Disentangling Task-Relevant Latent Dynamics for Reward Modeling

Reward design remains a major bottleneck for reinforcement learning in robotic manipulation. This paper proposes DiTaR, a framework that learns dense rewards from heterogeneous videos by disentangling task-relevant latent dynamics from visual distractions. DiTa, through an asymmetric two-stage procedure, suppresses task-irrelevant visual variation and trains a history- and language-conditioned rectified-flow model to capture task dynamics, leading to more reliable rewards and improved policy learning on MetaWorld-v3 and ManiSkill.