In this paper we resolved dominant modality effect by focusing on loss land scape and gradient modulation

Encouraging model in model-based reinforcement learning to converge to flatter minima in the loss landscape will result in better downstream policies