11715d433f6f8b9106baae0df023deb3-Paper-Conference.pdf

BC formulates imitation learning as a supervised learning problem. It needs no in-environment samples, but it suffers from the covariate shift issue [37], often leading totesttimeperformance degradation.