Cross-Domain Energy-Guided Diffusion Generation for Off-Dynamics Reinforcement Learning
Researchers have developed CEDGE, a novel framework for off-dynamics reinforcement learning that utilizes diffusion models to generate synthetic trajectories. This approach trains a diffusion model on source-domain data and then adapts these generated trajectories to a target domain using energy guidance. The energy guidance is designed to minimize distribution mismatches, allowing for efficient adaptation to new dynamics without retraining the diffusion model. Experiments show CEDGE improves trajectory generation for planning and enhances downstream policy learning. AI
IMPACT Introduces a new method for generating synthetic data in reinforcement learning, potentially improving policy learning in scenarios with mismatched dynamics.