dreds of mil­lions of miles would be required to demon­strate the safe­ty per­for­mance of autonomous vehi­cles, which is severe­ly inef­fi­cient. We dis­cov­er that sparse but adver­sar­i­al adjust­ments to the nat­u­ral­is­tic dri­ving envi­ron­ment, result­ing in the nat­u­ral­is­tic and adver­sar­i­al dri­ving envi­ron­ment, can sig­nif­i­cant­ly reduce the required test miles with­out loss of eval­u­a­tion unbi­ased­ness. By train­ing the back­ground vehi­cles to learn when to exe­cute what adver­sar­i­al maneu­ver, the pro­posed envi­ron­ment becomes an intel­li­gent envi­ron­ment for dri­ving intel­li­gence test­ing. We demon­strate the effec­tive­ness of the pro­posed envi­ron­ment in a high­way-dri­ving sim­u­la­tion. Com­par­ing with the nat­u­ral­is­tic dri­ving envi­ron­ment, the pro­posed envi­ron­ment can accel­er­ate the eval­u­a­tion process by mul­ti­ple orders of magnitude.