CHAMP: Integrated Logic with Reinforcement Learning for Hybrid Decision Making for Autonomous Vehicle Planning

A Cognitive Hybrid Autonomous Motion Planner (CHAMP) is developed for autonomous driving applications in challenging driving scenarios. The proposed hybrid planner unifies a hierarchical rule-based decision-making architecture with Reinforcement Learning (RL). For challenging intersection scenarios, RL agents are trained to replace a subset of the rules in the logical planner. The hybrid planner is systematically tested and benchmarked to demonstrate its effectiveness in handling challenging road scenario with congested and chaotic traffic conditions.

I. INTRODUCTION

The planning module in an Autonomous Vehicle (AV) software stack is responsible to generate detailed trajectory for some horizon in the future and uses perception, localization, and map data as inputs. It is often comprised of three main sub-modules, route planner, behavior planner and trajectory planner. Route planner generates a lane-level navigation route for the AV to reach the destination based on the static map. Behavior planner makes the high-level decisions on what the AV needs to do to follow the navigation route considering the dynamics on the road. Lastly, trajectory planner will generate a detailed trajectory to follow the behavior planner decisions while ensuring a safe, comfortable, and law-abiding maneuver for AV. In this paper, we focus on the behavior planner module and present a new approach to address the current gaps in AV planning tasks.