Safe and explainable Artificial Intelligence for safety-critical robotic systems

The increasing integration of Artificial Intelligence (AI) into robotic systems introduces both significant potential and critical safety challenges. As many robotic functions rely on opaque, data-driven models, ensuring transparency and trustworthiness has become essential for deployment in real-world environments. Explainable Artificial Intelligence (XAI) has emerged as a key research direction for addressing these challenges. In this work, we systematically examine XAI methods for robotic perception, planning, and control, drawing strong parallels to existing research in autonomous driving (AD). Following the structure of prior surveys, we analyze five major XAI paradigms interpretable-by-design models, interpretable surrogate models, interpretable monitoring, auxiliary explanations, and interpretable safety validation and discuss how each can be applied or extended to robotics. Furthermore, we highlight the limitations of current approaches, especially the fragility and inconsistency of post-hoc explanation techniques such as attention and saliency maps. Building on insights from XAI frameworks for AD, we introduce a modular architecture for robotics that integrates layered explainability with safety monitoring. This survey provides a unified conceptual foundation for developing safe and explainable robotics, offering guidance for researchers, designers, and policymakers seeking trustworthy AI-driven robotic systems.