An Explainable RL-Based Speed Adaptation Framework for Autonomous Driving Using a Custom CARLA Environment

Engineering notes

Key topics: autonomous driving, autonomous vehicle, motion planning, reinforcement learning, carla, planning, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This study proposes a reinforcement learning (RL)-based control architecture within a hierarchical reinforcement learning (HRL) motion planning framework, designed to generate interpretable and safe driving behaviors for autonomous vehicles. The architecture comprises a high-level RL agent responsible for selecting behavioral modes based on strategic objectives, and several specialized low-level agents that execute specific control tasks. The focus of this work is on the speed adjustment problem, where a dedicated low-level agent is trained to regulate vehicle speed according to road curvature and a given reference speed. Training is conducted within a custom-built CARLA simulation environment modeling realistic traffic scenarios and road geometries. The agent learns to decelerate before curves, optimize its speed according to curvature intensity, and maintain lane discipline to ensure safe and efficient navigation. Performance and generalization capability are evaluated by comparing the trained model with an analytical solution and by testing in unseen environments. Interpretability is enhanced through the computation of SHAP (SHapley Additive exPlanations) values, providing insights into how input features influence the agent’s behavior and contributing to transparency and trust in autonomous systems. The proposed framework establishes a robust and explainable foundation for behavior planning in real-world autonomous driving applications.

Links and sources

• Official / arXiv page

Comments