Reliable Multimodal Trajectory Prediction via Error Aligned Uncertainty Optimization

AbstractReliable uncertainty quantification in deep neural networks is very crucial safety-critical applications such as automated driving for trustworthy and informed decision-making. Assessing the quality of estimates challenging ground truth not available. Ideally, a well-calibrated model, should perfectly correlate with model error. We propose novel error aligned optimization method introduce trainable loss function to guide models yield good aligning Our approach targets continuous structured prediction regression tasks, evaluated on multiple datasets including large-scale vehicle motion task involving real-world distributional shifts. demonstrate that our improves average displacement by \(1.69\%\) \(4.69\%\), correlation \(17.22\%\) \(19.13\%\) quantified Pearson coefficient two state-of-the-art baselines.KeywordsReliable quantificationRobustnessMultimodal trajectory predictionError calibrationSafety-critical applicationsInformed decision-makingSafe artificial intelligenceAutomated drivingReal-world shift