Autonomous driving AI has evolved from perception-centric architectures to imitation- and reinforcement-learning-based decision frameworks, and more recently to end-to-end and explainable multimodal vision-language-action (VLA) models. VLA integrates visual, linguistic and action representations to make driving decisions explicit, supporting reliability and accountability in accident responsibility attribution. Physical AI is defined as a paradigm that models vehicle dynamics and physical constraints while unifying perception, decision-making and control, enabling real-world deployment. This presentation discusses key challenges, including the sim-to-real gap, physical constraints, safety validation and explainability.