Reinforcement Learning-based Decision-Making for Safe Motion Planning in Complex Driving Scenarios

Autonomous vehicles (AVs) face considerable challenges in decision-making and motion planning when operating in complex,dynamic environments, especially under safety-critical driving conditions. This research introduces SOLID−RL (StrategicOptimization for Learning in Intelligent Driving with Reinforcement Learning), a new technique that enhance and optimizeAV decision-making using reinforcement learning (RL) and dynamic programming. This methodology utilizes high-definition(HD) maps and environmental data to navigate urban scenarios. It employs a two-layer architecture: a high-level RL-baseddecision-maker that generates safe, rule-compliant actions and a low-level dynamic programming planner that optimizestrajectory generation. This method allows AVs to facilitate navigation in complex environments while adhering to traffic rulesand safety requirements. Simulations across urban scenarios, including intersections and overtaking maneuvers, demonstratethat SOLID−RL improves safety metrics, such as collision avoidance and path adherence compared to conventional methods.These findings contribute to the advancement of AV technology, offering a robust framework for safe and efficient autonomousnavigation in complex urban settings. This research paves the way for more reliable AV systems capable of handling thediverse challenges of real-world driving conditions.