DADiff: Diffusion-Driven Cross-Domain Policy Adaptation for Reinforcement Learning

DADiff addresses online dynamics adaptation in reinforcement learning, where a policy is trained with abundant source-domain data but can collect only limited target-domain interactions. The method uses a diffusion-based generative process for next-state prediction and estimates dynamics mismatch from discrepancies between source- and target-domain generative trajectories. It offers two adaptation variants: reward modification and data selection. The paper also derives a theoretical bound relating policy performance differences across domains to generative trajectory deviation, and reports experiments across multiple types of domain shifts.