Abstract: Imitation learning is increasingly utilized to improve driving performance using real-world data, yet ensuring the safety of its outputs remains a fundamental challenge. While differentiable ...