XAI-Guided Fourier Domain Generalisation for Pure-Simulation Sim-to-Real Deep Reinforcement Learning

XAI helps identify what matters for autonomous driving decisions: The framework uses explainability techniques to determine which visual features are actually important for the driving policy, allowing it to distinguish between task-critical elements (like road markings) and irrelevant visual variations (like lighting changes).

Simulation-only training reduces real-world deployment costs: Unlike existing methods that require collecting actual real-world data to bridge the sim-to-real gap, XFD works entirely within simulation, eliminating the need for expensive and time-consuming data acquisition from target deployment environments.

Targeted robustness is more effective than random perturbations: Rather than randomly altering simulated images, the method strategically modifies only the visual components that don't affect driving decisions, resulting in more efficient and effective transfer to real-world conditions compared to unguided domain randomization approaches.

Zero-shot real-world performance demonstrates practical viability: The experiments show the trained policies work in real-world autonomous driving scenarios without any additional fine-tuning or real-world data, suggesting the approach could reduce barriers to deploying AI systems in safety-critical applications.

Explainability serves a practical safety function: By making the AI's decision-making process transparent through attribution maps, the framework enables verification that policies are learning appropriate features, which is crucial for responsible deployment of autonomous systems.