Abstract
Trajectory prediction plays a vital role in the performance of autonomousdriving systems, and prediction accuracy, such as average displacement error(ADE) or final displacement error (FDE), is widely used as a performancemetric. However, a significant disparity exists between the accuracy ofpredictors on fixed datasets and driving performance when the predictors areused downstream for vehicle control, because of a dynamics gap. In the realworld, the prediction algorithm influences the behavior of the ego vehicle,which, in turn, influences the behaviors of other vehicles nearby. Thisinteraction results in predictor-specific dynamics that directly impactsprediction results. In fixed datasets, since other vehicles' responses arepredetermined, this interaction effect is lost, leading to a significantdynamics gap. This paper studies the overlooked significance of this dynamicsgap. We also examine several other factors contributing to the disparitybetween prediction performance and driving performance. The findings highlightthe trade-off between the predictor's computational efficiency and predictionaccuracy in determining real-world driving performance. In summary, aninteractive, task-driven evaluation protocol for trajectory prediction iscrucial to capture its effectiveness for autonomous driving. Source code alongwith experimental settings is available online.