HARL‐RSS: A Hierarchical Planning Framework for Autonomous Driving Based on Simplex Architecture and Responsibility‐Sensitive Safety

Today's autonomous driving technology heavily relies on large deep neural networks trained with deep learning, and it is challenging to achieve high levels of safety certification due to their complexity and black‐box nature. We present an approach to autonomous driving planning called Hierarchical Attention Reinforcement Learning Responsibility‐Sensitive Safety (HARL‐RSS), which consists of a high‐performance advanced controller, a safety controller, and a decision module that adaptively switches between the two controllers at runtime based on RSS. The advanced controller itself has a two‐level hierarchy, where a high‐level reinforcement learning (RL) agent adaptively selects among multiple low‐level RL agents, each trained for different driving styles, for example, aggressive versus conservative. Experimental results indicate that HARL‐RSS can strike an appropriate balance between safety and efficiency.