Exploiting Hierarchy for Scalable Decision Making in Autonomous Driving.

A major challenge in autonomous driving has been the intractability of planning algorithms. Research has largely focused on simple, short-term scenarios with few interacting traffic participants. We propose a hierarchical approach for long-horizon tactical planning in large-scale autonomous driving settings. Our approach exploits the locality of interactions with other agents by sequentially setting and accomplishing short-term goals involving fewer agents and hence is able to scale to more traffic participants. We demonstrate the effectiveness of our approach on an example highway driving problem where the ego vehicle must safely transit to the farthest lane in order to exit the highway at a designated exit.