arxiv Interpretable Self-Aware Neural Networks for Robust Trajectory Prediction