Directional Pullback Regularization for Long-Horizon Rollout of Latent World Models in Autonomous Driving

This research proposes Directional Pullback Regularization (DPR) to enhance the quality of long-horizon rollouts in latent world models for autonomous driving. By assigning a kinematics-aware anisotropic Riemannian metric to latent transitions, DPR improves displacement error, lateral motion consistency, and forward motion preservation across various training regimes, achieving the best performance with 3-step rollout supervision.