A Review of Trajectory Prediction for Autonomous Driving from Accuracy to Robustness and Trustworthiness

Abstract

Trajectory prediction links environmental perception, behavior understanding, and decision-making and planning in autonomous driving. Its value depends not only on geometric error on standard test sets but also on model stability in open traffic environments, probabilistic reliability, and contributions to planning safety. With advances in interaction modeling, high-definition maps, multimodal generation, and Transformers, trajectory prediction has evolved from temporal extrapolation for a single vehicle into a spatiotemporal reasoning task that integrates multi-agent interactions, road semantics, and behavioral uncertainty. However, open-loop metrics such as Average Displacement Error (ADE), Final Displacement Error (FDE), minimum Average Displacement Error (minADE), and minimum Final Displacement Error (minFDE) primarily reflect average prediction accuracy. They reveal little about failures under perception noise, map deviations, out-of-distribution scenarios, long-tail interactions, and adversarial perturbations, and they cannot determine whether model probabilities align with actual errors and planning risks. Taking integrated robustness and trustworthiness as its organizing theme, this review first briefly surveys the trajectory prediction task and representative methods. It then focuses on robustness under input perturbations, scene changes, cross-domain generalization, and adversarial attacks, and discusses uncertainty quantification and propagation, probabilistic calibration, failure detection, and prediction-planning closed-loop evaluation. We further develop a unified framework comprising input-level robustness, scene-level robustness, distribution-level trustworthiness, and planning-level trustworthiness, bringing prediction accuracy, probabilistic quality, calibration reliability, failure detection, and planning safety within a single evaluation framework. Finally, we outline future research directions in unified evaluation, propagation of multi-source uncertainty, calibratable multimodal prediction, long-tail risk, and joint closed-loop evaluation.