Explainable and Safe Learning-Based Autonomous Driving: A Survey from Modular to End-to-End Planning

Learning-based autonomous driving is reshaping the automotive industry, shifting from modular architectures to end-to-end systems that promise easier maintainability and scalability. However, a critical challenge remains: these approaches often lack explainability and safety. This survey addresses that gap by reviewing research aimed at enhancing both dimensions. We examine a decade of work across three key paradigms of learning-based autonomous driving: (a) separate prediction and planning, (b) joint prediction and planning, and (c) end-to-end planning. Our analysis highlights strategies including model-based physical priors, risk-aware prediction, hierarchical designs, safety shielding, interpretable neural architectures, multi-task losses, and LLM-driven reasoning. Notably, end-to-end planning has surged in recent years, often outpacing progress on interpretability. We therefore conclude by recommending hybrid approaches that integrate physical models and safety constraint, ensuring that learning-based autonomous driving remains both explainable and safe.