Human-Like Decision Making and Planning for Autonomous Driving with Reinforcement Learning

One of the main challenges faced by autonomous vehicles operating in mixed traffic scenarios pertains to ensuring safe and efficient navigation, particularly adhering to the implicit rules obeyed by human drivers. In this study, an Adaptive Socially-Compatible Hierarchical Behavior and Motion Planning (ASC-HBMP) framework is proposed to tackle the issue of socially-compatible navigation. ASC-HBMP comprehensively captures the attributes of other traffic participants to guide autonomous vehicles in devising human-like, safe, and efficient trajectories in a socially-compatible manner, striking a balance between safety and efficiency within complex multi-scenarios. Hierarchical Behavior and Motion Planning (HBMP) establishes driving tasks as high-level behavioral decision-making processes that emphasize efficiency, as well as low-level motion planning methods that prioritize safety. HBMP accepts the guidance provided by Adaptive Socially-Compatible Module (ASCM) to generate trajectories with diverse driving style characteristics. Finally, cross-platform simulation experiments are conducted on the SUMO and ROS simulators to validate the navigation performance and generalization capability of our approach in comparison to other baseline methods.