DyDiff: Long-Horizon Rollout via Dynamics Diffusion for Offline Reinforcement Learning
For offline reinforcement learning practitioners, DyDiff offers a way to leverage diffusion models as accurate dynamics models for long-horizon rollouts without online interaction.
DyDiff addresses the policy mismatch in diffusion models used as dynamics models for offline RL by iteratively injecting learning policy information, achieving improved long-horizon rollout accuracy and policy consistency. It outperforms existing methods on offline RL benchmarks.
With the great success of diffusion models (DMs) in generating realistic synthetic vision data, many researchers have investigated their potential in decision-making and control. Most of these works utilized DMs to sample directly from the trajectory space, where DMs can be viewed as a combination of dynamics models and policies. In this work, we explore how to decouple DMs' ability as dynamics models in fully offline settings, allowing the learning policy to roll out trajectories. As DMs learn the data distribution from the dataset, their intrinsic policy is actually the behavior policy induced from the dataset, which results in a mismatch between the behavior policy and the learning policy. We propose Dynamics Diffusion, short as DyDiff, which can inject information from the learning policy to DMs iteratively. DyDiff ensures long-horizon rollout accuracy while maintaining policy consistency and can be easily deployed on model-free algorithms. We provide theoretical analysis to show the advantage of DMs on long-horizon rollout over models and demonstrate the effectiveness of DyDiff in the context of offline reinforcement learning, where the rollout dataset is provided but no online environment for interaction.