Towards Learning Generalizable Driving Policies from Restricted Latent\n Representations

Abstract

Training intelligent agents that can drive autonomously in various urban and\nhighway scenarios has been a hot topic in the robotics society within the last\ndecades. However, the diversity of driving environments in terms of road\ntopology and positioning of the neighboring vehicles makes this problem very\nchallenging. It goes without saying that although scenario-specific driving\npolicies for autonomous driving are promising and can improve transportation\nsafety and efficiency, they are clearly not a universal scalable solution.\nInstead, we seek decision-making schemes and driving policies that can\ngeneralize to novel and unseen environments. In this work, we capitalize on the\nkey idea that human drivers learn abstract representations of their\nsurroundings that are fairly similar among various driving scenarios and\nenvironments. Through these representations, human drivers are able to quickly\nadapt to novel environments and drive in unseen conditions. Formally, through\nimposing an information bottleneck, we extract a latent representation that\nminimizes the \\textit{distance} -- a quantification that we introduce to gauge\nthe similarity among different driving configurations -- between driving\nscenarios. This latent space is then employed as the input to a Q-learning\nmodule to learn generalizable driving policies. Our experiments revealed that,\nusing this latent representation can reduce the number of crashes to about\nhalf.\n