Sensitivity Shaping for Latent Modeling

The paper introduces support-conditioned control-sensitivity regularization for generative dynamics models to enhance out-of-distribution (OOD) transition detection in robotic systems. By promoting sensitivity to control input changes in high-support training regions, the method preserves control-induced variation while reducing unstable extrapolation. Experiments demonstrate that this approach improves OOD detection and enhances the safety of closed-loop planning in tasks like vision-based obstacle avoidance and real-robot navigation, which is crucial for practitioners aiming to deploy reliable robotic systems.