Autonomous Navigation without HD Prior Maps

Abstract

Most fielded autonomous driving systems currently rely on High Definition (HD) prior maps both to localize, and to retrieve detailed geometric and semantic information about the environment. This information is necessary to enable safe operation of many downstream driving components including, prediction, planning, and control. However, this requirement has raised issues with scalability, confining autonomous systems to small test regions where such detailed maps can be maintained. Furthermore, the reliance on HD maps can prevent autonomous vehicles from realizing human-like flexibility to both explore new areas and successfully navigate in rapidly changing environments or weather conditions. In this thesis, we present MapLite, an autonomous navigation system using only Standard Definition (SD) prior maps, in conjunction with onboard perception to directly infer the necessary HD map online. We also explore the use of a Localizing Ground Penetrating Radar (LGPR) for precise localization using stable underground features that are robust to changing weather conditions. Together, these methods can reduce the requirement for HD prior maps and bring autonomous navigation closer to human levels of flexibility and robustness.