An End-to-End Decision-Aware Multi-Scale Attention-Based Model for Explainable Autonomous Driving

Explainability is critical for autonomous vehicle safety: The paper emphasizes that without understanding why self-driving cars make decisions, we cannot identify when they might fail or use them safely in real-world conditions—a fundamental requirement for responsible deployment.

Current explanation methods are unreliable: Existing approaches to explaining autonomous vehicle behavior suffer from "flawed reasoning and unreliable metrics," meaning we may have false confidence in systems we don't actually understand well, which poses serious safety risks.

Decision-specific explanations matter: The proposed model provides tailored explanations for each individual driving decision rather than generic explanations, suggesting that truly trustworthy autonomous systems need to justify their choices in context-dependent ways.

Better evaluation metrics are needed: The paper introduces a new "Joint F1 score" metric specifically designed for explainable AI, highlighting that standard performance measures alone are insufficient—we need metrics that actually measure whether explanations are accurate and reliable, not just whether predictions are correct.

Generalization and robustness testing strengthens trust: By validating the model across multiple datasets (BDD-OIA and nu-AR), the research demonstrates the importance of testing explainability systems in diverse scenarios before real-world deployment, rather than relying on single-dataset validation.